OVERVIEW

You get your blood drawn to detect cancer early.

You get insurance on your IVF procedure.

You automate launching a consumer brand.

You buy futures on GPU prices.

You employ babysitters for your AI agents.

You shop in physical stores the way you do in ecommerce.

This month we rename the “Data & AI Trends” report with a label more true to our purpose: the Innovation Monitor. Our goal is to keep leaders informed about what’s happening in data and AI use cases, at the edge: where the investment bets are being placed, where the world’s fiercest entrepreneurs, engineers, and scientists are sweating like newbies. We do this so you can monitor the flow of innovation in this insanely fast-paced market.

And what a month to make that change: in May, a fat $98.6 billion was dropped on data and AI companies. As always, we track the companies (54 this month) that raised more than $30MM and are doing something specific with data that you can borrow from, for your business.

This month, the lessons are super-abundant. The Healthcare section alone is mind-bending--if nothing else, skip down and scan to make you think you’re living at the cusp of something utopian.

The Top 10 Single Raises

The biggest individual financings this month among the use-case companies we profile:

1. Mobility Global: $2.5B (Senior Notes / Debt). Auto data.

2. Isomorphic Labs: $2.1B (Alphabet Funding/Capitalization). Drug development.

3. RoboStrategy: $2B (Committed Equity Facility). Robotics fund.

4. Sierra: $950MM (Venture). Customer service.

5. Tempus: $400MM (Public Company). Oncology data.

6. ICEYE: $347.9MM (Credit Facility). Satellite data.

7. Decart: $300MM (Venture). World models for video.

8. Cowboy Space: $275MM (Series B). Data centers in space.

9. Credibly: $260MM (Securitization / Debt). Small business credit.

10. Blitzy: $200MM (Venture). Coding agents.

Trends This Month

1. Data is king

This month is a reminder that data is the fuel for AI, and there’s really no replacing it. Massive, sometimes rare, sometimes public data drives the value in many of thing month’s raises. CARFAX, part of the $2.5 billion Mobility Global spinout, holds billions of vehicle records keyed to your VIN, scraped for decades from DMVs, body shops, police reports, and auctions. Tempus has the sequenced tumors and treatment outcomes of tens of millions of cancer patients. RADAR records how shoppers and products move through 1,400 stores. Daloopa pulls finacial data from public sources and sells it to the LLMs as source data. Isomorphic Labs aims to accelerate drug discovery through the big public Protein Data Bank. Consensus gathers 220 million scientific papers for search. Vi Labs helps hospitals interpret clinical data. Gaia is creating financial services for IVF treatments. Data--gathering it painstakingly over decades or with lightning fast automation--provides the fives senses for our models, and without it, they fly blind. For our readers: a reminder that the data your business runs on deserves your attention. If you’re not using it yet, you should.

2. Value in the grubby wiring of workflow

When we published our piece on the AI Sandwich--a framework for interpreting AI business models--we might have underplayed the value of workflow integrations in the “top bread” layer. Sierra processes customer returns. Commure says it closes more than 85% of a hospital’s billing and claims with no human touching it. Pace runs insurance operations straight through the mainframes carriers were supposed to retire two decades ago. Nova and Tessera climb into decades-old SAP code and rewrite it. All of them rent the language model from the same three labs; the value they provide is in managing the last mile, the grubby wiring into the old systems of record, where (in our 7D parlance) the data and decision parts of the value chain can actually be deployed. For a leader like you, these investments highlight how valuable it is to smooth the friction points in your business.

3. Agentic “total cost of ownership”

As we’ve noted, you can tell a technology has truly arrived when a generation of companies get funded purely to keep an eye on the original tech. This is happening now for agents: Gray Swan runs a global contest where 15,000 hackers try to jailbreak frontier models, then sells the resulting attack playbook back to Anthropic, OpenAI, and Meta—a coordinated red team action big as a Civil War reenactment. Cinder moderates content for the big AI labs, policing what their own models say. Geordie gives a security team a control screen for every agent operating in a business. Judgment grades whether your agents actually did the job you gave them. Frame fights AI-written phishing with AI-written phishing drills. Every new thing you let an agent do… opens a fresh way for it to go wrong. These startups are there to sell you a safety plan. This implies that the total cost of ownership of AI is ballooning--you have to buy figurative insurance and backup generators for the software. This needs to be factored into dreamy cost-cutting projections when assessing the return-on-project value for AI.

4. GPUs as trade-able commodity

Ornn is a one-company trend this month. Backed by the company that owns the New York Stock Exchange, they are building a futures market for renting GPUs—the same way wheat and crude have futures—because compute prices are now known to swing enough to blow a budget mid-training. If you have plans that assume cheap, on-tap compute, it might be worthwhile pressure-testing that and locking in capacity… talking to your CFO about hedging, the way you if would hedge fuel at an airline. What will be securitized next in this new commodity system… data center capacity? Model safety? Model speed?

THE COMPANIES

Finance: The Back Office Goes Agentic

$2.99 billion invested across 10 companies

Finance pulled in the most money of any sector we cover. While this is tilted towards financial plays the RoboStrategy fund, from a use case perspective it’s basically two ideas: software that reads messy unstructured, financial documents faster than a person can; and automating risk bets for insurers and credit.

RoboStrategy. RoboStrategy is a publicly traded fund that gives investors exposure to private robotics and embodied-AI companies such as Figure and Apptronik. It builds no technology itself. The figure is a committed facility to sell its own stock over time, not cash raised. ($2,000MM, May 15, 2026)

Credibly. Credibly underwrites working-capital loans to small businesses using their bank-transaction and cash-flow data rather than personal credit scores, trained on the repayment history of more than 60,000 prior loans. The figure is a securitization and warehouse financing, not equity. ($260MM, May 12, 2026)

Capchase. Capchase finances B2B purchases, paying vendors upfront while their buyers pay in installments; an AI agent reads a sales quote and returns an approved payment link in under a minute. Most of the figure is a credit facility for lending. ($174MM, May 27, 2026)

Reserv. Reserv is an AI-first, outsourced claims administrator for property-and-casualty insurers. It migrates a carrier’s historical and open claims, much of it unstructured paperwork, into a single database, then runs what it calls explainable AI to handle routine claims from first notice of loss through settlement, while its roughly 500 human adjusters take the tricky files. “Traceability” matters here because a regulator or carrier can demand to know why any decision was made. The company says it handles about 500,000 complex claims a year and wants to reach 30 million within four years. KKR led the round. ($125MM, May 4, 2026)

Elliptic. Elliptic maps blockchain addresses to real-world actors and scores crypto transactions for sanctions and money-laundering risk; it is adding AI agents to handle the repetitive investigative work compliance analysts now do by hand. ($120MM, May 12, 2026)

Fonoa. Fonoa automates indirect tax — VAT and GST — for global digital companies like Uber and Netflix: it tracks rule changes across 190-plus jurisdictions, calculates the right tax on each transaction, issues e-invoices, and files the returns. It doesn’t remit the money itself; the business still pays the authority.. ($110MM, May 28, 2026)

Moment. Moment, founded by former Citadel Securities quants, automates the manual back-office work behind fixed-income and wealth management — building and allocating trades across client accounts, wrangling data, and the like — for firms that together oversee more than $10 trillion. ($78MM, May 19, 2026)

Daloopa. Daloopa uses machine learning to extract financial figures from filings and investor decks for more than 5,500 companies, links each number to its source document, and pipes that data into AI finance tools through connectors to ChatGPT, Claude, and Perplexity. ($47MM, May 28, 2026)

Pace. Pace builds AI agents that run insurers’ back-office workflows across legacy systems, including green-screen mainframes, configured in plain language; it says it has completed more than 250,000 such workflows. ($46MM, May 27, 2026)

Ornn AI. Ornn wants to make renting computing power something you can trade, the way oil or wheat is traded. It starts with a reliable price — a published benchmark for what it actually costs to rent Nvidia’s top AI chips (the H100, H200, and Blackwell), built only from deals that actually closed rather than the sticker prices vendors advertise, and carried on the Bloomberg Terminal that traders already watch. Based on that it’s creating futures contracts that let a buyer lock in a rate today for compute they’ll use later, the way an airline hedges fuel. They settle against the average price over a stretch of time, unlike normal future contracts where you settle up on a set date. These swings can be serious--Ornn’s index saw the rental price of Blackwell chips jumping 48% in two months which would be enough to wreck the budget on a long training run for a model. Days before this raise, the company that owns the New York Stock Exchange said it would launch its own compute futures based on Ornn’s prices. ($33.7MM, May 26, 2026)

Healthcare: Liquid Biopsies

$2.91 billion invested across 7 companies

This month the innovation is accelerating in really mind-bending ways. You really need to read every item here. Massive specialized data sets (Isomorphic, Tempus); modeling drugs in software, not labs.

Isomorphic Labs. Isomorphic Labs spun out of Google DeepMind to turn AlphaFold, the system that solved protein folding, into a drug-design factory. Its models go beyond predicting a protein’s shape… to predicting how a candidate molecule binds to it, how tightly, and whether it will behave like a real drug… work that normally requires slow, costly lab experiments. The training data is the public Protein Data Bank plus the company’s own drug-design data, and the aim is to start qualifying early drug candidates using software, before anyone starts incurring lab costs. It has deals with Eli Lilly and Novartis, and CEO Demis Hassabis describes the mission, with a straight face, as “solve all disease.” Drug discovery costs billions and fails about nine times in ten, so cutting risk out of the earliest phase is worth real money. Massive mission, massive raise. ($2,100MM, May 12, 2026)

Tempus. Tempus sequences patients’ tumors and feeds the results, plus access to AI tools, back to oncologists who are deciding what to do next. Founder Eric Lefkofsky started it after his wife’s cancer diagnosis showed him how little data informed her treatment. The data asset (claiming 450 unique data connections as sources) is one of the largest collections of linked molecular and clinical records in oncology, tens of millions of de-identified patient journeys, more than a million of them with genomic data, gathered because roughly half of US oncologists order its tests. A doctor ordering a Tempus test gets back not just the sequencing but a match to therapies and trials, and the company is now training foundation models on that data. Pharma companies pay for the same data to design trials and find patients, a real revenue line. (Tempus is already public, so treat the May figure as a capital-markets event rather than a fresh venture round.) ($400MM, May 8, 2026)

Vi Labs. Vi runs industry-specific AI agents on top of a large store of clinical, behavioral, and operational health data, performing real-time tasks such as patient navigation and physician next-best-action recommendations for hospitals, drugmakers, and wellness companies. ($145.7MM, May 19, 2026)

Gaia. Gaia trains models on millions of anonymized fertility records to forecast a patient’s odds of a successful IVF birth, then uses that forecast to underwrite financing that refunds or discounts failed cycles. It’s like outcome insurance on a procedure. The founder built it after his own multi-year, six-figure fertility journey. ($100MM, May 20, 2026)

Commure. Commure runs a system of AI agents across hospital operations, transcribing the doctor-patient visit into a billable note and then handling billing and claims. The company claims to complete more than 85% of revenue-cycle work without a human. ($70MM, May 19, 2026)

ClearNote Health. ClearNote Health offers a test (a blood draw or “liquid biopsy”) that looks for signs a cancer is already present but not yet causing symptoms or showing up on normal workups. It’s aimed at pancreatic and ovarian cancers specifically, because those are notorious for being found too late--when they’re hard to treat. Instead of reading only the genome, it reads the epigenome, tagging a chemical mark on DNA called 5-hydroxymethylcytosine that shows when cancer starts. The lab pulls the DNA fragments carrying that mark out of a blood sample, sequences them, and trains machine-learning models on the profiles of patients known to have or not have, cancer. Its pancreatic test (as an example) fuses that epigenomic signal with genomic and other markers, and the company reports strong accuracy in high-risk patients such as new-onset diabetics. Catching these cancers months earlier than with normal methods can be the difference between a treatable case and a fatal one. ($52MM, May 27, 2026)

Genesis Molecular AI. Genesis, formally Genesis Therapeutics, pairs molecular machine-learning models with experienced drug hunters to design small-molecule medicines. Its models predict the properties that decide whether a candidate becomes a drug, how tightly it binds, how selective it is, whether it’s safe and manufacturable, so chemists can search through vast chemical variation and scenarios without actually making every molecule. Big Pharma clients hand Genesis targets they wants drugs for; pay Genesis to run their AI (”GEMS” platform) against them; and also provide their own proprietary experimental lab data in to train Genesis’s models. ($40MM, May 20, 2026)

Automotive Data: CARFAX Unleashed

$2.50 billion invested in 1 company

One company fills this section, but it’s a heavyweight, and an old friend. For those of us who remember Polk… well now it (plus other assets) is Mobility Global: the automotive-data business being spun out by S&P Global. It’s a lesson in what a decades-old data moat is worth.

Mobility Global. Mobility Global is the automotive-data arm S&P Global is spinning off into its own NYSE company, and it owns data brands most car buyers have used without realizing it: CARFAX, Polk, automotiveMastermind, and Market Scan. CARFAX builds a history for nearly every vehicle by ingesting billions of records from more than 100,000 sources, including DMVs, police and fire departments, service shops, and auctions, all keyed to the VIN. Polk is effectively the census of who owns which car. AutomotiveMastermind scores which of a dealer’s customers is most likely to buy next. Almost none of this is flashy generative AI; it’s predictive analytics on proprietary records built up over decades, and the moat is simply that no competitor holds the same data. The figure is debt raised to stand the company up on its own, not a venture round. ($2,500MM, May 19, 2026)

Media and Marketing: Fixing Customer Service at Last

$1.89 billion invested across 5 companies

My focus in this group is Sierra, which is aiming to finally fix customer service with AI after some genius back in the 80s (bet it was McKinsey) decided that corporations could save money by torturing customers with unfindable customer service numbers and long wait times.

Sierra. Sierra, founded by former Salesforce co-CEO Bret Taylor and ex-Google executive Clay Bavor, helps big companies build AI agents that handle customer conversations across chat, voice, email, and messaging. Rather than bet on one model, it routes each task to the best of more than a dozen, fast models for lookups… cautious ones for sensitive moments… and then supervisor agents on top to enforce company policy. The agents are grounded in each client’s own knowledge and integrated into their back-end systems. Sierra says it serves more than 40% of the Fortune 50; customer service is a giant, measurable cost… and arguably a source of customer dissatisfcation… so an agent that resolves the whole issue, in theory, should create both savings and value. The round valued the two-year-old company at roughly $15 billion. ($950MM, May 4, 2026)

Decart. Decart makes AI “world models” that turn a live video feed into an interactive, explorable scene in real time. ($300MM, May 18, 2026)

ElevenLabs. ElevenLabs trains its own deep-learning speech models for text-to-speech, voice cloning, and dubbing. Note the date: its headline 2026 financing was a $500MM round in February, and the May activity appears to be a roughly $100MM employee share sale rather than new capital. ($100MM, May 6, 2026)

Astrocade. Astrocade, founded by two PhD brothers, lets anyone build a playable game from a text prompt by generating the art, music, and mechanics together; its catalog of more than 75,000 user-made games feeds back into the models. ($56MM, May 5, 2026)

Nectar Social. Nectar Social, founded by two sisters, former Meta-ites, runs AI agents on licensed social-platform data to manage community, engage shoppers, and tie social activity back to revenue. ($30MM, May 14, 2026)

AI Tools: Coders, Slackers, Rockets

$0.962 billion invested across 10 companies

This is the grab-bag for companies whose product is AI itself rather than any one industry, and the theme is agents that do work. Cowboy Space is the attention grabber. Solar powered data centers in space.

Cowboy Space. Cowboy Space, founded by Robinhood co-founder Baiju Bhatt, is building rockets whose upper stages double as solar-powered orbital data centers, on the bet that power, not chips, is the binding constraint on AI and that the grid’s multi-year connection queue makes space the faster route. ($275MM, May 11, 2026)

Blitzy. Blitzy converts an enterprise’s entire codebase into a reference database, then runs thousands of coding agents, calling foundation models more than 100,000 times per run, to generate and validate millions of lines of code for legacy systems. ($200MM, May 5, 2026)

Sigma Computing. Big companies keep their data in a “cloud warehouse” — a giant central database run by the likes of Snowflake or Databricks. Normally only technical staff can pull from it, so everyone else waits on reports or copies bits into their own spreadsheets, which is slow and a security risk. Sigma gives non-technical employees a familiar, Excel-style screen to explore that warehouse data directly, without writing code and without copying anything out. Because Sigma’s queries (and now its AI agents) run inside the warehouse instead of pulling data into a separate tool, the company’s existing rules about who’s allowed to see what apply automatically. ($80MM, May 18, 2026)

Viktor. Viktor, built by former Meta engineers, is an AI assistant you talk to right inside your company’s chat app — Slack or Teams — like messaging a coworker. What makes it more than a chatbot is that it’s plugged into the other software the company runs on, so instead of just answering, it can go do things: pull a real report from the actual numbers, or even spin up a small internal tool. Over time it learns how your specific company works — your systems, your jargon, your routines — so its help gets sharper. In one example it caught a company burning $10,000 a week on wasted ad spend and flagged it. ($75MM, May 19, 2026)

Vapi. Vapi gives developers the infrastructure to build voice agents for phone calls, combining speech-to-text… a language model… and text-to-speech… behind a low-latency API. It reports handling more than a billion calls for customers including Ring and New York Life. ($50MM, May 12, 2026)

Monaco. Monaco sells an AI-native sales platform for startups that replaces a traditional CRM and its add-on tools, automatically building a target market, running outbound comms, and logging interactions. ($50MM, May 12, 2026)

Dust. Dust, founded by a former OpenAI engineer, lets companies build AI assistants connected to internal knowledge and tools and share them across teams in a common workspace. ($40MM, May 18, 2026)

Asteromorph. Asteromorph, a Seoul lab staffed largely by young science-olympiad winners, is building autonomous AI systems meant to generate scientific hypotheses and organize discoveries in biology and chemistry into a structured database. ($31MM, May 29, 2026)

Luel. Luel runs a marketplace that pays more than 500,000 contributors across 96 countries to produce multimodal training data, then sells audit-ready datasets to AI labs. ($31MM, May 15, 2026)

Polsia. Polsia claims its AI agents can run most of a business, including coding, sales, marketing, and support, and says the agents even ran this fundraise. The claims are largely company-supplied and have drawn skepticism. ($30MM, May 25, 2026)

Defense and Space: Sensors That Make New Data

$0.59 billion invested across 4 companies

The defense and space investments went to companies that generate physical-world data from satellites, and the ocean.

ICEYE. ICEYE operates the largest fleet of radar satellites, which see through cloud and darkness and applies AI to the returns for all-weather military intelligence and rapid flood maps for insurers. ($347.9MM, May 21, 2026)

HavocAI. HavocAI builds autonomous surface vessels and software to let one operator command thousands of them… the target client is a big one: the US Navy has a plan to field large numbers of cheap uncrewed boats in the Pacific by 2030. ($100MM, May 12, 2026)

Sidus Space. Sidus Space manufactures small satellites and sells space-based data-as-a-service. ($100MM, May 27, 2026)

Quartermaster AI. Quartermaster starts with a shipping problem: the system meant to track ships, called AIS, is self-reported and easily switched off or spoofed, which is how smuggling and sanctions evasion happens. So: it mounts weather-hardened cameras and radios on ships’ masts and fuses the feeds into a single, independent picture of ocean traffic, a crowd-sourced hive mind across hundreds of vessels, with computer-vision models on board identifying and classifying nearby ships. The imagery the network gathers becomes a second product: scarce real-world ocean data sold to marine-autonomy and robotics companies that use it like Go-Pro or Google cars data is used for autonomous vehicles. To get hardware onto ships cheaply, it makes the gear appealing to mariners, whose vessels then have more sensors for their environment, for safety. ($43MM, May 20, 2026)

Professional Services: More Agent Babysitters

$0.47 billion invested across 10 companies

Enter. Enter became Latin America’s first AI unicorn by chasing a problem unique to its market: Brazil has roughly 80 million active lawsuits, about eight times the US caseload, almost all of it filed in a fully digital court system. Enter’s agents ingest those filings and work the whole litigation cycle, analyzing suits, organizing evidence, flagging fraudulent serial claimants, recommending settle-or-fight, and drafting defenses. Because the court data is so voluminous and repetitive, a model tuned to Brazilian procedure beats generic legal tools. It processes more than 250,000 new cases a year for clients such as Itau, Santander, and Nubank, each of whom faces thousands of near-identical consumer suits. One co-founder argued a case before Brazil’s Supreme Court as one of the country’s youngest lawyers. ($100MM, May 5, 2026)

Tessera Labs. Tessera uses teams of AI agents to automate one of the most dreaded jobs in big companies — upgrading the core software that runs the business, like SAP. Every change is logged and rule-checked, so a company can trust software to touch the systems it normally guards most closely. ($60MM, May 6, 2026)

Frame Security. Frame Security generates company-specific phishing simulations and training with AI, replacing the static annual security course, on the premise that AI now writes the personalized attacks employees fall for--so it might as well also provide the training on the other side. ($50MM, May 11, 2026)

Lexroom.ai. Lexroom builds a legal AI for civil-law Europe, starting in Italy, grounding every answer in a curated base of millions of verified statutes and cases. ($49.9MM, May 19, 2026)

Cinder Technologies. Cinder turns a platform’s written content rules into AI agents that moderate abuse and fraud in more than 100 languages; its customers include OpenAI and Black Forest Labs, meaning AI labs now buy tools to police their own models’ output. ($41MM, May 12, 2026)

Gray Swan. Gray Swan, spun out of Carnegie Mellon’s AI-safety group, sells the tools frontier labs use to stress-test their models. It runs a global, gamified contest where more than 15,000 researchers try to break the latest models, an approach which has produced over a million real-world attack examples. That attack data trains a red-teaming agent which hammers models before release. Anthropic, OpenAI, and Meta are customers, because they have to show their models resist jailbreaks before shipping. As companies put agents into production and open new ways to be attacked, the idea is theses defenses will be valuable to enterprise. ($40MM, May 28, 2026)

FLUX. FLUX supplies Japanese companies the engineers and tools to adopt AI, spanning consulting, an AI-driven recruiting service, and marketing automation, for enterprises short on labor and in-house AI talent. ($37.7MM, May 29, 2026)

Judgment Labs. Judgment Labs evaluates AI agents by examining the full chain of an agent’s decisions and tool calls (not just the final output), then feeds that analysis back to improve agents. ($32MM, May 12, 2026)

Nova Intelligence. Nova Intelligence, led by a former DeepMind engineer, points agentic AI at the custom SAP code running payroll and supply chains, rewriting it for the mandated 2030 migration off the legacy ERP platform; one customer estimated the work went about five times faster. ($31.5MM, May 5, 2026)

Geordie AI. Geordie AI gives security teams a real-time view of every AI agent running inside a company, what data each can reach, and how it behaves, plus controls to rein them in. ($30MM, May 28, 2026)

Retail: Store Data Like Web Data

$0.23 billion invested across 2 companies

RADAR may be a game changer: a multiple-category (eg CPG, DIY) retail data generator.

RADAR. RADAR mounts sensors in store ceilings that read RFID tags and combine them with computer vision to take a full inventory snapshot every eight seconds, accurate to about 99% and down to where each item sits on the floor. RADAR also does “just walk out”-style checkout: as shoppers pick items up, they’re added to carts automatically, and the shopper is charged the moment they walk out the door. The bigger story here may be the data business: by capturing how products and people move, RADAR says it’s building one of the largest datasets of in-store behavior, the kind of signal e-commerce always enjoyed and physical retail never did. It’s deployed with American Eagle and Gap’s Old Navy, 1,400 stores in total. The founder studied RFID under the MIT professor credited with inventing it, and once turned his apartment into an Airbnb to keep the company alive. ($170MM, May 19, 2026)

Zyg. Zyg, built by the founders of the ad-tech company ironSource, is software that tries to run almost everything it takes to grow a small online brand: building the website, making the ads, buying the ads, winning repeat customers, even handling shipping. It does this with more than 60 AI agents that all draw on the same shared pool of the brand’s data, so what the ad side learns informs the store side and vice versa. Because every part sees the same data, Zyg can forecast things like how many customers will stick around and how much product will sell, and before a brand even launches it scores how likely the product is to take off, so a founder doesn’t pour money into a flop. The aspirational idea here, is that one person with a good product can go up against a big brand without hiring a marketing department. ($60MM, May 5, 2026)

Education: Teaching AI, With AI

$0.20 billion invested across 3 companies

The education investments split between teaching people AI skills… and building AI tutors.

KIDZ AI. Not Kars4 Kids, kids. KIDZ AI is the new name for Classover, a small public online-tutoring company that has rebranded around AI — promising AI tutors, a kids’ robotics curriculum, and even its own cloud-computing capacity. ($100MM, May 22, 2026)

Multiverse. Multiverse, founded by Euan Blair, retrains companies’ existing employees in AI and data skills through apprenticeship-style programs, on the logic that reskilling current staff is cheaper than hiring scarce AI talent. ($70MM, May 15, 2026)

Consensus. Consensus is an AI search engine for science that draws on more than 220 million papers, ranks them by research quality, and only then uses a model to summarize with citations, keeping answers grounded in real literature. ($30MM, May 11, 2026)

Manufacturing: Cameras and Copilots

$0.07 billion invested in 2 companies

Maneva. Maneva runs computer-vision models on a factory’s existing cameras to spot defects and anomalies in real time and trigger action. Its CEO is a former DeepMind engineer who worked on video-to-action AI. ($38.4MM, May 13, 2026)

Cosmon. Cosmon builds an AI agent for mechanical engineers that handles time-consuming CAD edits, drawings, and setup with multimodal models that read geometry and physics. They claim every change is traceable and reversible. ($31.2MM, May 8, 2026)

EDITOR’S NOTE

Several times a month we deep-dive into a company that just raised money and we pull it apart with our “7Ds” framework to see exactly how it uses data and AI—and what you can learn from it. This week: Afresh, a San Francisco company that raised $34 million on April 21st to use AI in the grocery store “fresh” (think perishables) department.

We found the combination of data, AI, rotten meat and bruised peaches to be an irresistible combination. And also, the great data challenge of calibrating something as ephemeral as lettuce on a store shelf.

BANANAS, BANANAS

The US grocery business is pretty enormous. We have 45,000 supermarkets in the US, and about 40% of the business, in dollars, is perishables. Those supermarkets (and other formats) together sell about $487B in perishables… but here’s the thing. We throw away 147 billion pounds of food every year, which is about 30% of all the food we make.

In weight, that’s the equivalent of 500 billion bananas.

It means every American is throwing away 2.3x their body weight in food every year. Which is nearly 100% more than what Mexico dumps on a per-capita basis, and (of course) almost 2x what they flush in France.

When you put that in the context of grocery store margins--notoriously slim; The Food Industry Association puts the industry’s net profit margin at 1.7%--you see why food waste is a big deal. (It’s sometimes called “shrinkage” or “shrink” in the industry, a term that includes theft and other mishaps) A few points of spoilage can flip a store from profitable to losing money.

So that’s bad; lost profit. But there’s a darker angle. Only 2% of surplus food is donated (according a report by nonprofit ReFED.) About 85% rots in the field unharvested or ends up in compost or landfills… meanwhile, 47.4 million Americans--about 13.5% of households, including 14 million children--live in food-insecure homes (USDA). The food bank and pantry system led by an organization called Feeding America supports a network of more than 200 food banks and 60,000 pantries, which scoops up over 4 billion pounds of food a year… yet it still cannot close the gap.

One last cost to society: if you throw away 500 billion bananas… all the water you used to grow them.. the labor you used to plant and pick them… the gas (and pollution) you used to ship them… was also, all waste.

All told, we grow plenty of food for all. But we lose about a third of it between farm and fork. So anything that keeps good food in circulation… to be sold or donated… chips away at both problems.

Enter our company of the week, Afresh.

HOW DO YOU GET DATA ON A PEACH?

Why is fresh so wasteful? Because it is the hardest thing in the store to predict. A can of clam chowder is good for two years; a clamshell of strawberries lasts mere days. Demand swings with the weather, the holidays, the weekend forecast.

Our readers know that if prediction is the problem, then data is likely the solution. Yet accurate data on fresh food is devilishly hard to procure. A barcode-scanned box of cereal leaves a clean digital trail; a bruise-y peach gets quietly tossed, and that info never touches a data-gathering system.

One part of Afresh’s job, then, is making data that does not otherwise exist: getting store associates to count fresh inventory through an app. Then they combine it with the retailer’s own data.

They use this data for a pretty simple set of predictions and decisions: net of spoilage, what inventory will I have on a daily basis; and--again net of anticipated spoilage--what should I order, so I don’t over or under-order?

The company says their software now runs in more than 12,500 departments across 40 states, and that 2025 revenue grew 70% year over year. Investors include Spark Capital, Insight Partners, Innovation Endeavors, Maersk Growth, and, in the latest round, Just Climate and High Sage Ventures.

Afresh is not alone. There’s Shelf Engine (automates ordering of perishables); Invafresh (cloud software for fresh merchandising and replenishment); Freshflow and Aravita (both pitch AI order forecasting). While the sector might need some naming consulting, it’s projected to boom nonetheless: AI in food retail generally is projected to grow from $2.7 billion in 2024 to $13.4 billion by 2030.

WHAT IT MEANS FOR YOU

• Look for your own “latent data.” Afresh uses Point of Sales (POS) product scans, and store delivery records… data that was already there… to feed new models. What transaction or operational data are you already collecting and ignoring?

• Find the place where people still use a clipboard. As you’ll see in the origin story below, the Afresh founders saw a grocery store manager’s handwritten food order guide, and knew there was a gap to fill. Where in your operation are high-dollar decisions still made on memory, guessing, scribbled notes?

• When the data does not exist, consider gathering it. Afresh built a food counting app that seems to be pretty efficient, even if it’s gathered by store employees. Is there a critical input you would have to create from scratch--but where owning that data would open new possibilities? Remember, the data asset--even if it expires at the end of the week like in a lettuce section--is still valuable for the future because it’s predictive.

• Keep a human in the loop. Afresh recommends order sizes; an associate confirms them--or overrides them. Which decisions should stay a recommendation for a person to sign off on, vs. become (or remain) fully automated?

THE 7Ds OF DATA INNOVATION FOR AFRESH

COMPANY OVERVIEW

Afresh builds software that tells grocery stores how much fresh food to order, make, and stock. Its AI reads a store’s sales, deliveries, and inventory, and produces an item-by-item order recommendation for perishable departments--produce, meat, seafood, deli, bakery--and, increasingly, for the rest of the store.

Date of Founding: 2017. Headquartered in San Francisco.

CEO / Founder Profile: Matt Schwartz, the CEO, came to the problem through food and sustainability rather than retail. He and co-founder Nathan Fenner met at Stanford’s Graduate School of Business, where Fenner brought an engineering background and Schwartz was researching how to build a more sustainable food system. Schwartz’s stated view is that fixing the food system is the single smartest way to improve human and environmental health.

Founding Story: In their second year of business school, Schwartz and Fenner ran an independent study on the fresh food supply chain, working from a thesis that fresh was the future of grocery yet barely touched by modern technology. To test it, they walked into a local grocery store. The produce section was out of stock on a lot of items. They found the department manager, who showed them how he ordered: paper, filled in by hand each day, for hundreds of items, equating to millions of dollars of product per year at a multi-billion-dollar chain. They founded Afresh in 2017; co-founders also include Volodymyr Kuleshov and Sha Sebastian.

Total Investment Raised: Afresh raised a $12 million Series A in July 2020 (led by Food Retail Ventures, with Innovation Endeavors, Maersk Growth, and Baseline Ventures), then a $115 million Series B in August 2022 that the company said brought total funding to $148 million. In April 2026 it added $34 million co-led by Just Climate--the climate fund of Generation Investment Management--and High Sage Ventures.

What the Company Does: The platform now includes six products. The core is store-level ordering: each morning it hands a department manager a recommended order for every item. In addition there is inventory management, in-store production planning (how many rotisserie chickens or muffins to make), and distribution-center forecasting (what the warehouse should buy). Afresh has pushed beyond fresh into center store, frozen, and general merchandise. It does this without dedicated hardware or shelf cameras--no robots, no fixed sensors--relying on the retailer’s data and counts entered by associates.

How the Company Makes Money: Afresh sells enterprise software to grocery retailers, generally as a subscription priced by the number of stores and departments deployed; exact pricing is not public.

Top Named Clients: Albertsons (a chainwide deployment across its banners, tied to a goal of cutting food waste 50% by 2030), WinCo Foods, Heinen’s, Save Mart, Bashas’, Cub Foods, Stater Bros., Meijer, Wakefern, and Fresh Thyme.

DETAIL - THE 7Ds OF DATA INNOVATION

1. DEMAND — “Who is their user?”

The Fresh Department Manager. The produce, meat, or deli manager who runs one perishable section of one store. Often a long-tenured, first-line employee, close to the floor, who knows the regulars and the rhythm of the week. They are rewarded based on in-stock levels, shrinkage, and how the department looks (full shelves, nothing rotting). Their day starts early, walking the section, eyeballing what’s left, judging what will sell, and placing their orders.

The Store Director. The person accountable for the whole store’s profit and loss, measured on total shrinkage, labor, and of course sales. They juggle dozens of departments, and fresh is the riskiest one on the floor.

The Corporate Fresh Merchant. The chain-level category leader who sets assortment (grocery word for the menu of food you sell). They want item-level visibility across hundreds of stores at once. They also own public sustainability targets--like that Albertsons committment to cutting food waste 50% by 2030, which has made them a USDA Food Loss and Waste Champion.

2. DILEMMA — “What makes the goal difficult?”

For the department manager, fresh ordering is tricky: over-order and it spoils within days; under-order and the shelf is empty by afternoon meaning missed sales and sniffy customers. There’s also complexity: a single section can carry hundreds of items, each with its own shelf life and demand curve; shrinkage for fresh fruit runs from 4% for bananas 40%+ for other items. The order tends to get placed under time pressure, with deliveries running on fixed schedules.

For the store director, spoilage hits that 1.7% margin, on top of the labor cost and tedium of counting fresh produce by hand. High employee turnover means the institutional knowledge of “how much we usually sell” walks out the door with every departure.

For the corporate merchant, there is no item-level visibility of stock and shrinkage across stores, making it harder to hit goals.

3. DATA — “What data can be accessed?”

Afresh runs almost entirely on data the retailer already produces--plus inventory counts that it captures in-store. Partners send their existing files to Afresh’s Azure-based cloud; Afresh builds a custom transformation layer for each feed that maps the raw fields to its internal schema, so the grocer doesn’t need to set up new tools or file formats.

Onboarding is a one-time transfer of historical data, then a daily transfer of fresh inputs. Four data sources feed the platform:

POS sales history. Item-level scan data from the checkout. This includes each SKU, how many units sold, date, and price, including promotions and markdowns. It comes from the chain’s existing POS system (Afresh uses the current feed and is POS-agnostic). It’s delivered first as a multi-year historical extract, then is refreshed daily.

Inventory counts. On-hand quantity for each fresh item. Most systems never record this for fresh, so Afresh generates it. Store associates count all the items in a department using the Afresh app--the virtues of which is that is supports several people counting at once to get it done quickly. It also flags a suspected misscan or off-entry as it is typed. Counts feed the inventory model and are validated by corporate through a web portal. Example: a clerk enters 200 cases of strawberries, the app sees that this is 10x the usual count and recent sales, and flags it.

Supply-chain and ordering records. Purchase orders, delivery schedules, shipment fill rates (how much of an order the vendor or distribution center actually shipped), vendor lead times, order guides (client-specific product menus), and the inventory in the data center. These come from the retailer’s ERP and warehouse management system (WMS). Afresh clusters ERP, WMS, and POS records for the same product into one entity, so an item is recognized once and tracked through each form it takes--by the case, by the item, or by weight. Afresh reconciles mismatched SKUs, units of measure, and shelf lives across the systems. Product attributes like pack size are drawn from the retailer’s item master.

External signals. External calendar data (day of week, holidays) and weather, that tell you a holiday weekend, and a warm forecast, might predict a berry spike that the trend from prior weeks of sales alone might miss.

4. DERIVATION — “How do analytics and AI make the data usable?”

Afresh applies machine learning across three layers:

Inventory estimation. The actual counts for produce shift around because spoilage, miscounts, and misscans go unrecorded. The company uses a hidden Markov model (the second appearance of this model in a DataStory newsletter in two weeks!) to estimate a hidden state--true inventory--from observed signals such as sales, deliveries, and periodic manual counts. For example Monday’s count is 10 units and Tuesday’s delivery adds 5, so true Tuesday inventory falls between 0 and 15; the model assigns a probability to each value using forecasted sales, shipment fill rates, and spoilage rates… then recalibrates against the next physical count. The output is an estimated actual, on-hand quantity, with an uncertainty range, for every item, each day.

Demand forecasting. An item-level, store-level, daily prediction of units that will sell, conditioned on price, promotions, perishability, seasonality, and weather. The model returns a forecast plus a confidence range so a volatile item (such as berries for pie before July 4) is handled differently from a steady seller like bagged carrots.

Order optimization. A decision engine converts the inventory estimate and demand forecast into a recommended order quantity per item. It also factors in the cost of a stockout vs. the cost of spoilage, optimizing for margin. It can simulate scenarios--like a different delivery day--before making recommendations. Afresh reports managers accept the recommended order, as-is, 97% of the time.

5. DELIVERY — “How does the workflow improve?”

The product lives in an app the department already uses. It’s scheduled into the morning ordering routine. Where the manager used to walk the aisle and make guesses, then hand-write an order guide, they now open the app to a recommended order, adjust the few things they know better than the model, and submit. It replaces paper order guides, spreadsheets, and end-of-day manual counts.

There’s also a web portal for the corporate team.

6. DECISION — “What decision is enabled?”

The platform affects three decisions:

How much to order. The daily question for every item, answered with a number instead of a guess. Customers report out-of-stocks down as much as 80%.

How much to produce. For in-store kitchens and bakeries, how many units to make so you don’t sell out; or end up dumping all the donuts at closing.

What to buy upstream. At the distribution center, how much to stock and ship so stores are covered. Also avoid warehouse spoilage.

According to the company, clients running on Afresh report food waste down 25% or more, produce operating margins up over 40%, inventory turns up about 7%, and roughly 31% less time spent taking inventory.

7. DESTINATION — “What does success look like?”

Afresh has helped its customers arrive when the fresh department isn’t wilty; it’s full but not overstocked; and less food ends up in the dumpster. This helps customers get all the items from their shopping list; the stores and chains grow revenue and margin; and broadly, we have a healthier food system.

SOURCES

Company, funding, and clients

Afresh, “Afresh Raises $34M to Scale AI Across the Grocery Industry,” PR Newswire, April 21, 2026. https://www.prnewswire.com/news-releases/afresh-raises-34m-to-scale-ai-across-the-grocery-industry--driving-fresher-food-stronger-margins-and-less-waste-302747794.html

Afresh, “Afresh Secures $115 Million in Series B Funding,” PR Newswire, August 4, 2022. https://www.prnewswire.com/news-releases/afresh-secures-115-million-in-series-b-funding-and-rolls-out-its-fresh-food-technology-to-thousands-of-stores-across-the-us-301598519.html

Innovation Endeavors, “Meet Super Evolution drivers Nathan and Matt: Co-founders of Afresh.” https://www.innovationendeavors.com/insights/meet-super-evolution-drivers-nathan-and-matt-co-founders-of-afresh

Afresh, “Sharing Our $115M Series B — and Why We Started Afresh.” https://www.afresh.com/resources/sharing-our-115-million-series-b-and-why-we-started-afresh

Progressive Grocer, “Albertsons Cos. Opts for Afresh’s AI Platform.” https://progressivegrocer.com/albertsons-cos-opts-afreshs-ai-platform

Grocery Dive, “Albertsons to deploy AI-powered food waste-reduction tech chainwide.” https://www.grocerydive.com/news/albertsons-to-deploy-ai-powered-food-waste-reduction-tech-chainwide/617855/

Technology, data, and results

Afresh, “Accelerated time to value and low IT lift: the new standard for deploying fresh technology.” https://www.afresh.com/resources/accelerated-time-to-value-and-low-it-lift-the-new-standard-for-deploying-fresh-technology

Afresh, “Afresh’s Artificial Intelligence” (platform overview). https://www.afresh.com/platform/artificial-intelligence

Afresh, “The Future of Fresh Inventory: How Afresh’s Machine Learning Model is Reimagining Inventory Management.” https://www.afresh.com/resources/the-future-of-fresh-inventory-how-afreshs-machine-learning-model-is-reimagining-inventory-management

Aaron Stern, “Making it Count: Probabilistic Inventory in Grocery,” Afresh Engineering (Medium). https://medium.com/afresh-engineering/making-it-count-52c3b5b459c7

Afresh, “Afresh Adds Inventory Management Solution to AI-Powered Fresh Tech Platform.” https://www.afresh.com/resources/afresh-adds-inventory-management-solution-to-ai-powered-fresh-tech-platform

Supermarket News, “How Afresh is revolutionizing grocery perimeters with AI.” https://www.supermarketnews.com/grocery-technology/how-afresh-is-revolutionizing-grocery-perimeters-with-ai

Food waste, hunger, and grocery margins

ReFED, “Food Waste Data — Causes & Impacts.” https://refed.org/food-waste/the-problem/

ReFED, U.S. Food Waste Report 2025. https://refed.org/downloads/refed-us-food-waste-report-2025.pdf

USDA Economic Research Service, Food Availability (Per Capita) Data System — Food Loss. https://www.ers.usda.gov/data-products/food-availability-per-capita-data-system/food-loss

USDA, Food Waste FAQs. https://www.usda.gov/about-food/food-safety/food-loss-and-waste/food-waste-faqs

Feeding America, “Ten Ways Food Waste Hurts the Environment.” https://www.feedingamerica.org/hunger-blog/ten-ways-food-waste-hurts-the-environment

Food Industry Association (FMI), “Grocery Store Chains Net Profit.” https://www.fmi.org/our-research/food-industry-facts/grocery-store-chains-net-profit

Grocery Dive, “Grocery industry profit margins fall to pre-pandemic levels: FMI.” https://www.grocerydive.com/news/grocery-industry-profit-margins-fall-to-pre-pandemic-levels-fmi/720517/

Competitors and market

CB Insights, “Top Afresh Alternatives, Competitors.” https://www.cbinsights.com/company/afresh-technologies/alternatives-competitors

CB Insights, “Invafresh — Products, Competitors, Financials.” https://www.cbinsights.com/company/invatron-systems

EDITOR’S NOTE

For weird reasons (okay, I was going to write a sci-fi screenplay) I got obsessed with Elon Musk’s company Neuralink and the dystopian notion of drilling a hole into the human skull to implant a device that connects our brains into the internet. (What could go wrong?). But there are aspect of this idea that are indeed utopian: what if we could reduce the friction between our desires, directions, demands… and what the robot world can do for us?

A slightly more comfortable path for linking humans to tech is simply teaching technology to understand our speech—via Automatic Speech Recognition (ASR).

Today’s case study, the services company Avoca--which four weeks ago raised $125MM at a $1B valuation--is a wonderfully humble expression of an ASR use case. Their tools are plugged into local service companies like My Plumber Plus in Northern Virginia, to take customer calls with a persona named “Sarah” and book emergency pipe fixes.

VOICE TO DATA

But how did we get here? And why is booking plumber service visits an invocation of one of the most important AI trends right now?

Like most AI breakthroughs, getting ASR to the point where it enables easy conversation came from breakthroughs in statistics and compute.

Pre-cloud attempts to accomplish this looked like the IBM program in the 1980s… the Tangora System!... an effort to create a voice-to-typewriter tool. (Named for Albert Tangora who--fun fact--holds the world record for accurate typing at 147 words per minute. I imagine Tangora looking and acting like Marty Mauser in Marty Supreme.) Use the IBM Tangora to dictate to your typewriter!

This system combined linguistics (breaking a chunk of speech into “triphones,” the smallest packet of human speech--three phonemes, or sounds--like C-A-T) with predictive stats that guessed the larger word. The program leader, Fred Jellinek, was famously said to quip, Every time I fire a linguist, the error rate goes down. Kind of mean--we love our linguists!--but it was a signal of the approach was to come.

This moment arrived (after a long journey) in September 2022, when OpenAI released a model called Whisper that--in a post-cloud, post-GPU world--could be, and was, trained on 680,000 hours of audio in 99 languages. That’s about 78 years of talk.

Adding speed came next. Early voice assistants felt laggy because they worked in a relay: one model turned the speech into text; then the text was handed to an LLM to decide what to say; a third turned the answer back into sound. In May 2024, OpenAI’s GPT-4o folded the ASR AI directly into a multi-modal system: a model that hears and responds in audio, without translating back and forth to text.

So, while Avoca’s competitive set, narrowly defined, includes Sameday, ServiceAgent.ai, Netic (also AI phone-answering for contractors), Rilla (records and coaches field-sales), Hatch (automates texting), and ServiceTitan (the software most large contractors already run their business on)… we can place Avoca in a broader context of the bridge created between human and technology via speech.

This is having an impact on medicine--we mentioned company Abridge which turns doctor-patient visits into EHR notes--but also on Waymo and autonomous driving (you need to talk to your robot car), and just the general speed of turning ideas, commands, meetings… into data, to put it to work.

Here is how Avoca chases the problem, in our 7Ds framework:

1. Demand. The customer service reps who answer the phones; and the service company owners.

2. Dilemma. Calls go unanswered, which represent missed revenue.

3. Data. The contractor’s own sales training information and scripts; their own phone calls recordings; their scheduling records; their lead sources.

4. Derivation. A low-latency voice model trained on trades use-cases, plus scoring models that grade every call.

5. Delivery. The AI answers the phone and writes the job onto the existing ServiceTitan board.

6. Decision. Create service appointments and follow ups; also enable marketing effectiveness analysis.

7. Destination. Booking rates that match the best human reps, and recovered lost revenue in missed calls.

COMPANY OVERVIEW

Avoca builds AI agents that work the phones for service businesses. The agents answer inbound calls, texts, and web chats around the clock; book jobs into the company’s scheduling software; and run outbound campaigns to fill open service slots. The company started in heating and cooling, plumbing, and electrical, and is extending into roofing, restoration, auto, moving, and junk removal.

Date of Founding: 2023. Headquartered in New York City, with an office in Santa Barbara.

Co-CEOs. Tyson Chen and Apurva Shrivastava met at an MIT poker night. Chen studied computer science at MIT and was a product manager at the self-driving startup Nuro. Shrivastava was also an MIT computer scientist and built products at Apple and the engineering tool Retool.

Founding Story. They began by building software to catch missed calls for restaurants. Wandering a restaurant conference in Texas, they were approached by Rescue Air, a Dallas heating and air company. They spent an afternoon with the team and saw the same missed-call problem was worth far more in the trades. For three months in 2023 they built a product for Rescue Air, which then introduced them to their first handful of customers. The company is named for the Avoca, a river in County Wicklow, Ireland, formed where two rivers meet.

Total Investment Raised. $125 million Series B led by Meritech and General Catalyst, announced in April 2026 at a $1 billion valuation.

What the Company Does. When a call comes in, Avoca’s agent answers in a few seconds, in natural speech. It can give a price range, handle an objection, schedule a visit, and flag a true emergency. Second, there’s an outreach function: it follows up on estimates that never closed, re-engages old customers, and answers web and ad leads within seconds. A third element listens to the company’s human reps and scores their calls. Avoca holds a SOC 2 Type 2 certification (for service company security assurance) and runs a human-in-the-loop program in which staff review and correct the AI’s work.

How the Company Makes Money. Avoca sells software subscriptions to service businesses. Pricing is not public. The company reached eight figures in annual recurring revenue in 2025 and says it is on track to book $1 billion in jobs for its customers this year.

Top Named Clients. More than 800 businesses, including Turnpoint, Sila Services, Goettl, and 1-800-GOT-JUNK?. Avoca has partnerships with ServiceTitan, the trade-franchise network Nexstar, and the payments company Clover.

THE 7Ds

1. DEMAND

Who is their user?

Customer Service Representative

The CSR sits in the front office and answers the phone. The job pays modestly, turns over often, and runs busiest when the weather is worst and things are frantic.

The Owner

The owner carries the cost of every missed call and every empty slot on the schedule. Most owners staff for an average day, which leaves them short on the days with heaviest demand.

2. DILEMMA

What makes the job hard?

Demand is spiky

A heat wave or a cold snap can triple call volume in a day, and the front office that was right-sized in April is buried in July. Calls come at night, on weekends, and during the storms that cause the very emergencies people are calling about.

Supply (of CSRs) is limited

The phone is slow, manual work; a rep can hold only one conversation at a time. A full-time receptionist runs $35,000 to $45,000 a year, and covering nights and weekends takes several of them.

Missed revenue

When the lines are full, the overflow rolls to voicemail, and the caller—who has a flooded basement and no patience—dials the next company on the list.

3. DATA

What data is used?

Avoca runs on data the contractor already owns--more latent data.

Scripts and sales materials

A body of data is assembled for each customer when Avoca is installed. It includes the business’s own scripts, greeting, and tone of voice; its price list, with the $89 tune-up… the trip charge… the financing offers. Its hours, service area, and the jobs it does and does not offer. Its rules for scheduling and for routing a real emergency. Avoca’s implementation team turns this into a custom knowledge base so the agent can answer service-specific questions the way a trained rep would.

Their own calls

The contractor’s own recorded calls feed in as well—the way its best reps handle a price objection becomes material that the AI and the company’s other reps both learn from. Every inbound call is recorded and transcribed, and from each transcript the system pulls a structured record: who called, the problem in their own words (I like this), any equipment they mention, how urgent they sound, whether it looks like a safety risk, and what happened—for instance an appointment booked, or routed to a technician. A call about an air conditioner “making a grinding noise” with a “burning smell” might come out the other side tagged as Carrier unit, high priority, safety risk, with a one-line summary for a rep or a manager to scan

Scheduling system

That call is matched against the scheduling system, usually ServiceTitan or Housecall Pro, which holds the customer’s history, past jobs, open estimates, the equipment on file, membership status, lifetime value, and the live picture of which technicians are free.

When the phone is answered, the system knows whether the caller is a gold-plan member worth $4,000 or a first-timer who just clicked a Google ad.

Lead sources

Web forms, Facebook, Yelp, Angi, and Google’s Local Services Ads—each one a signal that a new prospect has raised a hand somewhere other than the phone.

Most of this lived in scattered systems and in the heads of a few veteran reps before Avoca pulled it into one place.

4. DERIVATION

How does the AI make the data usable?

Voice agent

It pairs a speech model that answers fast enough to feel live with a language model tuned to the vocabulary of the trades, and it is configured for each customer from the knowledge base above—the scripts, prices, and rules that tell it how this particular business runs a call. The agent improves over time. As Tyson Chen puts it, a human rep might make the same mistake ten times, while the agent makes it once, because each corrected call updates the system for every call after it.

Scoring

Models listen to every call, human or AI, and grade it on things like how well the rep handled an objection, how much empathy they showed, and how many filler words they used.

Outbound

Logic that decides who to call back, when, and how persistently, and books new work for technicians who are free. People review samples of the agent’s calls and correct them, and those corrections feed back in. Avoca describes a loop in which its human reps train the AI… and the AI, in turn, trains the reps.

5. DELIVERY

How does the workflow change?

• Avoca sits on top of the contractor’s existing phone line and scheduling software. The AI answers the call, and the booked job appears on the same ServiceTitan dispatch board the team already works from, so nothing new has to be learned.

• Front-office managers get every call scored, instead of the handful they used to spot-check by hand.

• Owners get a dashboard that traces marketing tactics to the invoices.

6. DECISION

What decisions does it enable?

• The owner can decide where to spend the marketing budget, because the path from ad to paid job is visible.

• They can fill slow days by turning on outbound follow-up rather than hiring.

• They can stop paying for after-hours coverage that can be handled by robots with endless capacity.

• The front-office manager can decide who needs coaching, and on what, from the full record instead of a sample of a few calls.

7. DESTINATION

What does success look like?

The owners have reduced friction between themselves and their customers, and have complete capture of those interactions to mine for improved service.

8. WHY IT MATTERS TO YOU

You may never sell a furnace, but Avoca’s playbook transfers to almost any business that talks to its customers.

• Your conversations are unused data. Every call, meeting, and service visit is speech you could capture, structure, and act on. Start by asking what gets said out loud in your business that no one ever writes down.

• Reduce friction. Create data with spoken interactions, and make your customer’s path to working with you faster, with ASR.

• Cut the step where people have make records of work instead of doing it. Avoca books the job without anyone typing into the CRM, the same way ambient scribes pulled doctors out of endless notes taking follow-ups.

• Play to the niche. Avoca won by serving the trades that Silicon Valley skipped. The boring, underdigitized part of a market has less competition and more room to grow than the obvious one.

SOURCES

VOICE TECHNOLOGY & FRICTION

OpenAI, “Introducing Whisper,” September 2022. https://openai.com/index/whisper/

“Whisper (speech recognition system),” Wikipedia. https://en.wikipedia.org/wiki/Whisper_(speech_recognition_system)

Baevski et al., “wav2vec 2.0: A Framework for Self-Supervised Learning of Speech Representations,” NeurIPS 2020. https://proceedings.neurips.cc/paper/2020/file/92d1e1eb1cd6f9fba3227870bb6d7f07-Paper.pdf

OpenAI, “Hello GPT-4o,” May 13, 2024. https://openai.com/index/hello-gpt-4o/

American Medical Association, “Doctors work fewer hours, but the EHR still follows them home.” https://www.ama-assn.org/practice-management/physician-health/doctors-work-fewer-hours-ehr-still-follows-them-home

“Pajama Time: Working After Work in the Electronic Health Record,” PMC. https://pmc.ncbi.nlm.nih.gov/articles/PMC6712097/

Tebra, “How documentation became the top cause of physician burnout.” https://www.tebra.com/theintake/ehr-emr/how-documentation-became-top-cause-of-physician-burnout

VOICE AI LANDSCAPE

Nextiva, “50+ Conversational AI Statistics for 2026.” https://www.nextiva.com/blog/conversational-ai-statistics.html

Ringly, “47 voice AI statistics for 2026: market size, growth, and trends.” https://www.ringly.io/blog/voice-ai-statistics-2026

Fortune Business Insights, “Conversational AI Market Size, Share | 2026-2034.” https://www.fortunebusinessinsights.com/conversational-ai-market-109850

American Hospital Association, “6 Health Systems Enhancing Care Delivery with Ambient AI Scribes,” April 14, 2026. https://www.aha.org/aha-center-health-innovation-market-scan/2026-04-14-6-health-systems-enhancing-care-delivery-ambient-ai-scribes

Restaurant Dive, “Wendy’s to deploy drive-thru AI to over 500 restaurants this year,” 2025. https://www.restaurantdive.com/news/wendys-deploy-digital-menu-boards-drive-thru-ai-500-restaurants-2025/746977/

BusinessWire, “CES 2025: SoundHound AI Debuts Its First Ever In-Vehicle Voice Assistant,” January 7, 2025. https://www.businesswire.com/news/home/20250107847181/en/CES-2025-SoundHound-AI-Debuts-Its-First-Ever-In-Vehicle-Voice-Assistant-With-On-The-Go-Food-Ordering

COMPANY & FUNDING

Avoca, “Why We Are Building,” April 27, 2026. https://www.avoca.ai/blog/avoca-raises-125m-series-b-1b-valuation

Avoca press release, “Avoca Raises $125M+ at $1B Valuation,” PR Newswire, April 27, 2026. https://www.prnewswire.com/news-releases/avoca-raises-125m-at-1b-valuation-to-power-americas-services-economy-with-ai-302753962.html

Allie Garfinkle, “How a chance encounter in Texas sparked a $1 billion Kleiner Perkins-backed AI startup,” Fortune, April 27, 2026. https://fortune.com/2026/04/27/avoca-ai-agents-missed-calls-hvac-plumbing-roofing-kleiner-perkins-chen-shrivastava-braswell/

Avoca company website. https://www.avoca.ai/

Y Combinator, Avoca company page. https://www.ycombinator.com/companies/avoca

Tyson Chen, LinkedIn. https://www.linkedin.com/in/tysonchen17/

Apurva Shrivastava, LinkedIn. https://www.linkedin.com/in/apurva-shrivastava-138b19a0/

PRODUCT, MARKET & COMPETITORS

Avoca, AI CSR for HVAC. https://www.avoca.ai/industries/hvac

Avoca, Coach (real-time call scoring). https://www.avoca.ai/coach

Avoca on the ServiceTitan marketplace. https://marketplace.servicetitan.com/partner/Avoca-AI

Housecall Pro, “The Hidden Costs of Missed Calls for Home Service Business Owners.” https://www.housecallpro.com/resources/missed-calls/

Signpost, “How Much Revenue Do Home Services Businesses Lose from Missed Calls?” https://www.signpost.com/blog/how-much-revenue-do-home-services-businesses-lose-from-missed-calls/

Sameday, AI virtual receptionist for ServiceTitan. https://www.gosameday.com/

EDITOR’S NOTE

As the former head of data for large ad firms, I’ve been on countless client calls with sellers who… I mean, some of them… while perky, and great at securing a bagel platter… never seemed to have it occur them to do deep research about their clients. This pained me. It was like: these clients have such interesting businesses. You’re really going to walk in there and ask for money, armed only with the knowledge of what they bought from you last quarter? I should be more sympathetic, I know: every Account Exec carries a big book, they’re under pressure, they have all that nasty Salesforce logging to do. Yet it always felt like: this could be done better. Maybe a lot better. And when I think back to the scale of this operation… scores of sellers in so many markets… it feels like a huge gap.

The company Actively has what may be an elegant solution: tying together the data that already exists in a B2B organization into an intelligence tool for B2B salespeople, to make them more effective. Not replace them with a bot, but give them cyborg powers to expand and extend what they can accomplish.*

(*This “it’s not X it’s Y” sentence was written by a human.)

A startup I met recently claimed to have 50,000 AI agents--fifty thousand!--scurrying around the web like ants gathering data. Actively raised $45 million last month so that in your sales org, every account can get its own AI agent.

What does that agent do? It supports one account around the clock. The agent reads the call notes from past reps; scans the web for buying triggers—a funding round, a job change, a product launch. It decides who the AE (or sales development rep) should call, and why. It drafts the email. Then it hands the work to a human seller in the morning.

A couple things I like about this, as a case study for you:

1 - The data… the account history, the stakeholder map of the client, the Gong call transcripts… already belongs to the company. The company is just not getting value from it. Yet. It’s what we call, latent data. So Actively is sweeping in and creating value from pre-existing assets.

2 - This is also an important example of the fine line startups must play in the world of AI business models, especially ones that involve activating a client’s own data. What’s to stop the client from crafting its own homegrown tool that does this exact thing? (Some B2B sales companies are doing just that, with internal IT or product teams.) This is a live test! How fast can Actively turn the innovation crank handle, develop more features and more scope, and bring in exclusive data, to stay ahead of competitors, client DIY, and vertical integration from Big Tech?

Why are they chasing it, then? Actively’s argument is that Sales is the most expensive function in a B2B company: thirty to forty cents of every dollar of revenue at a B2B company goes back into go-to-market. Sales engagement tools are tricky to size as a space—CRM is $70B+ depending on estimates—but the vision seems to be that there’s an opportunity for somebody to be the Cursor (coders) or Harvey (lawyers) for B2B sellers. Cyborg tools for knowledge workers.

Other players have made the same observation. 11x.ai sells “Alice,” an AI account exec (reported $350 million valuation). Artisan runs “Ava” and put up “Stop Hiring Humans” billboards along SF highways. Qualified has “Piper.” Clay, Apollo, Outreach, Amplemarket, and Regie.ai compete in adjacent slices—data enrichment, outbound automation, and dialers.

So what’s their angle?

Actively’s 7Ds of Data Innovation:

• Demand. Sales reps and their leaders—SDRs, AEs, CROs at high-growth B2B companies

• Dilemma. Too many accounts, too little context, too much manual research (or none at all sometimes).

• Data. The company’s own CRM records, call transcripts, product usage, intent signals, and whatever is on the open web, combined into a single data set to train the agent.

• Derivation. Custom reasoning models—company-specific fine-tunes built on top of rented LLMs.

• Delivery. A morning hitlist for each rep, embedded in Salesforce, Salesloft, or Outreach.

• Decision. Which accounts to call today and what to say.

• Destination. Higher win rates, less burnout, more pipeline per head

COMPANY OVERVIEW

Actively builds AI agents for B2B sales teams. The product assigns one agent to every account in the customer’s market. The agents run in parallel. Each one builds a memory of what is happening inside its one account. They recommend who to call, they draft the message, and then queue up the work for human sellers and managers.

Date of Founding: 2022.

CEO Profile

Mihir Garimella is in his early twenties. He grew up in Pittsburgh and went to Stanford for both undergrad and graduate work in artificial intelligence. At fourteen he won the top prize in his age category at the Google Science Fair for “Flybot,” a tiny emergency-response drone inspired by how fruit flies dodge swatters (!). CNN profiled him as one of “tomorrow’s heroes.” Co-Founder Anshul Gupta is the other founder and runs the revenue side of the company. He has a master’s in computer science (AI) from Stanford and was a research peer of Garimella’s.

Founding Story

The two met at Stanford. They watched ChatGPT change the AI world in 2022 and 2023 and started asking: which corporate function spends the most money and has the worst AI tooling? They chose sales. A seller with 200 accounts can only seriously focus on a handful in any given week. The other 195 sit dormant. Garimella and Gupta thought reasoning-based AI agents could change that.

Total Investment Raised to Date: $68 million.

Seed: $5M, led by First Round Capital.

Series A: $17.5M in April 2025, led by Bain Capital Ventures.

Series B: $45M in April 2026, co-led by TCV and First Harmonic. Bain Capital Ventures, First Round, and Alkeon joined. $250M post-money valuation.

What the Company Does

For each customer, Actively builds custom AI agents that read everything the customer knows about each account—Salesforce records, Gong call recordings, Outreach emails, marketing-automation behavior, intent data from tools like 6sense, plus public web signals like a new funding round. The agents reason over that material and decide which accounts are worth calling today, who to contact, why now, and what to say. The output lands in the rep’s existing workflow.

How the Company Makes Money

Annual enterprise subscriptions. Pricing is not public. Third-party comparisons of competing AI sales platforms put enterprise contracts in this segment between $30,000 a year and the high six figures, depending on seat count and integrations. Actively targets the high end—companies with dozens of SDRs, hundreds of AEs, and complex sales.

Top Named Clients

Company-sourced: Ramp (corporate cards), Samsara (connected operations hardware), Ironclad (contract management software), Greenhouse (recruiting software), Justworks (HR for small businesses), Verkada (security cameras), Attentive (SMS marketing), Navan (corporate travel). Samsara is the biggest deployment: 1,000 people across sales, account development, RevOps, and customer success.

1. DEMAND

Who is their user?

Priority User 1: The Sales Development Rep (SDR)

The SDR is the entry-level role in B2B sales. Twenty-something. First or second job out of school. Sits in a noisy bullpen and works the phones. Compensation is mostly tied to “meetings booked”—convincing a stranger at a target company to sit on a 30-minute call with an account executive.

The classic SDR day:

- 8am. Dig through Salesforce notes and LinkedIn for 50 companies to email.

- Late morning. Fire off 100 emails.

- Afternoon. Make 50 cold calls.

- Evening. Log everything in Salesforce.

I’ve never worked in an org with this role, but the general take seems to be: it’s a burnout role because you churn and churn and don’t see the results of your work, and don’t see a goal, or an end in sight.

Priority User 2: The Account Executive (AE)

The AE is the next rung up. They own the revenue target. They take the meetings the SDR booked, run discovery calls, write the proposal, negotiate, and close. At fast-growing B2B companies a full-cycle AE is also expected to source pipeline too.

Priority User 3: The Chief Revenue Officer (CRO)

A 2025 GTM benchmarks report found that 63% of CROs have little or no confidence in how their Ideal Customer Profile is defined. Yikes. (We see that too, at the DataStory Group.) The executive on the hook for hundreds of millions in revenue often does not trust the definition of “who do we sell to” coming out of their own marketing team.

2. DILEMMA

What makes achieving that goal difficult?

Obstacles for the SDR

A modern SDR is given 200 to 500 accounts in a territory. Cold-call response rates run well under 5%. Even the best SDRs convert only a small fraction of their outreach to booked meetings. High volume plus high rejection equals exhaustion and a feeling of pointlessness.

Obstacles for the AE

The AE has too few accounts but too much context to keep straight inside each one. A typical enterprise deal can run six to twelve months, requiring input from ten stakeholders. When an AE inherits an account from a departed colleague, that context disappears.

Obstacles for the CRO

The CRO is responsible for fixing all the dynamics above, which can make them feel more like a data engineer—SF, Gong, notes, metrics, reports, KPIs—than the sales strategy leader they signed up to be.

3. DATA

What data can be accessed to achieve the goal?

Actively does not own a proprietary outside dataset. The data lives inside the client’s own tools. There are four main types.

Data Type 1: CRM Records.

The CRM is the system of record for the sales team. It stores every account, contact, deal, interaction the company logs.

Data Type 2: Conversation Data.

Sales calls have been getting recorded and transcribed for about a decade. Gong is the dominant tool; there’s also Chorus, Clari, and Avoma. Every customer call becomes a transcript which is a potential goldmine where the client expresses their needs, competitor stresses, and budget issues.

Data Type 3: Intent and Engagement Signals.

These are signals from third-party data providers that try to predict when a company is about to buy. The biggest is 6sense. They aggregate anonymous web behavior—what blogs are being read from which corporate IP ranges, what comparison pages are getting traffic—to flag accounts that look “in market.” Marketing-automation tools like Marketo and HubSpot add: who filled out a form, attended a webinar, downloaded a whitepaper. Also there’s the public web—funding rounds, big client wins, executive hires, layoffs, mergers, product launches, press coverage.

Data Type 4: Internal Enablement Material.

Most B2B companies keep a library of internal sales materials—how to position against each competitor, talk tracks for each persona, case studies, pricing playbooks. This usually lives in Notion, Confluence, or Highspot.

4. DERIVATION

How do analytics and AI make the data usable?

Layer 1: Custom Reasoning Models.

For each client, the team builds a custom model that combines off-the-shelf reasoning models from OpenAI and Anthropic, trained on the data above.

Layer 2: Per-Account Memory.

The second layer is the context for the agent. Actively assigns one persistent agent per account. That agent’s memory updates whenever something new happens—a new call recorded, an email opened, a free-trial signup. The memory persists across AE changes, and territory reassignments.

Layer 3: Multi-Agent Execution.

Once the memory and the model are in place, agents move into execution: research, prioritization, contact selection, message drafting, and finally a hand-off to a human seller or to an automated sequence in Salesloft or Outreach.

5. DELIVERY

How does the workflow improve for the user?

Actively’s interface sits inside the tools sellers use—Salesforce, Outreach, Salesloft, Gong—and inside a web app called the Agent Inbox. SDRs get a queue of priority prospects.

The AE gets a daily meeting-prep digest before calls: account summary, recent news, suggested talking points, last-touch history.

The CRO gets a dashboard: which accounts are stalling, win rate trends.

6. DECISION

What decision is enabled by using the data?

Decision 1: Which accounts to work on today

In a 500-account territory, an SDR can only seriously work twenty or thirty in a day. Picking the right twenty is a high value decision in the SDR’s workflow.

Decision 2: Who at the account to contact?

Inside any large account there are dozens of plausible buyers. Those agents help choose based on who has been mentioned in past calls, who responded to past marketing emails, who recently changed jobs--then recommends a primary contact.

Decision 3: How to allocate sales investment?

For the CRO, the question is “can my existing reps cover more accounts with the same effort?” to deliver more pipeline with flat headcount. A classic optimization challenge--match your supply to your available demand.

7. DESTINATION

What does success look like?

Sales is a great area for data-driven insights--because you can test them. Give a seller a toolkit conferring cyborg superpowers, and you should be able to see the results show up immediately: in higher win rates, more senior relationships, bigger budgets per relationship. In short, higher productivity.

And, one would think, higher happiness in the sales team. If you can show up at work and just be an athlete… not carry the Gatorade and launder the uniforms… and not only that, have a robot voice in your head telling you where to run on the field… why am I talking in these extended sports metaphors?... that feels like a major win.

8. SOURCES

COMPANY SOURCES

1. Actively, “Intelligence-Led Revenue,” April 20, 2026. https://www.actively.ai/blog/intelligence-led-revenue

2. Actively company website. https://www.actively.ai/

3. Actively, Customers page. https://www.actively.ai/customers

4. Actively, Samsara customer story. https://www.actively.ai/customers/samsara

5. Actively, Justworks customer story. https://www.actively.ai/customers/justworks

6. Actively, Ironclad customer story. https://www.actively.ai/customers/ironclad

7. Actively, SDR solutions page. https://www.actively.ai/solutions/sdr

8. Mihir Garimella, personal website. https://mihir.garimella.io/

9. Mihir Garimella LinkedIn. https://www.linkedin.com/in/mgarimella/

10. Anshul Gupta LinkedIn. https://www.linkedin.com/in/agupta24/

FUNDING & THIRD-PARTY PRESS

11. BusinessWire, “Actively Raises $45M Series B,” April 28, 2026. https://www.businesswire.com/news/home/20260428810008/en/Actively-Raises-$45M-Series-B-to-Scale-Intelligence-Led-Revenue-Platform

12. Pulse 2.0 (Amit Chowdhry), Series B coverage, April 29, 2026. https://pulse2.com/actively-45-million-raised-to-scale-intelligence-led-revenue-platform-powered-by-per-account-ai-agents/

13. Sofia Chierchio, “Meet The $250 Million Startup Challenging Salesforce With AI Agents,” Forbes, April 28, 2026. https://www.forbes.com/sites/sofiachierchio/2026/04/28/meet-the-250-million-startup-challenging-salesforce-with-ai-agents/

14. Charles Rollet, “Actively AI raises $22.5M to offer sales ‘superintelligence,’ says AI SDRs failed,” TechCrunch, April 2, 2025. https://techcrunch.com/2025/04/02/actively-ai-raises-22-5m-to-offer-sales-superintelligence-says-ai-sdrs-failed/

15. BusinessWire, Series A coverage, April 1, 2025. https://www.businesswire.com/news/home/20250401198986/en/Actively-AI-Raises-$22.5M-to-Maximize-Revenue-for-Top-Sales-Teams-with-GTM-Superintelligence

16. Bain Capital Ventures, “Actively AI is Launching a New Era of GTM Superintelligence,” April 2, 2025. https://baincapitalventures.com/insight/actively-ai-is-launching-a-new-era-of-gtm-superintelligence/

FOUNDER PROFILES

17. CNN Business, “Mihir Garimella is making drones that go where humans can’t,” February 8, 2018. https://www.cnn.com/2018/02/08/tech/mihir-garimella-drones-tomorrows-hero/index.html

18. Singularity Hub profile of Mihir Garimella, August 25, 2017. https://singularityhub.com/2017/08/25/this-inspiring-teenager-wants-to-save-lives-with-his-flying-robots/

19. The GTM Engineer (Substack), interview with Anshul Gupta. https://thegtmengineer.substack.com/p/the-account-as-a-unit-and-finding

20. Orum Podcast, “Anshul Gupta on AI and tryin’ to triage.” https://www.orum.com/podcasts/anshul-gupta

COMPETITORS & AI-SDR LANDSCAPE

21. TechCrunch, “AI SDR startups are booming, so why are VCs wary?” December 26, 2024. https://techcrunch.com/2024/12/26/ai-sdr-startups-are-booming-so-why-are-vcs-wary/

22. TechCrunch, “A16z- and Benchmark-backed 11x has been claiming customers it doesn’t have,” March 24, 2025. https://techcrunch.com/2025/03/24/a16z-and-benchmark-backed-11x-has-been-claiming-customers-it-doesnt-have/

23. Artisan AI “Stop Hiring Humans” campaign. https://www.artisan.co/blog/stop-hiring-humans

24. Qualified AI SDR (Piper) product page. https://www.qualified.com/ai-sdr

25. 11x.ai company website. https://www.11x.ai/

MARKET SIZING & CONTEXT

26. Market Growth Reports, Sales Engagement Platform Market Size Report. https://www.marketgrowthreports.com/market-reports/sales-engagement-platform-market-118278

27. Lusha, “Why Outbound Sales Development Reps Burn Out in 15 Months.” https://www.lusha.com/blog/outbound-sales-rep-burnout-causes/

28. Sales Gravy, “Why Your Best SDRs Burn Out by Month Four.” https://salesgravy.com/why-your-best-sdrs-burn-out-by-month-four-and-how-to-stop-it/

29. Orum, “How Sales Leaders Can Fight SDR Burnout.” https://www.orum.com/blog/fight-sdr-burnout

30. Fullcast, “RevOps for CROs.” https://www.fullcast.com/content/revops-for-cros/

COMPANY PROFILES

31. PitchBook, Actively AI 2026 Company Profile. https://pitchbook.com/profiles/company/172150-84

32. Actively LinkedIn company page. https://www.linkedin.com/company/actively-ai/

OVERVIEW

April 2026 was the heaviest month of data and AI investment we have tracked yet. $85 billion poured into the sector, with the big numbers pumped by established whales like Intel spending $14 billion on a stake in its Ireland fab; and the compute splurge continuing with $25 billion raised in debt to keep building data centers.

BUT AS EVER WHAT WE CARE ABOUT at The DataStory is the other slice: the 40 companies who raised more than $30MM and are doing something specific with data or AI that an executive can actually learn from. They added up to about $21.2 billion.

And it is a whopper of a month:

THE TOP 10

Among the use-case companies we profile, the biggest individual financing actions this month:

1. Project Prometheus: $10 Billion (Venture)

2. ServiceNow: $4.000 Billion (Post-IPO Debt)

3. Upstart: $1.250 Billion (Forward-flow loan-purchase commitment)

4. Ineffable Intelligence: $1.1 Billion (Seed)

5. Pagaya: $1.1 Billion

6. Stats Perform: $750.0 Million (Debt + funding round, combined)

7. IKS Health: $670.0 Million (Acquisition debt)

8. Recursive Superintelligence: $500.0 Million (Series A)

9. K1x: $175.0 Million (Private Equity)

10. Rogo: $160.0 Million (Series D)

11. Aidoc: $150.0 Million (Series E)

TRENDS FOR MAY 2026

1. Fundamental shift in AI from training to self-learning

Something fundamental happened this month. Some of the biggest brains in AI, including the inventor of Alpha Go, are now betting on artificial intelligence that is not trained on data, in the conventional sense. They are self-teaching, observing our environment, experimenting, trying, improving. One way to connect these bets is through the concept of world models (if you haven’t read our primer on world models, now is a good time.) David Silver--the Alpha Go lead researcher, formerly of DeepMind--raised a $1.1B seed for Ineffable Intelligence to build “superlearners” that discover knowledge through experience--their own. It means he wants their models to poke at, test, trial and error their way through, our planet, our human experience--and get better at it than we are. Why? Per Wired, his preferred thought experiment is: Imagine going back in time and releasing a large language model in a world that believed the world was flat. Without being able to interact with the real world, the system, he says, would remain an avid flat-earther, even if it continued to improve its own code. This trend is echoed in the $10B bet from Jeff Bezos on the humbly-titled “Project Prometheus” focusing on physical AI--the interaction of models with real-world spaces. Then another big round at Recursive Superintelligence, a pre-product AI company developing--per Economic World--“self-teaching AI, a framework where models can improve their own reasoning capabilities without constant human intervention or new external data…effectively mimicking the way a human expert might practice a skill in isolation.” This is on top of the launches of Yann LeCun and Fei Fei Li’s (separate) world model companies in the last few months. I believe this is the beginning of AI Phase 2. Why? Models will no longer rely on a narrowly framed user premise, or the frame of training data assigned to it. Those previous models were like pets: trained, set up with a defined domain and acceptable choices. These new models are designed to take action on their own, test their understanding of reality, adjust, refine, repeat. To be truly autonomous, in other words. We are setting up our models on their sweet little toddler feet to bumble their way in the world, crash, learn. Only with this tech, they will be graduating from MIT… about…now… and will start changing the world by 2027. Yes, we are moving from AI 1.0, which resemble Pets, to AI 2.,0, which is Toddler AI.

2. Satellites create a “data-only replica of our home planet”

Five satellite or Earth-observation companies took funding in April. And you know I have a soft spot for satellite data because of its crazy American roots--in the USGS (US Geological Survey) dating back to the 1800s, which mapped our world starting with the Colorado River and the Grand Canyon, and later yielded Landsat satellites. Now, investors are funding satellites which do not take pictures taken from space, but instead generate new types of space data like radio frequency mapping which detect motion--in objects, and the atmosphere--which is so much richer, according to the company Spire (rhymes with spy-er; $70MM raise), that it means “we have reached a critical milestone in human history where it is possible to create a data-only replica of our home planet.” Yunyao Yunyao Aerospace ($73MM) does the same, for China. Sidus Space ($58.5MM) is focused on defense applications. Xoople ($130MM) is less about data collection and more about organizing space data for AI training and applications.

3. Lending goes agentic

This was the month that financial services / lending went cyborg, following the Legal profession. Upstart and Pagaya together moved $2.65 billion in AI-underwritten lending capacity. Neither is venture money. Both are something more interesting: institutional capital underwriting AI-underwritten consumer credit at AAA ratings. That is a vote of confidence from the most conservative buyers in finance that the models work. Add Rogo ($160MM for AI research for investment banks), K1x ($175MM for automating private-market tax data), Wealth.com ($65MM for AI estate planning), Strala ($51MM for end-to-end claims processing), and Acclaim ($40MM for voice AI in regulated banking) and you have one industry being rewritten very fast across underwriting, research, ops, and customer service.

4. Voice AI finds its voice

Three meaningful raises this month for AI customer service--Netomi ($110MM), Avoca ($125MM), Acclaim AI ($40MM)--and what they share is that none of them lead with text, but with voice. Avoca puts voice agents into HVAC and plumbing call centers and reportedly handles 24/7 inbound on behalf of trade businesses. Acclaim built specifically for banking and fintech, where every word said to a customer is potentially a regulated communication. Netomi is the broadest play, handling voice, plus chat, plus social across enterprise communications with customers. The big change here is that the models are now good enough at conversation, hand-off, and compliance that regulated industries are willing to deploy them on inbound. Again, a human resources implication--what happens to the humans who work in call centers?

SECTOR & COMPANY PROFILES

AI, Information Infrastructure & Tools--$14.95B

See comments above: Fundamental shift in AI from training to self-learning

Project Prometheus

Project Prometheus is the AI lab Jeff Bezos and Vikram Bajaj launched out of stealth in November 2025--Bezos’s first operational role since stepping down from Amazon in 2021. The April round closed at $10 billion at a $38 billion valuation--all without selling anything. The focus is physical AI: models trained on real-world experimental data, robotics interactions, and engineering data rather than on text. Their customer targets are manufacturing, aerospace, drug discovery, and industrial design. Bajaj is the operating co-CEO with a physical chemistry PhD from MIT, ran the Wing and Waymo programs at Google X, and co-founded AI drug-discovery lab Xaira Therapeutics. The wildcard is that Bezos is separately reportedly raising up to $100 billion for a holding company that would acquire stakes in industrial firms--engineering, design, manufacturing--and it’s reasonable to infer that these companies would leverage Prometheus--kind of like the new AWS and cloud computing, only for Physical AI.

Venture Round, $10.0B, April 23, 2026.

ServiceNow

ServiceNow borrowed $4B under a short-term unsecured term loan to help fund its Armis acquisition, later moving to refinance with a bond sale to continue retrofitting its 20-year-old workflow platform with AI agents.

Post-IPO Debt, $4.0B, April 17, 2026.

Recursive Superintelligence

Recursive Superintelligence is a four-month-old London-incorporated AI lab co-founded by Richard Socher (former Salesforce chief scientist) and Tim Rocktäschel (former principal scientist at Google DeepMind). Their mission is “recursive” self-improvement--automating the AI development pipeline (data selection, training, post-training, and research direction) so the models build themselves without a human in the loop. Socher calls this “the third and perhaps final stage of neural networks.”

Series A, $500MM, April 18, 2026.

Hightouch

Hightouch is the data activation company. The focus is “agentic marketing”: AI agents that read the CRM data layer, decide what to do, and execute across channels--for customers like PetSmart and Grammarly running personalization across millions of loyalty members. The big insight from the Hightouch trajectory is that data activation, identity resolution, and decisioning can be the new brain of customer marketing.

Series D, $150MM, April 29, 2026.

Omni

Omni provides an AI analytics platform with a semantic data layer that connects enterprise data warehouses to BI tools and AI applications, betting that the missing piece in most companies is consistency across queries and AI prompts.

Series C, $120MM, April 23, 2026.

BLP Digital

Swiss company BLP Digital provides agentic AI automation for enterprise resource planning.

Secondary Market, $50MM, April 22, 2026.

Yongyue Intelligence

Chinese AI company Yongyue Intelligence focuses on intelligent technology solutions and automation systems for the domestic enterprise market.

Series A, $50MM, April 22, 2026.

Noon

Noon creates AI tools for product design, to close the gap between design and the engineers.

Venture, $44MM, April 1, 2026.

Resolve.ai

Resolve.ai is an AI platform that autonomously investigates production incidents without paging a human.

Series A, $40MM, April 16, 2026.

Financial Services: The Agentic Credit Score--$3.14B

See “Lending goes agentic” above.

Upstart

Upstart is an AI lending marketplace that partnered with banks and credit unions to use machine learning instead of FICO scores for consumer loans. This means we are not labeled with a score, but are fluidly assigned a risk score based on our data. Upstart claims 90 percent of loans go through with no human intervention, and the model evaluates more than 2,500 data points per applicant. The company hit $1 billion in annual revenue in 2025 and returned to profitability. Co-founder and CEO Dave Girouard is a former president of Google Enterprise--Upstart was an early Thiel Fellowship company.

Post-IPO Debt (Forward Flow), $1.25B, April 29, 2026.

Pagaya

Pagaya is an AI-underwriting partner that supports dozens of consumer lenders, banks, and fintechs. Lenders send Pagaya the borrowers they would otherwise decline; Pagaya’s models re-underwrite them, approve the credit-worthy ones, and then bundle the loans into asset-backed securities sold to institutional investors.

Post-IPO Debt (ABS), $1.1B combined, April 6 and April 24, 2026.

K1x

K1x is a fintech that digitizes and automates private market tax data processing--the K1s and 1099s that flow back from private equity funds, partnerships, and real estate vehicles into tax preparation cycles.

Private Equity, $175MM, April 16, 2026.

Rogo

Rogo is the Harvey of investment banking--an AI research and analysis platform for bankers and asset managers, automating the diligence, market research, and modeling workflows that junior analysts have always done overnight.

Series D, $160MM, April 29, 2026.

Wealth.com

Wealth.com provides AI-powered estate and tax planning software for financial advisors and wealth management firms, with the bet that estate planning is one of the most under-automated white-collar functions in the country.

Series B, $65MM, April 16, 2026.

Strala

Strala is an AI-powered insurance technology company providing automated claims processing for carriers and managing general agents.

Series A, $51.3MM, April 2, 2026.

Acclaim AI

Acclaim AI is a voice-first AI customer experience platform built specifically for banking, fintech, and other regulated industries where every recorded word is required for compliance.

Series A, $40MM, April 1, 2026.

Healthcare: New Markets for Old Molecules--$970MM

April is the month medical AI stopped being about narrow point solutions and started consolidating into platforms.

IKS Health

IKS Health raised $670MM in debt financing to fund its acquisition of US healthcare technology firm TruBridge--a $565MM deal that pushes the combined entity to over 2,000 healthcare organizations and 150,000 clinicians, with the strategic goal of pumping electronic health records into IKS’s agentic AI care platform.

Acquisition Debt, $670MM, April 23, 2026.

Aidoc

Aidoc is a clinical AI company that builds image-reading models for radiology, cardiology, and stroke departments. What’s distinctive is CARE, their clinical foundation model--the FDA cleared CARE as a triage system covering more than ten indications from a single architecture, the first time a foundation model has been approved for multiple acute conditions at once. The platform now gathers data from 60 million patient cases annually across nearly 2,000 hospitals. The goal is for the AI to see the image and write the radiology report, which a radiologist then reviews and signs. The CEO compares it to what Cursor and Claude Code are doing for software engineers. Diagnostic errors contribute to roughly 400,000 U.S. deaths a year; so this could help those nice doctors in the Pitt relax a little bit.

Series E, $150MM, April 29, 2026.

AcuityMD

AcuityMD is a medical device “commercial intelligence” platform--like CRM for medical sales. It ingests claims data, hospital affiliations, physician practice patterns, and outcomes data to help medtech sales reps and marketers figure out which doctors to call on and what to say.

Series C, $80MM, April 21, 2026.

Biossil

Toronto-based Biossil exited stealth in April with $70MM in equity from OpenAI, Founders Fund, and Staircase Ventures. The thesis is: instead of using AI to invent new molecules--which most AI biotech startups have failed at--Biossil uses AI to scavenge through drugs that failed in late-stage human trials and find reasons they should be revived. I wrote about one such case in my book--see the chapter on “Matching: the Coupon and the Cancer Genome” about the repurposing of a drug larotrectinib and its impact on NTRK mutations. Anyway: Biossil’s platform reads publicly available data on each molecule (papers, raw data, press releases, securities filings), uses embedding models to plot those molecules against disease genetics, and identifies candidates for repurposing the drugs. The company has already licensed and is running late-stage trials on a Johnson & Johnson sickle cell drug whose original endpoint was a miss. The company has trials going for sickle cell disease, idiopathic pulmonary fibrosis, glioblastoma, breast cancer, and Alzheimer’s.

Venture, $70MM, April 23, 2026.

Entertainment, Marketing & Advertising: Data for Sports Betting— $793MM

Stats Perform

Stats Perform is a British sports AI company and is behind almost every sports data product you encounter--the Opta player stats inside broadcast graphics, the real-time scoring in fantasy apps, the betting odds models used by sportsbooks. It’s now the first-ever official worldwide betting data and betting streaming rights distributor for FIFA--including all 104 matches of the 2026 World Cup, the 2027 Women’s World Cup, and seven other FIFA tournaments through 2029. The data side: they tracks 500,000+ matches and 3,900 competitions annually, generating effectively infinite data points per game that feed prediction models used by Vegas, settlement engines for prediction markets, and AI products inside leagues themselves. I don’t love this sector: listen to my hero Michael Lewis’s podcast, season 5 of “Against the Rules,” and you’ll see why.

Debt + Funding Round, $750MM combined, April 10, 2026.

Bluefish

Bluefish is an AI marketing platform that helps brands manage how they appear across AI-powered search and discovery channels--the same category Profound and Brandlight are building in, but with a focus on Fortune 500 and their needs (security, compliance).

Series B, $43MM, April 14, 2026.

Defense, Government & Space: Next-Gen Satellites— $431MM

See above: Satellites create a “data-only replica of our home planet”

Xoople

Spanish Earth-observation startup Xoople has spent seven years building enterprise distribution, integrating its EarthAI platform into Microsoft Azure and Esri’s geospatial tools while running on government-collected Sentinel-2 data. The new round funds Xoople’s first proprietary constellation, with sensors being built by L3Harris that Pirondini claims will deliver data “two orders of magnitude better” than current commercial satellites. The pitch is that AI agents need a verified, continuous time-series of physical change on Earth--”Earth’s system of record”--and that pitch succeeded while the company has not yet flown its own hardware.

Series B, $130MM, April 6, 2026.

Scout AI

Scout AI is a defense technology company developing AI foundation models for autonomous unmanned warfare and multi-domain military operations.

Series A, $100MM, April 29, 2026.

Yunyao Aerospace Technology

Chinese commercial satellite operator Yunyao Aerospace runs a low-Earth-orbit nanosatellite constellation focused on global weather data collection--the same category as Spire on the other side of the world.

Series B, $72.6MM, April 2, 2026.

Spire Global

Spire is a publicly traded (NYSE: SPIR) space-to-cloud analytics company that operates a constellation of LEMUR nanosatellites doing RF sensing, radio occultation (c’mon, keep up with the lingo) weather data, ship and plane tracking, and spoofing/jamming detection. The company sold off its maritime business in 2025 and has been rebuilding around weather, agriculture, and defense data, with recent contracts from NOAA and Deloitte. One focus of the new strategy is agriculture: Spire launched an integrated soil-moisture and weather product that combines 40 years of historical data, with daily satellite observations, and 45-day AI forecasts, sold via API to insurers and growers.

Post-IPO Equity, $70MM, April 8, 2026.

Sidus Space

Sidus Space designs, manufactures, and operates modular satellites and AI-driven space data products, including Earth observation as a service for government and commercial clients.

Post-IPO Equity, $58.5MM, April 19, 2026.

Professional Services: Focus on Cybersecurity, Legal— $425MM

Cybersecurity continues from last month, filling out a category that is supports physical access and digital security. Modus is taking on audit/accounting workflow. Legora is back for a smaller follow-on, which is interesting because in March they had just raised $550MM. Patlytics is a new entry--the IP-law version of what Harvey did for general law.

Deepdun

Deepdun is a cybersecurity company providing AI-powered edge security and DDoS protection for enterprise and cloud environments.

Series A, $100MM, April 21, 2026.

Modus

Modus is an AI-native audit technology platform building automated accounting and financial compliance for enterprises.

Series A, $85MM, April 7, 2026.

Artemis

Artemis is a cybersecurity startup developing AI-powered protection against AI-generated threats and attacks--defense built specifically for the new offensive capabilities Anthropic’s Mythos demonstrated last month.

Series A, $55MM, April 15, 2026.

Legora

Stockholm-based legal AI platform Legora--the European version of Harvey--added a $50MM Series D round to extend the $550MM March raise.

Series D, $50MM, April 30, 2026.

Alcatraz AI

Cupertino-based Alcatraz AI builds physical access control hardware that uses facial authentication instead of badges--its flagship product, the Rock, replaces card readers and turnstiles. The technical hook is that the system does not store facial images: it converts each face into an encrypted mathematical representation bound to a credential, so there is no biometric database to breach, and it complies with GDPR, CCPA, and Illinois BIPA out of the box. Founder Vince Gaydarzhiev was a former Apple Face ID engineer. Customers include AI data centers (the company says it grew 300% in that vertical in 2025), Fortune 100 campuses, major US airports, NFL teams, and universities. The story Alcatraz tells is that tailgating--somebody politely holding the door open for an unauthorized person--is the number-one cause of physical security breaches, and badge-and-passcode systems are useless against it.

Series B, $50MM, April 1, 2026.

Actively AI

Actively AI is a sales-intelligence company that uses AI reasoning models trained on a company’s top human sellers to identify and prioritize the highest-value accounts, and run “per-account agents” that synthesize all the relevant signals into a strategy for each named target for enterprise B2B sales.

Series B, $45MM, April 28, 2026.

Patlytics

Patlytics is an AI-powered patent intelligence platform automating patent drafting, infringement analysis, and invalidity research workflows--the patent-law version of what Harvey built for contracts.

Series B, $40.5MM, April 8, 2026.

Retail / Consumer: Voice-y--$355MM

Four companies, three of which are voice-AI customer service plays in different industries (Avoca for trades, Netomi across enterprise, plus Sona for frontline workforce management).

Avoca

Avoca is an AI workforce platform providing 24/7 voice-AI customer service agents for home service businesses--HVAC, plumbing, roofing, electrical--where 30-50% of inbound calls historically go unanswered after hours.

Series B, $125MM, April 27, 2026.

Netomi

Netomi is an AI-first customer service platform that automates support across voice, chat, email, and social channels for enterprise customers, with a focus on industries where call volume and language complexity have historically blocked automation.

Series C, $110MM, April 30, 2026.

Wiliot

Wiliot makes battery-free Bluetooth sensors--sticker-sized tags called “IoT Pixels” that harvest energy from ambient radio waves and report continuous location and condition data into Wiliot’s cloud. The platform processes billions of sensing events monthly across millions of products to basically do all the inventory tracking we thought RFID tags were going to do, decades ago. At the time I worked at Nielsen, with the grocery data business teams, and we thought RFID tags were going to blow up CPG data and measurement. Maybe IoT Pixels will now. The April round was a $75MM strategic minority investment from Avery Dennison, the global materials science and labeling giant, which also took a board seat and committed to being Wiliot’s preferred manufacturing and commercial partner. The thesis is that supply chains still run on barcode scans and intermittent visibility, and that continuous item-level data--particularly for retail, logistics, and food--is the missing layer that AI ops platforms need. Also the sensors check whether refrigerated items are kept cold.

Corporate Round, $75MM, April 27, 2026.

Sona

Sona is an AI scheduling, HR, payroll, and compliance system.

Series B, $45MM, April 1, 2026.

Transportation / Warehousing: Loop’s Supply Chain Bet--$95MM

Just Loop this month! A good one, see below.

Loop

Loop is an AI platform for logistics and supply chains, started by two ex-Uber operators (Matt McKinney and Shaosu Liu) to focus on freight audit and payment, now now they’re expanding across the entire back office. The Series Clands at a moment when tariffs, supplier diversification, and higher energy costs are forcing every multinational to wrestle with their supply chain data. Loop’s core technology is DUX, a family of models and agents built for logistics--they ingest the messy mix of PDFs, EDI feeds, emails, and ERP/TMS (transportation management system) /WMS (warehouse management system) data that runs supply chains and convert it into intel useful for finance and operations. Customers include Olipop, Kendra Scott, Outset Medical, Clemens Food Group, and Dot Foods. Loop is betting that the back-office of physical commerce is under-digitized.

Series C, $95MM, April 17, 2026.

Energy: The Hunt for Hot Rocks--$40MM

Zanskar

Zanskar is a geothermal energy company that uses AI-powered subsurface exploration to discover and develop new geothermal resources--applying machine learning to seismic, geological, and historical drilling data to predict where geothermal heat exists below ground. Geothermal is a pretty interesting space--see good New Yorker article.

Debt Financing, $40MM, April 16, 2026.

Agriculture: AI Takes On Broccoli--$34MM

Afresh

Afresh is the AI platform for grocery ordering, started in 2017 by Stanford MBAs Matt Schwartz and Nathan Fenner after they walked into grocery stores and watched produce managers do inventory on paper printouts. The April round of $34MM was co-led by Just Climate (Generation Investment Management’s climate-focused arm) and HighSage Ventures. The product is built around the fact that fresh food is uniquely hard for AI--shelf life is measured in days, demand swings with weather, and inventory data is incomplete because items lose weight, get damaged, and get mis-scanned. Afresh’s models are tuned for short shelf life and imperfect data. The platform is deployed in 12,500+ departments across 40 states, with customers including Albertsons, Meijer, Wakefern, and Stater Bros. They claim to have prevented more than 200 million pounds of food waste to date.

Venture, $34MM, April 21, 2026.

As we do each month, we offer a case study about a company that’s raised major funding, to see what cutting-edge tactics you can learn from using our “7Ds” framework. This month we profile WHOOP, the sports wearable company which raised $575 million in March at a $10.1 billion valuation. Maybe it’s a strange choice for The DataStory: it’s B2C; and here at the DataStory, we are not athletes. But WHOOP reveals why biometrics are fascinating.

BIOMETRICS BEDAZZLE US

In biometrics the journey from raw reality (my pulse) to metrics (my heart rate variability, or HRV) to action (“Justin, you have AFib and need to see a doctor”) is stark, vivid, and kind of strange—the measurement is not out there in some grocery store; it’s happening right there on my wrist and my chest.

Second, by stepping outside our usual frameworks—about say financial or sales data—biometrics in general, and WHOOP in particular, remind us of several important principles:

1. The importance of passive measurement. The biometrics industry’s central insight is that the richest signal comes from continuous, unobtrusive collection. No friction. Compare that to altrenatives for healthcare data collection: episodic at best, and requiring the patient to show up and get stuck with a needle. In your business this is the equivalent of purchase data, browsing patterns, dwell times, and product use--all passive signal where the customer “votes with their feet” and doesn’t filter or add response bias.

2. The value of data-gathering infrastructure. Over time WHOOP claims to have gathered 24 billion hours of physiological data. That’s an asset, a moat, and an attraction for strategic investors Abbott Labs and the Mayo Clinic.

3. The gap between metrics and diagnosis. In biometrics, you can gather a signal (HRV), you can even interpret into a score (“Strain”)… but to make a diagnosis, now you need FDA approval. Thematically we see that with many AI companies that use metrics to surface events that require attention—and draw humans into the loop.

4. The gap between population and individual-level decisions. Most medical inferences come from population studies. The leap from “in a population of 100,000 people, this signal correlates with this condition” to “your specific reading means you have this condition” is pretty vast. This is true in nearly every analytics-driven business scenario. Your churn model is right about the population; it’s less right about any individual customer. Your employee engagement survey reflects something true at the org level; it tells you much less about a specific person or team. To leave that gap is to leave a lot of value on the table—measured in revenue and profit in one world, in healthy vs. sick people in this one.

THE BODY AS DATA

Zooming back in on biometrics, it’s fascinating to think about the body as a zone of measurement opportunities. I walked the floor of CES with a healthcare executive who told me their company was looking at gait as a signal for serious conditions. I didn’t get it at the time. Now I do. We created this (selective) table to show the scope of what’s happening out there in biometrics:

With all this data being gathered, these physical systems are becoming legible to algorithms in a way that’s—needless to say—unprecedented in our history.

RACING PULSE

This is happening fast, and gaining speed, right now. Apple Watch and Fitbit irregular-rhythm notifications constitute a massive largest cardiac screening deployment. In 2024, my alma mater Samsung, and Apple, both received FDA clearance to detect sleep apnea from the wrist. The FDA cleared the first OTC continuous glucose monitor that same March. Apple put a clinical-grade hearing aid into AirPods. South Korean researchers are predicting cardiovascular disease from retinal photographs. An Innolitics analysis counted 295 AI/ML-enabled medical device clearances in 2025 alone.

It’s also happening at tremendous scale. How many people are walking around with a biometric device on right now? Like all market size info, this is hard to source; but based on best available: Roughly 563 million smartwatch users globally, before adding smart rings and other biometric devices. Apple Watch has an installed base of ~170–200 million people. Garmin has shipped ~320 million devices since 1989, with 20M+ units sold in 2025. Fitbit (now Google) has ~33 million monthly active users. Oura sold its 5.5 millionth ring in 2025. WHOOP has 2.5 million users. Add it up (not sure how to factor in duplication) and it could be an installed base of biometric devices is north of… half a billion.

OK, TO WHOOP

Below, you’ll read about WHOOP, and where they fit in. One thing I’ll note is the boundary WHOOP is pushing, linking consumer wearables to into medical territory. The Series G round was led by Collaborative Fund, but there were strategics in this round, Abbott Laboratories (the global leader in continuous glucose monitoring) and Mayo Clinic. WHOOP MG includes an FDA-cleared ECG feature and Heart Screener that detects signs of atrial fibrillation. And their Advanced Labs product actually prompts to user to head into a Quest Diagnostics lab for medical testing.

Where would this lead? Sports tech, consumer tech, bridging the gap to medical treatment proper. You can almost see the Jetsons robot doctor propped up in the corner of your home like a Dyson vacuum, ready to take your vitals and scold you for your high sugar.

A quick competitive snapshot: Oura is the closest direct rival—a similar subscription model, $1B revenue this year tracking toward $2B, and Dexcom (Abbott’s CGM rival) already in the cap table. Apple and Garmin dominate the broader wrist market and have added recovery metrics. Fitbit launched a Daily Readiness Score. Voice biomarker startups, retinal-imaging AI, breath VOC analysis—these are different signal classes but the same general idea.

TL/DR

Here is what you will find in this case study:

Demand: Athletes, executives, and health-curious consumers.

Dilemma: A desire to maximize performance.

Data: Five sensors, 100 readings/second, 100MB per user per day, 24 billion hours of physiological data, plus journal entries and lab biomarkers.

Derivation: Cleaning data into metrics like HRV. Proprietary scoring models compress the day into Strain, Recovery, Sleep. A LLM-based Coach (GPT-4) adds the conversational layer.

Delivery: A screenless band you never take off, plus an app that briefs you in the morning and gives recaps at night.

Decision: When to push, when to rest, when to sleep, when to see a doctor.

Destination: Personalized health and performance optimization.

WHOOP OVERVIEW

Date of Founding: 2012, at the Harvard Innovation Labs.¹¹

CEO Profile

Will Ahmed grew up on Long Island. Egyptian-American. Captained the men’s squash team at Harvard. Graduated 2012 with an A.B. in government.¹² He sits on the Board of Fellows of Harvard Medical School. Has meditated 20 minutes a day for eight years.

Founding Story

Ahmed was overtraining as a squash player—pushing hard, getting fitter, then suddenly hit a wall and didn’t understand why. He wrote an academic paper on continuously measuring the human body, and turned that into a business plan. He founded WHOOP in 2012 with John Capodilupo (engineering) and Aurelian Nicolae (mechanical). Their strategy was to start with elite athletes—Michael Phelps, LeBron James—whose fame would jumpstart the brand. The first commercial WHOOP shipped in 2015.

Funding History

$958 million total across nine rounds. Most recent: $575M Series G, March 2026, $10.1B valuation. Led by Collaborative Fund. Strategic investors: Abbott Laboratories and Mayo Clinic. Other institutional money: Qatar Investment Authority, Mubadala, IVP, Foundry, Accomplice. Athletes: Cristiano Ronaldo (also global ambassador), LeBron James, Rory McIlroy, Niall Horan, Reggie Miller. SoftBank put in $200M in 2021 at a $3.6B valuation.

What the Company Does

WHOOP sells a fabric strap with five sensors that’s worn 24/7 and can even embed in clothing. The wireless charger snaps onto the band so it never has to come off—which removes the most common reason people abandon wearables.¹¹ The data flows to an app showing three core scores: Strain, Recovery, and Sleep (more on these and other metrics, below).

How the Company Makes Money

They sell annual subscriptions from $199/year to $359/year. After the initial 12-month commitment, members can switch to monthly ($25–$40). Hardware is essentially free; free upgrades to new devices come with active membership. Members who lapse lose app access; the device stops tracking. WHOOP Unite (B2B) sells aggregated wellness analytics to corporations and military organizations. They hit a $1B revenue run rate at end of 2025. 2.5M members globally.

Top Named Clients

Athletes: Cristiano Ronaldo, LeBron James, Michael Phelps, Patrick Mahomes, Virat Kohli, Rory McIlroy, Tiger Woods, Justin Thomas, Nelly Korda. Crown Prince of Dubai posts his recovery scores on Instagram. Team partnerships: Ferrari F1 (announced January 2026), Paris Saint-Germain through 2029, the PGA Tour, the WTA, MLB.¹¹²⁹

DIAGRAM

The DataStory’s own 7 D’s framework, applied to WHOOP.

1. DEMAND

Who is their user?

Priority User 1: The Performance Athlete.

Athletes either paid (NFL, PGA Tour, Premier League) or personal (CrossFit athletes, marathoners, cyclists). They worry about overtraining and injury.

Priority User 2: The Health-Focused Executive.

Late 30s to 50s, busy professional, often a CEO. Sleep is irregular. Travel disrupts everything. They are high-octane enough to think of their performance as something to be professionally optimized.

Priority User 3: The Health-Curious Consumer (newer cohort).

Especially women. Female members were up 150% YoY in early 2026 and engage with WHOOP AI ~30% more than men. Interested in sleep, hormonal cycle, stress, and long-term healthspan. WHOOP launched a Women’s Health Specialized Blood Biomarker Panel specifically for this group. Strangely no mention of a hypochondriac segment.

2. DILEMMA

What makes achieving that goal difficult?

Most users in these segments already own a wearable—an Apple Watch, Fitbit, Garmin, or Oura. The challenge WHOOP addresses is collecting biometric data and then turning it into useful guidance for the individual.

The Athlete.

Trains year-round and wants to know whether to push hard or take a recovery day.

The Executive.

Travels frequently and sleeps irregularly. Their need is to make personal changes around travel, alcohol, or workload, and get guidance.

The Health-Curious Consumer.

Often a woman. Most wearables were originally designed around male athletic-performance metrics and don’t take into consideration effects of the menstrual cycle, perimenopause, and pregnancy.

3. DATA

What data can be accessed to achieve the goal?

Five sensors gather 100 measurements per second, totalling 100 megabytes of data per user per day—roughly the equivalent of streaming 25 minutes of Netflix every day from your wrist. (What do you call that movie—Inside Out?) Across its user base, WHOOP claims to have accumulated 24 billion hours of physiological data. That is the dataset that Abbott and Mayo Clinic invested in. The data flows from four pipelines:

Data Type 1: Photoplethysmography (PPG) heart rate data.

PPG = shining light into the skin and measuring how much reflects back. As blood pumps through the wrist, blood volume under the sensor changes with each heartbeat, and the reflected light changes with it. From these signals, WHOOP derives heart rate and heart rate variability (HRV)—the small variations in time between consecutive heartbeats, which is interpreted in multiple ways.

Data Type 2: Motion, temperature, and ambient light.

Accelerometers measure user movement.

Skin temperature sensors measure sleep (temp drops slightly in deep sleep) and illness or menstruation (temp rises slightly).

Ambient light sensors check the physical context of the user (dark bedroom vs. sunlit room).

Data Type 3: User Journal and behavioral data.

WHOOP asks users to enter 30 seconds of context per day: alcohol, caffeine, late meal, meditation, screen time before bed, travel, illness. By correlating journal entries with biometric outcomes—“what happens to your recovery on days you drink alcohol?”—WHOOP can interpret the sensor info, or at least have a hypothesis to work from.

Data Type 4: Lab biomarkers (Advanced Labs).

Users can order actual blood panels in the app. Reminiscent of crazy Theranos, I know—note that the wearable doesn’t claim to prick the skin and administer the test itself. The service sets up the user with a Quest Diagnostics appointment, and results come back into the app. A standard panel might include 65 biomarkers including vitamin D, cholesterol, glucose, hormones, inflammation. A Women’s Health Panel launched in April 2026 focused on hormonal markers.

4. DERIVATION

How do analytics/AI make the data usable?

Three layers very similar to our Analytics Pyramid: Layer 1 makes the raw data usable. Layer 2 compresses it into scores. Layer 3 turns the numbers into a conversation

Layer 1: Signal processing and sleep staging.

Sleep stage measurement is inferred from continuous PPG, motion, and temperature… to classify each 30-second window as REM, deep, light, or awake.

Layer 2: Proprietary scoring models (Strain, Recovery, Sleep).

“Strain” is a 0–21 score representing the intensity of heart rate over 24 hours, weighted by the duration of that intensity. The 0–21 scale echoes the “Borg scale of perceived exertion,” which sounds like a magic item from Dungeons & Dragons, but is instead a sports science metric developed in the 1960s.

“Recovery” comes from three primary inputs: HRV during the previous night’s sleep, resting heart rate, and the amount of time the user spent in restorative sleep stages. Then the user gets a traffic light: Green (67%+) = ready to rock. Yellow (34–66%) = moderate effort. Red (0–33%) = rest day.

Sleep is a 0–100% score derived from a model that personalizes recommendations to the individual based on the day’s strain, prior sleep debt, and naps. The output is a specific recommended bedtime. Naps, bedtime… this is my kind of app.

Layer 3: WHOOP Coach (generative AI on OpenAI GPT-4).

This layers conversational AI on top of the data. The member asks: “why is my recovery low this week?” The system pulls relevant data and sends it (the company claims, anonymized) along with the question to an LLM partner. (According to my sandwich framework, WHOOP is a bread company.) The returning answer combines the LLM’s general knowledge (of physiology in this case) with the user’s specific data points, and produces a personalized answer in plain language.

Also, the user gets a “Daily Outlook” each morning and “Day in Review” in the evening.

Lastly, the company uses AI to estimate physiological age (as opposed to chronological age). They do this based on nine biomarkers: heart rate zones, sleep consistency, VO2 max, lean body mass, others.

5. DELIVERY

How does the workflow improve for the user?

Hardware

The band is engineered for 24/7 wear that’s pure measurement, no engagement at all. The wireless power source snaps onto the band so members charge it without taking off the band itself.

Software

Daily rhythm:

• Morning. The user opens the app, sees Recovery score. Daily Outlook briefs them on the day ahead based on Recovery, schedule, training goals.

• During the day. The strain target displays on the home screen. The app prompts the user to confirm activity (run, lift, yoga) so the algorithm attributes heart rate spikes correctly.

• Evening. Day in Review: strain, sleep, behaviors. The user logs Journal entries (alcohol, caffeine, etc.). Then the Sleep Need calculation outputs a recommended bedtime.

• Anytime. The WHOOP Coach is available for ad-hoc questions.

6. DECISION

What decision is enabled by using the data?

WHOOP enables lots of small daily decisions but cumulatively, you can see how they add up.

• Training. Hard workout today, or easy day?

• Sleep. What time should I go to bed?

• Behavior. Another glass of wine?

• Health. Should I see a doctor? WHOOP’s respiratory-rate algorithm flagged early COVID-19 cases, and crazily, is documented in academic literature.

7. DESTINATION

What does success look like?

For the health-conscious consumer, the exec, the athlete, interacting with a smart device—especially one tuned their bodies—is about optimization.

Which is strange to say when you come from adtech—a multi-hundred-billion-dollar industry built by the largest tech companies in the world—focused on getting people to click ads. This is optimization of health, performance, even longevity.

Just as ad-tech-ers create a closed loop between ad exposure, ad impact, and action taken on an ad campaign (shift impressions to gaming apps!), WHOOP creates a loop between biometrics and actions… decisions… the user can take.

Although arguably the device industry is still missing a slice—namely, the actual outcome. Was the executive more effective when they went to bed at 10:38pm? Was the user healthier when they detected they had Covid (maybe that one’s obvious)? Was the athlete able to achieve greater personal records? (Although atheletes can log performance over time, so it can be seen in trends.) The biometrics are—by my take—proxies.

But in any case, the wonder of this space, I believe, is the personalized feedback loop: feeding you back data on your own body in real time, and interpreting it for decisions. Our bodies already provide this information in the form of mood and a million other nuanced signals—says the MBA who hasn’t taken a biology class since 12th grade AP Bio—but not with explanation. With these devices, you can filter the signals through algos designed by people who did take bio in college and get informed counsel. To optimize your health all around.

8. GLOSSARY

Heart Rate Variability (HRV): Small variations in time between consecutive heartbeats.

Photoplethysmography (PPG): Measures blood flow by shining light into the skin and detecting how much reflects back.

Strain (WHOOP-specific): 0–21 score for 24-hour cardiovascular load—heart rate intensity weighted by duration.

Recovery (WHOOP-specific): 0–100% score, calculated each morning. Built from HRV during slow-wave sleep, resting heart rate, and restorative sleep time.

Polysomnography (PSG): The standard sleep test. Electrodes attached to scalp, face, chest. Measures brain waves, eye movement, muscle tone, breathing.

Healthspan (WHOOP-specific): Estimates physiological age from nine biomarkers (heart rate zones, sleep consistency, VO2 max, lean body mass, etc.). Distinct from chronological age.

Advanced Labs (WHOOP-specific): The company’s blood biomarker testing program (launched 2025, U.S. only).

Hardware-as-a-Service (HaaS): Physical hardware delivered through recurring subscription, not one-time sale.

9. NOTES & SOURCES

1. Connie Loizos, “Whoop’s valuation just tripled to $10 billion,” TechCrunch, March 31, 2026. https://techcrunch.com/2026/03/31/whoop-valuation-10b-series-g-fundraise/

2. “Whoop Raises $575 Million at $10 Billion Valuation on Its Way to an IPO,” Bloomberg, March 31, 2026. https://www.bloomberg.com/news/articles/2026-03-31/whoop-raises-575-million-at-a-10-billion-valuation-on-its-way-to-an-ipo

3. WHOOP, “WHOOP Raises $575 Million at $10.1 Billion Valuation,” Press Center, March 31, 2026. https://www.whoop.com/us/en/press-center/whoop-announces-series-g-funding/

4. Collaborative Fund, “WHOOP” investment letter, March 31, 2026. https://collabfund.com/blog/whoop/

5. Elise Reuter, “Whoop raises $575M, adds Abbott as strategic investor,” MedTech Dive, March 31, 2026. https://www.medtechdive.com/news/whoop-raises-575m-adds-abbott-as-strategic-investor/816229/

6. “Whoop raises $575M—Abbott, LeBron James, Cristiano Ronaldo among the investors,” Cardiovascular Business, March 31, 2026. https://cardiovascularbusiness.com/topics/healthcare-management/healthcare-economics/whoop-raises-575m-abbott-lebron-james-cristiano-ronaldo-among-investors

7. Blythe Karow, “A Proclamation on the Future of Wearables: The Real Reasons Abbott and Mayo Invested in WHOOP,” Substack, April 2026. https://blythekarow.substack.com/p/a-proclamation-on-the-future-of-wearables

8. “Whoop raises $575M series G, Abbott comes on board amid hiring spree,” Fierce Biotech, April 1, 2026. https://www.fiercebiotech.com/medtech/whoop-raises-575m-series-g-abbott-comes-board-amid-hiring-spree

9. “Whoop concludes $575m fundraise, adding Abbott as strategic investor,” Medical Device Network / GlobalData, April 1, 2026. https://finance.yahoo.com/sectors/healthcare/articles/whoop-concludes-575m-fundraise-adding-104744037.html

10. “Whoop Raises $575 Mn Series G, Abbott Joins as FDA Eases Regulatory Path,” Digital Health News, April 2, 2026. https://www.digitalhealthnews.com/whoop-raises-575-mn-series-g-abbott-joins-as-fda-eases-regulatory-path

11. “Whoop (company),” Wikipedia, accessed May 2026. https://en.wikipedia.org/wiki/Whoop_(company)

12. “Will Ahmed,” Wikipedia, accessed May 2026. https://en.wikipedia.org/wiki/Will_Ahmed

13. WHOOP, “About Us.” https://www.whoop.com/us/en/about/

14. Harvard SEAS, “How WHOOP founder Will Ahmed found his voice as an entrepreneur,” September 23, 2024. https://seas.harvard.edu/news/how-whoop-founder-will-ahmed-found-his-voice-entrepreneur

15. WHOOP, “Membership Options | Compare Plans & Features.” https://www.whoop.com/us/en/membership/

16. PriceTimeline, “WHOOP Membership Pricing Changes,” updated September 25, 2025. https://pricetimeline.com/data/price/whoop-membership

17. Bellenger CR, et al., “Wrist-Based Photoplethysmography Assessment of Heart Rate and Heart Rate Variability: Validation of WHOOP,” Sensors (Basel), May 20, 2021. https://www.mdpi.com/1424-8220/21/10/3571

18. WHOOP, “WHOOP Proven Most Accurate Wearable in Heart Rate & HRV Measurements,” October 27, 2022. https://www.whoop.com/us/en/thelocker/whoop-proven-most-accurate-wearable-in-heart-rate-heart-rate-variability-measurements/

19. “WHOOP App Features Tested,” Liveworksleep, February 27, 2026. https://liveworksleep.com/whoop-app-features/

20. WHOOP, “WHOOP Unveils the New WHOOP Coach Powered by OpenAI,” BusinessWire, September 26, 2023. https://www.businesswire.com/news/home/20230926899032/en/

21. WHOOP, “New AI guidance from WHOOP connects every part of your health,” October 16, 2025. https://www.whoop.com/us/en/thelocker/new-ai-guidance-from-whoop/

22. Rolls Performance Lab, “WHOOP 5.0 & WHOOP MG: AI-Powered Performance, Recovery, and Health Optimization Explained,” May 25, 2025. https://www.rollsperformancelab.com/news-and-articles-1/whoop-launches-5

23. “What’s Behind Whoop’s Soaring $10 Billion Valuation,” Kavout Market Lens, April 2026. https://www.kavout.com/market-lens/what-s-behind-whoop-s-soaring-10-billion-valuation

24. “Case Study: Whoop Business Strategy & Device-as-a-Service Success,” Circuly, January 27, 2026. https://www.circuly.io/blog/case-study-whoop-business-strategy-device-as-a-service-success

25. “Whoop revenue, valuation & funding,” Sacra, March 21, 2026. https://sacra.com/c/whoop/

26. WHOOP partnered with Central Queensland University researchers to validate its COVID-19 detection algorithm; identified 80% of positive cases by day three of symptoms. See “Whoop (company),” Wikipedia.

27. WHOOP, Advanced Labs and Healthspan product information. https://www.whoop.com/

29. WHOOP became the official health and fitness wearable partner of Ferrari Formula One team, January 2026. See “Whoop (company),” Wikipedia.

30. Apple Watch and Fitbit/Pixel Watch irregular rhythm notifications now constitute the largest deployed cardiac screening tool in human history. See Apple Heart Study and FDA AI/ML clearance database.

31. Samsung Galaxy Watch (February 2024) and Apple Watch (September 2024) received FDA approval for sleep apnea / breathing disturbance detection. Roughly 25M American adults have OSA, the majority undiagnosed. https://www.apple.com/newsroom/

32. FDA, “FDA Clears First Over-the-Counter Continuous Glucose Monitor” (Dexcom Stelo), March 2024. https://www.fda.gov/news-events/press-announcements/fda-clears-first-over-counter-continuous-glucose-monitor

33. Apple, “Apple introduces groundbreaking health features” — AirPods Pro 2 with iOS 18 hearing aid feature, September 2024. https://www.apple.com/newsroom/2024/09/apple-introduces-groundbreaking-health-features/

34. Mediwhale’s Reti-CVD and Reti-CKD use machine learning on retinal images to predict cardiovascular and kidney disease. https://www.mediwhale.com/

35. FDA, “Artificial Intelligence and Machine Learning (AI/ML)-Enabled Medical Devices.” 295 clearances in 2025. https://www.fda.gov/medical-devices/software-medical-device-samd/artificial-intelligence-and-machine-learning-aiml-enabled-medical-devices

38. DemandSage / Counterpoint Research, “Smartwatch Statistics 2026: Global Users & Market Share.” 562.86M smartwatch users worldwide; market valued $35.29B; Apple holds 23% market share. https://www.demandsage.com/smartwatch-statistics/

39. SQ Magazine, “Apple Statistics 2026.” Apple Watch active users surpassed 170M with 92% retention; 54M units shipped in 2025; Apple Watch health features triggered 45,000+ emergency calls. Cf. Above Avalon (Neil Cybart) estimate of 200M+ Apple Watch installed base. https://sqmagazine.co.uk/apple-statistics/

40. “Apple Q4: Why Wearables Revenue Declined Despite Record User Growth,” the5krunner, October 30, 2025. Apple Watch active installed base at all-time high; 95% U.S. customer satisfaction; half of Q4 buyers were new to the product. https://the5krunner.com/2025/10/30/apple-q4-apple-watch-revenue-new-users/

41. “How many active Garmin Connect users are there in 2026?” the5krunner, May 1, 2026. Garmin shipped ~320M products since 1989; record 20M+ units in 2025; Q4 2025 fitness revenue +42% YoY. https://the5krunner.com/2026/05/01/garmin-connect-users-2026/

42. Business of Apps, “Fitbit Revenue and Usage Statistics (2026),” March 9, 2026. ~33M monthly active users; 137M registered users; 4.3M units sold in 2025 (down 21.8% YoY); Fitbit Premium ~2.2M paying subscribers. https://www.businessofapps.com/data/fitbit-statistics/

43. Oura press release, “Oura Surpasses 5.5 Million Rings Sold,” BusinessWire, September 22, 2025. 5.5M rings sold cumulatively; ~3M sold in 2025 alone; ~2M paying subscribers; $1B revenue run rate. https://www.businesswire.com/news/home/20250922351288/

44. IDC research, cited in Sherin Shibu, “Smart Ring Maker Oura Expects Sales to Grow to $2B Next Year,” Entrepreneur, November 11, 2025. Oura holds ~80% of global smart ring market. https://www.entrepreneur.com/business-news/smart-ring-maker-oura-expects-sales-to-grow-to-2b-next-year/499465

I’ve spent the last months drinking lots of coffee with leaders of data and AI companies. Companies at different sizes, different lifestages. I see a clear pattern: nearly every business model corresponds to a part of a sandwich.

Below I’ll explain this framework and point out where there is risk--and bounty--in the role a company chooses in this three-layer stack.

DO I NEED TO DESCRIBE A SANDWICH?

Why not: Top layer of bread. Middle section let’s call the meat. Bottom layer of bread.

My theory is that all three layers of the sandwich exist in any value proposition that spans the delivery of data… to some user… to answer some question.

But few companies span the whole divide between data and user. Most are either bread companies or meat companies.

TOP BREAD: User interface layer. This is where the user gets an answer to some question. It might be an app on a handheld device--a device someone’s carrying around a warehouse (“do we have snow shovels in inventory?”) or a hospital (“what were the Glycated Hemoglobin results for patient Smith?”). This layer is home to what product folks call the use case (eg, “help retail workers find items on a large display floor to increase efficiency per task and decrease customer wait times”).

MEAT: Analytics layer. Where the data is processed to provide that answer. This layer might be a large language model (LLM) that’s reading text and analyzing it. Or it might be an algorithm that is crunching numbers in a statistical model. This is the part that does the thinking. Call it the AI or intelligence layer. For instance, it might be the natural language interface so the retail worker can type, one handed, snw shovles, and get an answer back, “Aisle 15 Bin 6.” Or it might be calling a benchmarks database so the RN knows not just what patient Smith’s Glycated Hemoglobin results were, but whether they were normal or abnormal, and whether immediate action is needed.

BOTTOM BREAD: Data layer. This is the information that’s being processed to provide the answer in the UI. The data layer does not provide answers, but it does provide the input to answers. The data can originate multiple places. It might be the proprietary data of the company whose stack we are in--data they have carefully cultivated. Or data that’s pulled from multiple captive third party databases. Or maybe it’s data the company simply bought in a license deal. Or, it might be their client’s data--for instance the client’s customer data—used to make the model’s answers more specific, and to leverage the client’s own hard-won data asset.

So those are the layers. The part a company plays in the sandwich tells you almost everything about how they’ll compete, what their moat looks like, and whether they’re about to get eaten or do the eating.

BREAD COMPANIES

Bread companies own the top and the bottom layer of the data or AI company. They own the user interface. They own the client relationship. Generally they bring some data to the table, or sit on top of the client’s data. What they don’t own is the middle.

The meat, they rent.

A pre-AI example might be a web publisher: a web page or app (UI) with journalism (data). But the content management system, they will license.

In the AI era, if the company runs on an LLM, these firms definitely won’t see the need to build their own LLM. Way too expensive. Why not rent somebody’s else’s giant model for a per-token fee?

That LLM or analytics engine, once called, can now be run through your data layer, to give the user the answer they want. That’s how the “bread” layers work together for a complete service:

I own the user layer--I know what answer they want (top bread).

So I can explain to a model how to answer (rented meat layer).

And I have the data required to provide this answer (bottom bread).

A poster child here is Harvey. Harvey raised $200 million in March. A hundred thousand lawyers across thirteen hundred organizations are logging into it. And Harvey’s own website, its own blog posts, its own investor materials will say: it is built atop GPT, Claude, Gemini. It does not train its own frontier model. What it has is a product built for lawyers, by lawyers--over 10% of Harvey’s employees are lawyers, only I bet they don’t have to wear suits--who have explained to their tools what a lawyer’s workflow is like; and they have trained the models with data. In this case: massive libraries or legal contracts and documents.

Harvey is a bread company.

Another example in healthcare: Abridge turns doctor-patient conversations into clinical notes. These consultation notes, records of every session, are legally required, and by all accounts a huge pain to create and store. Abridge’s tool listens to doctor-patient conversations on a phone app (definitely user-facing); and summarizes the visit into clinical notes that dump out into systems like Epic--where electronic health records are stored. Abridge’s app is trained on 1.5 million medical encounters--the bottom of the bread, the data that trains the model to speak like a doctor. The actual language model that writes up the notes? It can come from any company with a great LLM service.

MEAT COMPANIES

Meat companies are the inverse--they bring the analytics, the model-building expertise, the platform that ingests the customer’s data and turns it into predictions, forecasts, optimizations, simulations… answers.

What they don’t bring is the industry knowledge. They don’t know your business when you hire them. They learn it in order to work with you.

The canonical meat company was early-phase Palantir, famous for its “forward-deployed engineers” who embedded at customers, learned the business in the field, and built custom applications. (Palantir has since grown businesses out of this that are more “full sandwich”--for instance in the military.)

A purer meat example is H2O.ai. With one of their products, you literally upload a CSV file and they will make predictions for you off your rows and columns. Or they’ll train a language model on your data.

An example closer to home for my adtech friends is Chalice. Chalice helps companies who aren’t equipped with sizable data science teams to compete in the world of performance ads. For performance advertising, your whole goal is to predict who will buy a product based on seeing an ad, or who will click on your ad, or whatever. Chalice’s clients are brands like Bayer who want to do performance advertising… but don’t have a staff of stats PhDs.

Why should my aspirin maker be good at digital ad algorithms? They shouldn’t be. Rent it from the meat company!

ONE SLICE OF BREAD

Here is the category I worry about. A lot of companies in this market aren’t full bread sandwiches. They’re one slice.

I think they’re toast.

Top bread companies have an application but are not focused on harvesting the data. Okay, they can answer a question in the moment. But how are they creating long term value? Are they storing data so they can build benchmarks? Are they observing trends and advising clients? Are they digging into the data to provide new answers to more questions? If they’re not: warning sign.

Pay attention to the data you’re handling. If you don’t have the rights to use it… figure it out.

Bottom-bread-only companies are where I come from in my career, the whole syndicated-data and data-broker complex. These are real businesses with real revenue, juicy margins and decades of customer relationships. Their model is “we have the data; you figure out what to do with it.” I believe great data will always have value.

But if you’re not close to the user--if you don’t have that top layer of bread--you’re not going to see closely how use cases evolve. You won’t be able to grow with the client.

At best, you’ll be at a disadvantage in client negotiations because you don’t know the value of what you’re selling. At worst, substitute data sources will eat into your position, bite by buttery bite.

MOATS AND RISKS

What I also worry about, walking away from these meetings with data and AI CEOs, is how these nice people are going to stay in business. How do they dig a moat around their business to protect it for the long term?

Think about the tumult in the software (SaaS) market right now as foreshadowing. SaaS is fighting for its life now, saying: “We have client relationships, historical workflow, but an ancient tech stack: Now what?” They are clinging to their complex client integrations for dear life--until they figure it out or get disintermediated.

For bread companies… you’ve earned access to the user. Getting close to them, seeing them hammer away at their jobs, observing what frustrates and thwarts them, is a privilege you have earned. Now you can take their expertise and program it into an app.

But… can’t anyone with a winning smile and a few weeks of user research do the same?

For bread companies I believe the moat solution is threefold:

• Market share. Get in there, sign the biggest names you can, and repeat as fast as possible. The more client logos, the more users… the more client knowledge you access.

• Innovation cycles. Customer knowledge won’t do you much good if you’re not turning it into product. Listen to customers and ship code. Fast. You’ll have a feature and skill library that will be hard to replicate, and it will get wider as well as deeper.

• Niche choices. I once heard Steven Leavitt of Freakonomics say that his father was advised as a young gastrointestinal resident to pick a speciality no one else would choose: flatulence. Whether this is apocryphal or not, you get the idea: pick a market that few are focused on. See Spectral AI--a healthcare company that tells doctors which sections of burn wounds are healable. That’s niche.

For meat companies, the risk is the hyperscalers. Amazon SageMaker, Google Vertex AI, Azure ML--these are all meat, and they come bundled with the cloud the customer is already paying for.

When your differentiator is “we make ML easier,” and AWS ships something also claiming “we make ML easier” at 40% of the price, you have a problem.

The answer for meat companies is to pull a Palantir: shore up your meat business with a full-sandwich business. Like Palantir and the military, get deep with a customer base and create a defensible position with user access, processing, and data.

For all these businesses, there are two generic threats:

• Big AI. Not just a threat for meat companies (making processing simple for enterprises). For Big AI, integrating into a specialized data source or toolkit (like Slack, or Zoominfo) or standing up a few vertical-specific data sources, is an easy add--and then they are threatening bread companies, too.

• DIY. Clients have their own data (obviously). Even more obviously, they have access to their own users. If the meat layer is rentable… well, then all you need is a few clever kids in your engineering team, and give them a long leash to disrupt your way of doing things… and in theory, you can package your own workflow solutions with less expense and risk than licensing it. (Fine: we all know too well the agony of trying to innovate inside a big company whose expertise is not software development. Still: good results are possible.)

LEARNINGS

If you are a CEO running an AI company, or an investor evaluating one, or an employee considering an offer, the question that matters more than “what’s the TAM” or “what’s the ARR” is:

How many layers of the sandwich does this company actually own?

One layer of bread, and you are exposed. You are just waiting for a substitute offering to make your life an anxious hell. You’d better get hold of that customer relationship, or that data, fast.

Two layers of bread, or the meat layer, and you have a business. It can be a big business. But you’d better be entrenching yourself rapidly, reinvesting your proceeds into defenses like exclusive data, or massive skill libraries, or tentacle-like enterprise integrations. Or all of the above. Move fast and be paranoid.

Three layers, and you have a defensible business. It’s like Freud said of therapy: its job is to turn neurotic misery into normal unhappiness. The data and AI space right now is neurotic. If you can get three layers, you evolve beyond the neurotic frenzy; you have a moat; and you just have the normal challenges of succeeding of business: making customers happy, growing revenue, managing margins, suppliers, and talent.

The deli is open, friends. The question is: what are you serving?

Overview

March 2026 saw $62 billion flow into data and AI companies, with fully $48 billion flowing to Amazon, CoreWeave, and Vertiv—in debt—to build data centers. (See our explainer.)

But… what you care about is the $11 billion that went to specific use cases you can learn from: 63 companies that raised more than $30MM in March, pushing the edges of innovation with data or AI. First let’s look at…

The Top 10 Single Raises

Note the colossal round at Shield AI (drones), and the billion dollar seed round… seed round!… for AMI, our friend Yann LeCun and his world models (we got ahead of that one with this explainer from January).

1. Shield AI: $2.25B (Series G + Debt)

2. AMI Labs: $1.03B (Seed)

3. Nxtra Data: $1B (Private Equity)

4. VAST Data: $1B (Series F)

5. WHOOP: $575MM (Series G)

6. Legora: $550MM (Series D)

7. Cloaked: $375MM(Series B)

8. Cambridge Mobile Telematics: $350MM (Private Equity)

9. AIsphere / PixVerse: $300MM (Series C)

10. TENEX.AI / Grafana Labs (Tie): $250MM each

And now the fun part; the interpretation…

Trends for April 2026

1. Healthcare keeps logging firsts

We are whipping past firsts and milestones in healthcare. First--companies are claiming to be able to automate the matching a drug to patient. Alto Neuroscience claims to be using brain scan data to do patient matching for psychiatric drugs?-- simply mental. Deep Intelligent Pharma uses genomics and real-world-data to do so for a span of other conditions. A lot of suffering can be avoided if this works. Second milestone? Utah is letting AI prescribe meds without humans-in-the-loop. (Is this real? It is.) Doctronic is the company that broke the scrip barrier and grew 15x in six months. Healthcare, after maybe financial services and adtech, is the richest data sector in the world. It would be nice to see some scalable applications helping developing countries with low physician/population ratios? In the meantime, business leaders can focus on the use case of what I’ll call “key-to-lock matching.” Patient, meet drug. How about: business, meet vendor? Renter, meet apartment? Couch potato, meet series? Many applications for sophisticated matching as long as the sophisticated data (we are talking about brain scans) is available.

2. Robots to watch the robots to watch the robots

You know we’ve reached some kind of critical mass when companies create AI technology whose sole job is to monitor other AI technology. Onyx monitors and schools your agents; JetStream Security manages “MCP sprawl” (great phrase). I suppose it’s not so different from me creating a fact-checking agent… but it feels different. Cyclical. Russian dolls. Perhaps it represent the simple solution to the “Alignment Problem” (the notion that our models can grow to desire and pursue things we wouldn’t want them to). We prevent Skynet from creating the Terminator by creating a Skynet babysitter bot. Then… who’s babysitting the babysitter? In your business, think about that babysitting opportunity seriously. Quality control? Error detection? Startup Qodo is doing just that for software quality. If agents can monitor agents, there’s nothing too abstract for you to monitor and poke at with a bot.

3. Legal: Bots kill all the lawyers

I’m so glad my dad’s buddy the law professor told me not to go to law school. Law, being made up entirely of words, is now in full-on, bottom-up, AI re-engineering mode. It’s a no-brainer to use massive amount of legal text to train LLMs to read and redline contracts--like Harvey in the US, or Legora in Europe (lots more languages). Crosby in New York is going for a cyborg approach combining agents (not in the CAA sense… or the CIA sense for that matter) with lawyers. And the category is so hot that Spellbook actually raised $40MM in debt financing specifically to acquire competitors. It will be interesting to watch this space for a glimpse of what a fully-AI-ified labor market transforms into. Will it indeed be the cyborg model--human talent paired with machines? Will law partners now solely act as salespeople, drafting behind them (no pun intended) an army of droid drafters? I wonder about how talent will settle. I feel like sales and relationship talent will get a new lease on life.

4. Cybersecurity in the month of Mythos

Weird that thirteen bloody cybersecurity companies raised rounds in March, totaling more than $1.1 billion… in the same news cycle that Anthropic’s Mythos came into the spotlight. (Am I being naive? Did one cause the other? I’m assuming cyber is just having a moment.) ICYMI Mythos is the AI model that can autonomously discover thousands of software vulnerabilities in existing infrastructure, some that have lain undiscovered for decades. This follows a November 2025 incident where a Chinese state-linked group used Claude Code itself to run 80-to-90-percent automated cyber-espionage against around 30 organizations. If cyber defense can be automated, so can cyber-attack; and vice-versa. For the business leader: there is a more constructive and collaborative paradigm approaching: agent to agent business dealings. If two bots can face off in eternal automated deathgrip like Godzilla and Mothra who never run out of quarters… can they also dance like Fred Astair and Ginger Rogers? There I am being naive again.

Sector & Company Profiles

AI, Info Infrastructure & Tools: The Oven Window - $3.66B

Maybe we’ve turned a corner - this month the largest investment category is AI that helps AI. A lot of monitoring tools: how sound is the data (Validio)? Can we monitor the health of data and applications in an enterprise (Grafana)? Can we monitor data flows in AI applications (Dash0)? Are our enterprise projects making progress and on time (Axiamatic)? We build this massive infrastructure; now we need to peer in the oven window and make sure the cake is not burning.

AMI Labs

Yann LeCun, Meta’s former chief AI scientist, launched AMI Labs to build world models — AI systems that understand physics and three-dimensional space instead of just predicting the next word in a sentence.

Seed, $1.03B, March 10, 2026.

Nxtra Data

Nxtra is the data center arm of Bharti Airtel, India’s second-largest telecom. It plans to triple capacity to 400 megawatts by 2027, making it the country’s sovereign AI infrastructure bet.

Private Equity, $1.00B, March 30, 2026.

VAST Data

VAST Data sells storage and data platform software for training massive AI models. Its architecture lets thousands of GPUs share the same pool of data without bottlenecks. Major AI labs need this kind of plumbing; customers include CoreWeave, xAI, and the hyperscalers.

Series F + Secondary, $1B, March 9, 2026.

Grafana Labs

Grafana Labs makes the open-source dashboards that engineers use to monitor applications, servers, and now AI models in production.

Series E, $250MM, March 12, 2026.

Dash0

Dash0 builds observability for agentic AI, where a single user request can trigger dozens of LLM calls, tool uses, and sub-agent invocations that need to be debugged when something goes wrong.

Series B, $110MM, March 23, 2026.

Axiamatic

An AI platform that ingests signals from 250-plus enterprise systems to keep ERP and digital transformation programs from running late.

Series A, $54MM, March 11, 2026.

Qdrant

An open-source vector database that powers the retrieval layer of many RAG systems, with 250 million downloads and enterprise customers including Tripadvisor, HubSpot, and Bosch.

Series B, $50MM, March 12, 2026.

Gumloop

Gumloop lets non-technical employees build AI agents by dragging and dropping. Shopify reportedly using the platform to automate more than 20 million actions across the company.

Series B, $50MM, March 12, 2026.

Standard Template Labs (STLabs)

AI agents that basically are designed to replace the IT department. All that help-desk work of resetting passwords, provisioning laptops, and access requests. The thesis is that every IT ticket looks roughly the same across every company, and a well-trained agent can resolve about 70 percent of them without any human intervention.

Seed, $49MM, March 17, 2026.

Highlight AI

An AI tool that sits on top of all your work apps, to give you a single interface with Slack, Zoom, Linear, and Claude Code etc. Research puts this coordination time at roughly 24 hours per employee per week.

Series A, $40MM, March 24, 2026.

Validio

Stockholm-based Validio continuously monitors enterprise data pipelines for bad or missing values--an MIT study found 95 percent of AI projects fail because of data problems feeding the model.

Series A, $30MM, March 5, 2026.

Defense, Government & Space: Drones - $2.25B

This section is short. One company: Shield. In an age where drones have quietly become the ICBMs of the age--that is, the conflict-defining weapon--Shield emerges a winner by manufacturing spooky looking aircraft (the X-Bat!) as well as AI systems (Hivemind!) for “mission autonomy”--a buzzword you can just hear being used 10,000 per day in the Pentagon. Massive raise.

Shield AI

Shield AI builds the autonomous pilot software that lets military drones fly in GPS-jammed and radio-jammed environments.

Series G + Debt, $2.25B, March 26, 2026.

Professional Services: The White-Collar Revolution - $1.85B

Wow, a flurry of professional services raises. More deals than any other category — 18 companies raising $1.85 billion. See commentary at the top of the report.

Legora (x)

Legora is a Swedish legal AI product that works inside Microsoft Word and helps transactional lawyers draft, review, and negotiate contracts. More than 250 firms use it, including Clifford Chance and Linklaters. Founder Max Junestrand started the company in 2022 at age 25, which tells you something about how fresh this market is. Legora’s edge against Harvey in Europe is that it trains on contracts in more than 20 languages.

Series D, $550MM, March 10, 2026.

TENEX.AI

TENEX.AI operates an AI-native detection and response service, running 24/7 security operations centers with agents; they claim response times in seconds vs. a human’s 15 minutes.

Series B, $250MM, March 31, 2026.

Harvey

Harvey is the best-known name in legal AI, used by 100,000 lawyers across more than 500 firms including most of the Am Law 100, and has now trained its own legal-specific models after originally building on OpenAI.

Series G, $200MM, March 25, 2026.

Kai

AI-native cybersecurity that claims to have rebuilt the cyber category from the ground up, with agents. I keep thinking about those knights who jump off the battlements in the Battle of Hogwarts.

Series A, $125MM, March 10, 2026.

ORO Labs

ORO Labs automates the intake-to-contract process for enterprise procurement, a $180 billion-a-year function at the Global 2000 level that still runs mostly on emails and PDFs. Customers include Coca-Cola, Siemens, and Novartis.

Series C, $100MM, March 12, 2026.

Juicebox

Juicebox builds AI agents that source candidates for recruiters, replacing the inbound pipeline (which now averages 250 applications per job) with outbound outreach to people who were not applying. Customers include Ramp, Cursor, and Cognition.

Series B, $80MM, March 10, 2026.

depthfirst

depthfirst, founded by technical leaders from DeepMind, Databricks, and Faire, trains specialized security models that find and fix software vulnerabilities.

Series B, $80MM, March 31, 2026.

Qodo

Qodo verifies AI-generated code before it hits production. Customers include Walmart, NVIDIA, and Ford.

Series B, $70MM, March 30, 2026.

Crosby

Crosby is a New York AI law firm — an actual law firm with partners and clients, not a software vendor. Focuses on commercial contracts with a stated goal of reviewing one in less than an hour, where most large firms take a week. The bet is that a fully AI-native firm can underprice traditional counsel by an order of magnitude and still make more per partner, because the AI does 95 percent of the drafting.

Series B, $60MM, March 31, 2026.

Linx Security

Linx uses AI to monitor and govern the identities and permissions of employees and vendors accessing data inside an enterprise.

Series B, $50MM, March 31, 2026.

Cylake

AI-native security platform that runs entirely inside a customer’s environment — for banks, defense contractors, and regulated agencies that cannot ship telemetry to a vendor cloud.

Seed, $45MM, March 5, 2026.

Spellbook

Spellbook, a Canadian AI contract-drafting tool used by nearly 4,000 law firms in 80 countries including Nestle and eBay, raised $40MM in debt from RBC specifically to acquire smaller legal AI competitors as the market consolidates.

Debt, $40MM, March 3, 2026.

RunSybil

AI agents that conduct continuous penetration tests on live software — finding vulnerabilities the way a human hacker would, but continuously and at machine speed.

Seed, $40MM, March 18, 2026.

Onyx Security

Now we have to build systems that monitor our agents. Onyx Security claims 80% of enterprises expose sensitive adata with their agents--hence we need Onyx’s “control plane for AI agents” to discover which agents are running; monitoring their behavior; and enforcing security.

Series A, $35MM, March 12, 2026.

JetStream Security

Gives enterprises visibility and control over the AI agents, MCP servers, and custom AI applications. Coined the phrase “MCP sprawl”--nice (or not nice).

Seed, $34MM, March 3, 2026.

Fig Security

Fig Security, founded by veterans of Israel’s Unit 8200, audits the resilience of operational data security infrastructure.

Series A, $30MM, March 3, 2026.

Qevlar AI

Paris-based Qevlar automates security investigations. When your security systems send off alers, many (apparently) are not real threat. Their tools qualify the real threats quickly for human analysts. Customers include Mercedes-Benz, Sodexo, and Orange Cyberdefense.

Series A, $30MM, March 10, 2026.

Lio

Lio (formerly askLio) automates enterprise procurement with AI agents that handle the full workflow — request prioritization, supplier comparison, negotiation, vendor onboarding, and purchasing.

Series A, $30MM, March 5, 2026.

Healthcare: AI Rx - $1.09B

See commentary at top of article.

WHOOP

The wristband tracks sleep, recovery, “strain” (love that), and stress. The real product is the coaching AI behind it. The reason WHOOP matters is data: the company has reportedly collected 5 billion hours of real human biometric data over more than a decade, which is fricking bananas. Founder Will Ahmed started the company in 2012 after studying NCAA squash players. That long time horizon--14 years!--for biometric data-gathering is an outstanding moat.

Series G, $575MM, March 31, 2026.

Qualified Health

Qualified Health builds a secure, compliant AI platform for hospitals to deploy their own AI applications.

Series B, $125MM, March 25, 2026.

Alto Neuroscience

Alto Neuroscience is a publicly traded “precision psychiatry company” that uses EEG brain scans and machine learning to predict which mental health drug will work for which patient before prescribing, rather than the traditional trial-and-error approach. If this is real… incredible. Founder Amit Etkin is a Stanford psychiatrist and neuroscientist.

Post-IPO Equity, $120MM, March 16, 2026.

Thesis Care

Thesis Care deploys AI agents to handle clinical operations and care management — the work overworked nurses and patient coordinators do between doctor visits — in a hybrid model with licensed clinicians.

Series A, $45MM, March 25, 2026.

Ease Health

Ease Health launched in March to build intake, scheduling, billing, and clinical documentation tools for outpatient behavioral health clinics, one of the largest and most underserved segments of U.S. healthcare.

Series A, $41MM, March 2, 2026.

Doctronic

Doctronic became the first AI system legally authorized to practice medicine in the United States when Utah’s AI Learning Lab regulatory sandbox cleared it in December 2025 to renew prescriptions for 190 medications without a human physician. (WTF) The company has grown fifteenfold in six months.

Series B, $40MM, March 24, 2026.

Adonis

Adonis builds AI-based revenue cycle management, learning from millions of claims per month to find and recover the 15 percent of potential revenue that providers typically lose to billing inefficiencies.

Series C, $40MM, March 25, 2026.

Deep Intelligent Pharma

Deep Intelligent Pharma is a Singapore-based company that uses AI to design and run clinical trials more efficiently. The platform combines real-world evidence, genomic data, and AI simulation to identify patient populations where a drug is most likely to work, shrinking the cost and timeline of traditional trials. Clinical trials typically cost $40–50MM per drug and take seven to ten years. If AI can cut either number by half, the economics of drug development change--which is why the company has partnerships with several large pharmaceutical firms and is building operations across Asia.

Venture, $40MM, March 24, 2026.

Waiv

Waiv builds AI-native precision tests that read digital histopathology slides and predict molecular biomarkers, treatment response, and prognosis for oncology patients.

Venture, $33MM, March 13, 2026.

Spectral AI

Spectral AI (NASDAQ: MDAI) uses multispectral imaging to tell trauma surgeons which parts of a burn or wound will heal and which will not.

Grant, $31.7MM, March 18, 2026.

Entertainment, Marketing, Advertising: Make a Data Face - $555MM

Five companies, half a billion dollars, and creative content made in minutes. Notable concentration of companies in China here.

AIsphere / PixVere

AIsphere, known to consumers as PixVerse, is a Beijing-based AI video generation company that lets users generate short videos from text prompts; it competes directly with Google’s Veo. The company reports more than 100 million users globally, about 60 percent outside China. PixVerse has become a default tool for social media creators because of its focus on character consistency — a hard problem in AI video where the same person is supposed to look the same across multiple shots. Founder Yaohui Wang previously worked at ByteDance.

Series C, $300MM, March 12, 2026.

Granola

Granola is a London-based AI meeting assistant that sits in the background of video calls and produces cleaned-up notes without a bot joining the meeting. Became the default for many tech workers in the last year.

Series C, $125MM, March 25, 2026.

VAST (3D)

Hangzhou-based VAST builds 3D foundation models for gaming, robotics, and industrial design, serving more than 3 million professional creators, 80 percent of them outside China.

Series A, $50MM, March 5, 2026.

Tripo AI

Tripo AI uses funds to expand its developer base — now 6.5 million creators and 90,000 developers — with new architectures that generate polygon meshes in seconds. A polygon mesh is pretty much what it sounds like. Like a wire-model of a character to be rendered.

Venture, $50MM, March 25, 2026.

Sett

Tel Aviv-based Sett generates and tests mobile game ads with AI agents at clients including Zynga, Playtika, and Papaya Gaming, built on the premise that 97 percent of creative concepts fail and the only way to find the 3 percent that work is to generate thousands of variants.

Series B, $30MM, March 31, 2026.

Manufacturing, Chips & Robots: Physical AI and digital doubles. $524MM

Six companies working on the physical layer of AI… or maybe more accurately, building out an uncanny valley between physical and IRL worlds, to make real things. Lightwheel trains robots (fake humans) with simulation software (fake reality). Normal Computing uses software to monitor every aspect of real-world hardware testing. Elite robots create “cobots” where humans and machines labor together in harmony. All focused on efficient building. I love the idea of simulation for working through bad ideas and mistakes--infomercial “before” clips processing infinitely in some data center so your knife can stop slipping on the real life tomato.

Lightwheel

Lightwheel is a Santa Clara company that builds simulation environments where robots learn to navigate the physical world before they are deployed in it. The product generates synthetic training data for robot vision and manipulation — thousands of hours of simulated work that would cost millions to gather in the real world. This sim-to-real transfer approach is the quiet reason robotics is suddenly working; founder Hang Su previously led autonomous driving research at Baidu.

Series A, $145.5MM, March 11, 2026.

D-Robotics

D-Robotics builds edge AI computing platforms that handle real-time perception and control for driverless vehicles and industrial robots without sending data to the cloud.

Series B, $120MM, March 16, 2026.

Elite Robots

Elite Robots makes collaborative industrial arms — “cobots”— that work alongside humans on Chinese factory floors, bridging the gap between fully automated production lines and manual labor by handling the dull, repetitive work while humans take exceptions.

Series D, $87MM, March 25, 2026.

Nominal

Nominal builds the software that hardware engineers at SpaceX-style companies use to test aircraft, satellites, autonomous vehicles, and fusion reactors. In ten months it has landed four of the five largest U.S. defense contractors as customers.

Series B-2, $80MM, March 5, 2026.

Normal Computing

Normal Computing builds AI chips based on thermodynamic computing — a different paradigm that uses physical randomness instead of suppressing it — with its CN101 chip targeting a 1000x energy efficiency gain for generative AI inference versus conventional GPUs.

Strategic, $50MM, March 25, 2026.

Dexmate

Dexmate builds dexterous robot hands for mobile humanoids, and its Vega robot has 36 degrees of freedom, handles 15 kilograms with both arms, and runs 20 hours on a battery; LG Technology Ventures took a stake in March.

Venture, ~$41MM, March 2026.

Retail / Consumer: Mystery Shopping - $445MM

I think Cloaked is a major privacy concept. A privacy avatar for the average consumer. Create and shed skins as a digital consumer, to obscure your real behavior.

Cloaked

Cloaked generates fake but working phone numbers, emails, and credit cards that consumers use in place of their real contact info, keeping the real identity clean even when an alias gets compromised or sold.

Series B, $375MM, March 19, 2026.

Lace AI

Lace AI listens to every inbound call at HVAC, plumbing, and roofing companies, identifying the 15-to-40-percent of revenue that leaks out through missed bookings.

Series A, $40MM, March 23, 2026.

Capacity

Capacity consolidates voice AI, quality assurance, and speech analytics into a single contact center platform serving more than 20,000 businesses.

Series E, $30MM, March 2026.

Transportation / Warehousing: Smart Streets - $440MM

Cambridge Mobile Telematics

Behind most of the U.S. insurance industry’s usage-based insurance programs. The phone in a driver’s pocket has accelerometers and gyroscopes that can detect braking patterns, phone use while driving, and cornering force; CMT collects that data from tens of millions of drivers and feeds it to State Farm, Allstate, Liberty Mutual, and dozens of other insurers. CMT has been repositioning itself around teen driver safety — you can see why. It’s a friendlier position than saying, in effect, “your insurance company is watching you.”

Private Equity, $350MM, March 24, 2026.

NoTraffic

NoTraffic replaces traditional traffic signal controllers with AI-powered units that adapt in real time, cutting average wait times by 20 to 40 percent in deployments across Phoenix and Houston.

Series C, $90MM, March 24, 2026.

Financial Services: The Back Office Gets a Brain (Again). $427MM

The financial services theme from last month’s report continues: every back-office function is getting an AI rewrite. Take the paperwork, give it to the agents, keep a human in the loop for the thorny cases.

9fin

London-based 9fin automates research for the bonds and loans market.

Series C, $170MM, March 31, 2026.

Next Technology Holding

Next Technology Holding (NASDAQ: NXTT) has product lines in hotel management monitoring, smart water systems, and AI-enabled liquid cooling for crypto mining.

Post-IPO Equity, $157MM, March 25, 2026.

Spade

Spade cleans up the raw transaction strings that flow through banks and fintechs, matching each one to a verified merchant with low latency. Customers include Stripe.

Series B, $40MM, March 24, 2026.

Worth AI

Worth automates the credit checks, KYC, KYB, and fraud assessment work that banks and fintechs do before approving a small business line of credit, processing over 1,200 data points per decision through an explainable AI engine.

Series A, $30MM, March 24, 2026.

Notch

Notch builds AI agents for regulated industries — starting with global insurers, large brokers, and financial institutions — where the auditability of each decision is important.

Series A, $30MM, March 25, 2026.

Education. $55MM

Kinder Institute for Urban Research

The Kinder Institute for Urban Research at Rice University received $55MM from the Kinder Foundation in March — the largest single gift to urban research in the United States in recent memory. The institute’s Houston Population Study is the longest-running urban study in America, now in its 44th year. The new funding will expand data collection, bring AI into the analysis, and extend the research to other metros. The unique value is time horizon: the Kinder Institute has tracked the same Houston neighborhoods since 1982.

Grant, $55MM, March 5, 2026.

Real Estate / Construction. $46MM

PAUL Tech

PAUL Tech is a Mannheim, Germany, company that sells heat-as-a-service to apartment building owners. The model is unusual: PAUL installs sensors on every radiator, a central AI control unit, and a retrofit of the building’s heating distribution, and the owner pays a monthly subscription instead of upfront capex. The AI runs predictive control on the building — learning each tenant’s schedule, weather patterns, and thermal mass — and typically cuts energy consumption by about 40 percent. The company’s pitch is that buildings are the largest source of carbon emissions in Europe, and retrofitting them at scale requires a financing model owners actually want.

Debt, $46.1MM, March 24, 2026.

Select Sources

Business Wire. “WHOOP Raises $575 Million at $10.1 Billion Valuation.” Business Wire, March 31, 2026. https://www.businesswire.com/

Business Wire. “Legora Secures $550 Million Series D at $5.55 Billion Valuation.” Business Wire, March 10, 2026. https://www.businesswire.com/

Marina Temkin. “Legora raises $550M at $5.55B valuation.” TechCrunch, March 10, 2026. https://techcrunch.com/

Kenrick Cai. “VAST Data raises $1B at $30B valuation.” Reuters, March 9, 2026. https://www.reuters.com/technology/

Business Wire. “CMT Raises $350M from TPG Rise Climate, Allianz X, State Farm.” Business Wire, March 24, 2026. https://www.businesswire.com/

South China Morning Post. “PixVerse-maker AIsphere raises $300 million Series C.” SCMP, March 12, 2026. https://www.scmp.com/

TechCrunch. “Lightwheel raises $145.5M Series A for robotics simulation.” TechCrunch, March 11, 2026. https://techcrunch.com/

Alto Neuroscience. “Alto Neuroscience Announces $120 Million PIPE Financing.” Alto IR, March 16, 2026. https://investors.altoneuroscience.com/

Alex Konrad. “AI law firm Crosby raises $60M at $400M valuation.” Forbes, March 31, 2026. https://www.forbes.com/

Kinder Institute. “Kinder Foundation Announces $55 Million Gift.” Rice University, March 5, 2026. https://kinder.rice.edu/

Business Wire. “Standard Template Labs Launches With $49 Million in Seed Funding.” Business Wire, March 17, 2026. https://www.businesswire.com/

PAUL Tech. “PAUL Tech secures €42.4 million in debt financing.” PAUL Tech News, March 24, 2026. https://www.paul.tech/

Tech in Asia. “Deep Intelligent Pharma raises $40M, hits $150M total.” Tech in Asia, March 24, 2026. https://www.techinasia.com/

Reuters. “Shield AI raises $2 billion at $12.7 billion valuation.” Reuters, March 26, 2026. https://www.reuters.com/technology/

Reuters. “Yann LeCun’s AMI Labs raises $1B in seed round at $3.5B valuation.” Reuters, March 10, 2026. https://www.reuters.com/technology/

Bloomberg. “Airtel’s Nxtra Raises $1 Billion from Alpha Wave, Carlyle.” Bloomberg, March 30, 2026. https://www.bloomberg.com/

TechCrunch. “Cloaked raises $375M Series B from General Catalyst.” TechCrunch, March 19, 2026. https://techcrunch.com/

Business Wire. “TENEX.AI Closes $250 Million Series B at Unicorn Valuation.” Business Wire, March 31, 2026. https://www.businesswire.com/

Business Wire. “Grafana Labs Raises $250 Million Series E at $6 Billion Valuation.” Business Wire, March 12, 2026. https://www.businesswire.com/

Bloomberg. “Harvey Hits $11 Billion Valuation in $200M Series G.” Bloomberg, March 25, 2026. https://www.bloomberg.com/

Financial Times. “9fin raises $170M Series C at $1.3B valuation.” Financial Times, March 31, 2026. https://www.ft.com/

GlobeNewswire. “Next Technology Holding Inc. Announces $157 Million Registered Direct Offering.” GlobeNewswire, March 25, 2026. https://www.globenewswire.com/

Business Wire. “Qualified Health Raises $125M Series B Led by NEA.” Business Wire, March 25, 2026. https://www.businesswire.com/

TechCrunch. “Granola raises $125M Series C at $1.5B valuation.” TechCrunch, March 25, 2026. https://techcrunch.com/

SecurityWeek. “Galina Antova launches Kai with $125M from Evolution Equity.” SecurityWeek, March 10, 2026. https://www.securityweek.com/

South China Morning Post. “Horizon Robotics spinoff D-Robotics raises $120M Series B.” SCMP, March 16, 2026. https://www.scmp.com/

Business Wire. “Dash0 Raises $110M Series B at $1B Valuation.” Business Wire, March 23, 2026. https://www.businesswire.com/

VentureBeat. “ORO Labs raises $100M Series C for AI procurement platform.” VentureBeat, March 12, 2026. https://venturebeat.com/

Business Wire. “NoTraffic Raises $90M Series C Led by PSG Equity.” Business Wire, March 24, 2026. https://www.businesswire.com/

Reuters. “Elite Robots raises $87M Series D for industrial cobots.” Reuters, March 25, 2026. https://www.reuters.com/

TechCrunch. “Nominal hits $1B valuation, raises $155M in 10 months.” TechCrunch, March 5, 2026. https://techcrunch.com/

SecurityWeek. “Depthfirst Raises $80 Million in Series B Funding.” SecurityWeek, April 1, 2026. https://www.securityweek.com/

TechCrunch. “Sequoia-backed Juicebox lands $80M.” TechCrunch, March 11, 2026. https://techcrunch.com/

TechCrunch. “Qodo raises $70M for code verification as AI coding scales.” TechCrunch, March 30, 2026. https://techcrunch.com/

Sharon Goldman. “Exclusive: AI startup Axiamatic raises $54 million.” Fortune, March 11, 2026. https://fortune.com/

TechCrunch. “Gumloop lands $50M from Benchmark.” TechCrunch, March 12, 2026. https://techcrunch.com/

TechTarget. “Qdrant raises $50M in funding to fuel vector database growth.” TechTarget, March 12, 2026. https://www.techtarget.com/

PR Newswire. “Normal Computing Raises $50M Led By Samsung Catalyst.” PR Newswire, March 25, 2026. https://www.prnewswire.com/

SecurityWeek. “Linx Security Raises $50 Million for Identity Security.” SecurityWeek, March 31, 2026. https://www.securityweek.com/

Pandaily. “VAST Secures $50 Million Series A.” Pandaily, March 6, 2026. https://pandaily.com/

PR Newswire. “Tripo AI Announces $50 Million in Funding.” PR Newswire, March 25, 2026. https://www.prnewswire.com/

Business Wire. “Thesis Care Announces $45 Million Series A.” Business Wire, March 25, 2026. https://www.businesswire.com/

SiliconANGLE. “Cylake launches with $45M.” SiliconANGLE, March 5, 2026. https://siliconangle.com/

Korea Herald. “LG CNS takes stake in Silicon Valley robot startup Dexmate.” Korea Herald, March 10, 2026. https://www.koreaherald.com/

SiliconANGLE. “Ease Health launches with $41M.” SiliconANGLE, March 2, 2026. https://siliconangle.com/

PR Newswire. “Adonis Raises $40M Series C.” PR Newswire, March 25, 2026. https://www.prnewswire.com/

Business Wire. “Highlight AI Raises $40 Million Series A.” Business Wire, March 24, 2026. https://www.businesswire.com/

Business Wire. “Spade Announces $40M Series B.” Business Wire, March 24, 2026. https://www.businesswire.com/

Casey Ross. “Doctronic raises $40 million.” STAT News, March 23, 2026. https://www.statnews.com/

Sharon Goldman. “RunSybil raises $40 million.” Fortune, March 18, 2026. https://fortune.com/

TechCrunch (via Lace AI). “Lace AI raises $40M Series A.” TechCrunch, March 23, 2026. https://techcrunch.com/

Globe and Mail. “Spellbook secures $40-million from RBC.” The Globe and Mail, March 4, 2026. https://www.theglobeandmail.com/

SiliconANGLE. “Onyx Security launches with $40M.” SiliconANGLE, March 12, 2026. https://siliconangle.com/

SecurityWeek. “AI Security Firm JetStream Launches With $34 Million in Seed Funding.” SecurityWeek, March 4, 2026. https://www.securityweek.com/

European Biotechnology Magazine. “Owkin spins out Waiv with $33M.” European Biotechnology, March 13, 2026. https://european-biotechnology.com/

GlobeNewswire. “Spectral AI Awarded $31.7 Million of BARDA Funding.” GlobeNewswire, March 18, 2026. https://investors.spectral-ai.com/

PYMNTS. “Worth Raises $30 Million.” PYMNTS, March 24, 2026. https://www.pymnts.com/

GlobeNewswire. “Notch Raises $30 Million.” GlobeNewswire, March 25, 2026. https://www.globenewswire.com/

TechCrunch. “Lio raises $30M from Andreessen Horowitz.” TechCrunch, March 5, 2026. https://techcrunch.com/

CTech (Calcalist). “Fig Security raises $30 million Series A.” CTech, March 3, 2026. https://www.calcalistech.com/

Validio. “Validio raises $30M Series A led by Plural.” Validio blog, March 5, 2026. https://validio.io/

CTech. “Sett secures $30 million Series B.” CTech, March 31, 2026. https://www.calcalistech.com/

Business Wire. “Qevlar AI Raises $30M.” Business Wire, March 10, 2026. https://www.businesswire.com/

St. Louis Business Journal. “Capacity advances Series E fundraise.” St. Louis Business Journal, March 2026. https://www.bizjournals.com/stlouis/

Anthropic. “Progress Report on Frontier AI Cybersecurity Capabilities.” Anthropic Blog, April 2026. https://www.anthropic.com/

AISI. “Pre-Deployment Evaluation of Mythos.” UK AI Safety Institute, April 2026. https://www.aisi.gov.uk/

CSO Online. “State-Linked Group Used Claude Code in Automated Cyber-Espionage Campaign.” CSO Online, November 2025. https://www.csoonline.com/

EDITOR’S NOTE

As a leader today you’re snowblind with AI news--but gawping at somebody else’s trillion dollar valuation won’t help you grow your business. At the DataStory we believe the best way to spark your innovative spirit is through stories--case studies--that show you in detail how companies on the bleeding edge of this AI market are doing it. Who are they serving? With what data? How do they put that data to work with analytics and AI? And most importantly--what can you learn from them?

In our case study this month, we profile Profound, a rare hot AI startup that’s based in New York, solving a very New York problem--that is, an advertising and marketing problem: How do brands adapt to a world of AI search?

What is AI search?

It’s when you don’t perform a traditional Google search to find a reliable roofer, or how to treat chronic anxiety, or how many calories in a White Claw. You ask ChatGPT. Or Perplexity, or Claude. Or you just get your browser’s AI summary.

You don’t click through to the webpage providing the originating information. This is called “Zero-Click Search” because people generally don’t click to the sources. We search, we read, we’re done. This phenomenon--that we all experience everyday now--is creating a multi-headed crisis.

First, Google and Bing search are getting disrupted--a huge deal to the marketing and ads ecosystem. Search ads command a 33% share of all US digital ads--an enormous slice--and an estimated $190B in the US in 2025. As of now, AI platforms comprise about 1% of all website referral traffic--seems small, but that’s off a huge and important base--and Gartner predicts traditional search will fall by 25%, yielding to AI search.

Crisis #2 is in the publishers whose traffic is being choked off, way up at the source.

The third crisis is to the brands--the roofer, the pharma company, the beverage company--whose products are not being surfaced in the AI search ecosystem--or at least, not in the way those companies are used to.

Brands spent thirty years figuring out SEO and currently spend (estimates vary) $100 billion a year on SEO software and services, getting web users to click on their blue link and their search ad. The game is changing.

This is the problem Profound is trying to address: helping brands navigate AI search. Investors seem to think this is a big deal. In 18 months Profound went from a four-person team to a billion-dollar valuation by building the most prominent platform to help brands monitor and improve how they appear in AI-generated responses. In the process they are defining the Answer Engine Optimization, or AEO, category.

Profound is not alone. Peec AI, Otterly AI, AthenaHQ, Scrunch AI, Semrush, and Ahrefs all offer related services. But today we’re focusing on Profound because of their big raise reported in February ($96MM), and more importantly, because we want to understand how they use data and AI to provide their services to brands.

In all of our case studies, The DataStory uses our own “7Ds of Data Innovation” framework to analyze how companies use data and AI. What does this look like for Profound?

COMPANY OVERVIEW

Profound is an enterprise software platform that helps brands understand and control how they appear in AI-generated search results. When a consumer asks ChatGPT which running shoes to buy, or asks Perplexity to compare CRM tools, Profound shows the brand if it was mentioned, what was said, and how to improve its position.

Date of Founding: 2024.

Founding Story: James Cadwallader had founded Kyra Media, a video content company focused on TikTok and YouTube. Dylan Babbs (CTO) came from maps and location engineering at Uber. The two met at South Park Commons, a community for early stage startups. In early 2024, they started asking a question: when someone types “What is the best running shoe?” into ChatGPT, which brand shows up — and why? Running hundreds of queries across ChatGPT, Perplexity, and Google’s AI answers, they found something alarming: Google rankings didn’t translate to AI visibility. Companies could spend millions on SEO and still be absent from the answers their customers were reading. No dashboard, no measurement tool, no playbook existed. Marketing executives confirmed it was a real and growing problem. By August 2024, the duo emerged from stealth with a working product and a seed round closed.

Total Investment Raised to Date: $155 million total. Seed: $3.5M (August 2024, led by Keith Rabois / Khosla Ventures). Series A: $20M (June 2025, led by Kleiner Perkins). Series B: $35M (August 2025, led by Sequoia Capital). Series C: $96M (February 2026, led by Lightspeed Venture Partners) at a $1 billion valuation.

What the Company Does: Profound gives marketing teams a dashboard to see how their brand appears across all major AI answer engines—ChatGPT, Perplexity, Google AI Overviews, Gemini, Claude, Grok, Microsoft Copilot, Meta AI, and DeepSeek.

The platform tracks how often the brand is mentioned; what the AI says about it; which web sources the AI cites; and how the brand compares to competitors.

Beyond monitoring, Profound now offers autonomous AI Agents that help marketers act on the data. These Agents can draft content, optimize messaging for AI crawlers, generate competitive reports, and run AEO campaigns.

How the Company Makes Money: Profound operates on a SaaS subscription model with multiple tiers such as: Starter: $99/month with basic prompt tracking on ChatGPT. Growth: $399/month with multi-engine coverage, competitive analysis, and Profound Agents. Enterprise: custom pricing with SSO, API access, role-based permissions, and dedicated support.

Top Named Clients: They claim over 700 clients. Named customers include Target, Walmart, Figma, MongoDB, Ramp, Chime, U.S. Bank, DocuSign, Plaid, Indeed, Deel, and Charlotte Tilbury.

1. DEMAND

Who is their user?

Priority User 1: The Head of Digital Marketing / VP of Marketing

This person runs the digital strategy for a mid-to-large-sized brand. Their job is to make sure customers can find the company online--whether through search, social media, or paid advertising. They manage a team that produces content, runs campaigns, and tracks performance metrics like website traffic, conversion rates, and cost per acquisition.

This person typically sits at a company with annual revenue above $100 million. They report to the CMO. Their budget decisions are scrutinized quarterly.

For twenty years, this job was, in part, organized around Google. The VP of Marketing knew the rules: invest in SEO to rank in organic results, buy Google Ads to appear in paid results, and track everything through Google Analytics.

Priority User 2: The Content / SEO Team Lead

This person manages day-to-day content creation and search optimization. They decide what blog posts to write, which keywords to target, and how to structure pages so Google ranks them highly. They work closely with writers, agencies, and analytics platforms like Semrush and Ahrefs.

LEARNING: Profound’s users are at the direct-response and quant-side of marketing, focused on how much investment is required to get what level of results.

2. DILEMMA

What makes achieving that goal difficult?

Obstacles for the VP of Marketing

The biggest challenge is visibility. In traditional search, a brand could check its Google ranking in seconds. They need data to justify new spending, and right now they are staring at a channel—AI search—that is growing fast but almost entirely unmeasured.

Another problem is attribution. Even when a brand suspects AI is driving sales—because customers mention ChatGPT in surveys, for example—there is no clean way to prove it.

Obstacles for the Content / SEO Team Lead

The content lead faces a more tactical dilemma. Traditional SEO has clear best practices: use the right keywords, build a network of links. AI search optimization does not have a settled playbook. Each AI platform has different preferences for the sources it trusts and cites. Content that performs well on one engine might be invisible on another.

The pace of change compounds the difficulty. AI models update constantly. What works this month may not work next month. Content teams must now maintain optimization across 10 or more AI engines simultaneously, while still managing their traditional SEO workload.

LEARNING: Digital marketing leaders need numbers, and results, to do their jobs and to justify marketing spend. AI search is opaque--leaving marketers without tools to either be effective, measure ROI in AI search, or justify budget decisions.

3. DATA

What data can be accessed to achieve the goal?

Profound’s platform runs on three distinct data pipelines. Each one collects a different type of information, from a different source. Together they give brands a picture of how they show up in AI search.

Data Type 1: Conversation Data They License

The first pipeline is bulk data that Profound purchases about what real people are actually asking AI platforms. Profound gathers “hundreds of millions of conversations every month from third-party data providers.” A company blog post explains the economics: early on, the team did not have much data and “couldn’t afford to buy that much of it.” As the company raised more capital, it scaled up these purchases.

Where does this data come from? Profound licenses anonymized conversations from “double opt-in consumer panels of real AI users.” All data is scrubbed of personal information. Real consumers have agreed to share their AI conversations anonymously. Profound does not name the specific panel providers.

I don’t know exactly what this means, but my interpretation is the following. If you’re a panel data provider, that means you are offering an incentive to thousands of people to sell you their data. You might have a “router sniffer” on their home wifi, that’s basically tracking all the internet traffic coming into and out of the home. Or, you might have this setup on the panelist’s phone, as an app that does the tracking. You’re paying them for the data--they sign a contract and they get whatever you’re paying them: a certain amount of cash per month, or a shot at a sweepstakes, or barter like discounts on goods and services.

Then--again I’m guessing--the panel company, or Profound itself, filters the web activity for records of the moments where the user pulled up an AI engine and performed a search. What Profound would be looking at is the data associated with me typing into ChatGPT “hwo many claories in a white claw” and looking at all the responses. Or better, I ask “hwo many claories in a hard seltzer” and I get a response that mentions White Claw, Truly, and High Noon. And now those answers get added to Profound’s count of “citations” for each of those brands.

Then, you do that thousands and thousands of times for all the panelists, then for all the user’s searches, and you get the millions number. The advantage of this data is that it’s opt-in, it’s a representative sample (the panel companies are careful to recruit panelists proportionate to the population), and it’s “in the wild,” which is to say, unbiased--presumably no one in the panelist household is pulling their punches on their GPT searches to appear more virtuous--for instance to hide their overweening interest in White Claw. (Asking for a friend.)

Data Type 2: Queries They Generate

The second pipeline is data that Profound generates. The company sends millions of prompts per day to AI platforms and records the full responses. Profound runs these queries across the AI platforms: ChatGPT, Perplexity, Claude, Microsoft Copilot, Google AI Overviews, Google AI Mode, Google Gemini, Grok, Amazon Rufus, Meta AI, and DeepSeek.

This is classic data-making. I used to quip--back when data was made by humans--that all data businesses are built on a lot of people doing something really boring. Like there was this old newspaper advertising market share business called CMR which--I am told--employed people to sit around clipping newspaper ads from newspapers all day long, and then the employees would measure the column lengths with a ruler. (Imagine all that ink on your hands. What did they wear? Smocks?) Now the machines gather the data. Software fires all the queries at the AI search engines and records the responses. Presumably the machines don’t get bored and wear whatever they want.

In any case, Profound parses the AI responses in detail: in particular the citations. Which URLs are cited, how often, in response to which prompts, and for which brands. Over time, the company has built a database of billions of citations. This data reveals which content assets are driving AI visibility and which competitor content is outperforming the brand.

Data Type 3: Client Data from their Own Websites

The third pipeline flips the direction to look at the interaction with the AI engines from the perspective of the brand’s own website. When a ChatGPT, Perplexity, or another AI agent’s bot scrapes a company’s web pages, Profound detects their crawlers and their visits. They record which pages were accessed, which AI bots were involved, and how frequently the bots returned.

Presumably Profound is looking for patterns--what content do the bots gravitate toward? Which engine seems to like what kind of words? All in the interest of steering the bots on the client’s behalf, toward whatever content the client most wants discovered; recording the highest traffic.

LEARNING: Profound uses what I would call a classic (in a good way) approach to data: a diversity of sources. There’s data they buy (where they apply their own “take” and vision to what is a commodity input); there’s data they make (which becomes IP proper); and data they get from the client, to provide additional value to the customer.

4. DERIVATION

How do analytics/AI make the data usable?

Profound processes its raw data through three distinct layers. Here is how each layer works.

Layer 1: Traditional Scoring, Benchmarking, and Trending

Profound runs straightforward analytics on their data. Classic counting. The most important metric is share of voice. Profound counts how many times a brand appears in AI-generated answers for a given set of prompts; then divides that by the total number of responses tracked; to get the brand’s share of the AI conversation in their category. This is benchmarked against competitors and tracked over time, so a CMO can see whether the brand is gaining or losing ground.

Then there’s “citation tracking”--identifying the websites that generate the most AI responses. The URLs are counted, deduplicated, categorized by domain and content type (eg blog posts, FAQs, product pages), and tracked over time.

Sentiment analysis adds a qualitative layer. Profound categorizes whether each mention is positive, neutral, or negative, and also flags it when the metric flips from positive to negative. If an AI platform starts describing a brand’s product unfavorably after a competitor launches a new feature, the dashboard surfaces that change.

These metrics are basic counts, percentages, trends—the same type of structured reporting that marketers rely on in classic marketing data. Their value is, they bring that familiar analytical rigor to a channel that previously had no measurement at all.

Layer 2: AI Reasoning Models

The second layer uses AI reasoning models—LLMs trained to work through complex problems step by step. The reasoning models analyze patterns across hundreds of thousands of AI citations to understand what makes the highest-performing content work.

An example: a of scan the top-cited pages might reveal that FAQ pages with concise, direct answers… for whatever reason… get cited 3x more often than long-form blog posts. These are observations a human analyst could not make at scale.

Entity extraction is another AI task. In this context it means identifying every brand, product, and organization mentioned in AI search text, and then normalizing and consolidating all the entities mentioned: e.g., the model must recognize that “Slack,” “Salesforce’s messaging tool,” and “the team communication platform” might all refer to the same entity. Sometimes this is called “Named Entity Recognition” which is a little more intuitive.

Layer 3: Generative AI

This is a new offering, and the first one that gets the company out of using AI to interpret data, and into helping clients respond in an automated way. In February 2026, Profound launched agents that use the analytics (mentioned in Layer 1) to generate actual marketing content. They research what AI platforms are picking up for a given topic, and then write content structured to match those citation patterns; track whether the output earns more citations; and optimize.

LEARNING: Profound uses a variety of data and AI techniques for the full cycle of their business: they use data and plain old counting; they use AI for extracting complicated insights from lots of text data; then employ a different mode of AI (generative) for developing work from the metrics. A reminder that creating strong intellectual property does not necessarily rely on a single stroke of wizardry. It can be a string of medium-clever applications across a value chain.

5. DELIVERY

How does the workflow improve for the user?

Profound’s interface is a web-based dashboard that a marketing team logs into daily, much like they would use Google Analytics or Semrush.

Insights shows how the brand appears across all tracked AI platforms. The marketer can see share of voice, citation trends, sentiment scores, and the specific prompts where their brand appears (or does not). Then “Prompt Volumes” identify high-volume questions in their category.

“Agent Analytics” provides the client view, showing which AI crawlers visit the brand’s website, what pages they read, and how frequently. Finally there is Profound Agents that draft optimized blog posts (or FAQs or what have you) based on citation analysis.

Profound integrates with enterprise systems like HubSpot, Google Workspace, Gamma, Vercel, and Parallel AI.

The manual workflow this replaces was painful. Before, a marketer who wanted to check their AI visibility had to manually type prompts into ChatGPT, Perplexity, and other platforms, one at a time, and record results in a spreadsheet. Doing this at scale was basically impossible.

LEARNING: Because AEO wasn’t really a thing before, it’s not making a current process more efficient; more like it’s creating a proper toolkit for a new need. The first mover advantage may be an advantage here--or might open space for a competitor to learn from their mistakes.

6. DECISION

What decision is enabled by using the data?

Profound enables three categories of decisions for AI search marketers:

Content prioritization. Prioritize content that builds a brand’s citation visibility in AI search.

Competitive response. When a competitor gains AI visibility in a key product category, the brand can see it happen and respond.

Budget allocation. The VP of Marketing can now present data on AI search share-of-voice to leadership and make the case for investing in AEO alongside traditional SEO and paid search.

LEARNING: This reminds us that, despite the AI luster, Profound’s is still a classic metrics business. Not much different than an old market share metric: get visibility into your competitive space, and by doing so equip yourself to take action.

7. DESTINATION

What does success look like?

Brands who use Profound’s tools are getting a window on a new ecosystem. Before, these brands lived in the land of old school search; now they live in the world of AI engines; and Profound’s metrics show the way.

In my book I called this a Dark Room problem.

It’s akin to information asymmetry. You have an opportunity; you can’t attack it if you don’t understand it--especially when you don’t understand the risks. Data that sheds light on the unknown… the dark spaces…for instance by scoring a creditor, or pricing a used car… helps to make a market. Even open up new spaces for innovation.

Here, the new market (AI search) came over the brands without their willing participation. Darkness falling. The metrics from Profound help shine the light on that new ecosystem. The whole problem sounds very marketing-y and Madison Ave and abstract until you put yourself in the position of that roofer whose phone suddenly stops ringing--no leads. Then, you want the light. Your livelihood depends on it.

Worth noting, for our readers, is how the company is working to sustain their leadership status in this new domain: they have launched “Profound University” for education and training in their services. And with their latest raise they claim they will establish certifications in AI search, as well as a marketplace for partners who are approved agent-makers.

In doing so they’re building community, services, a marketplace, around their leadership. The Trade Desk (TTD) did this with their “Edge Academy,” helping professionals understand the gnarly world of digital and programmatic advertising. As their former competitor, I always thought it was a brilliant stroke. It was probably a considerable investment. But the investment kept paying for itself in earned media, as spontaneous social posts popped up saying “I’ve been certified in X!” with the Trade Desk logo painted all over it. In Profound’s case, the more they cast themselves as AI search experts, the greater the sparkle on their market position.

LEARNING: When you create a marketplace, you are creating economic opportunity for your customers. It’s worth doubling down on that leadership with small acts of “selflessness”--eg with education and marketplaces--that help your clients, while further establishing your authority.

8. GLOSSARY

Answer Engine Optimization (AEO): The practice of structuring content so AI-powered platforms cite it in their generated answers. The AI-era equivalent of SEO.

Answer Engine: An AI system that generates a direct answer rather than showing a list of links. ChatGPT, Perplexity, and Google AI Overviews are answer engines.

Share of Voice (AI): The percentage of AI-generated responses in a category that mention a specific brand. Higher share of voice means the brand appears more frequently.

Prompt Volume: An estimate of how many users ask a specific question to AI platforms. Similar to keyword search volume in traditional SEO, but for AI conversations.

AI Crawler / AI Agent: A bot from an AI platform that visits websites to gather information used in generating answers. Similar to Google’s web crawler.

Citation (in AI context): A reference to a specific web page that an AI system uses as a source. Being cited is the AI equivalent of ranking in Google search results.

Reasoning Model: A large language model trained to work through complex problems step by step. Used by Profound to analyze patterns across large datasets.

Zero-Click Search: A search where the user gets information directly from the results page or AI answer without clicking through to any website.

9. SOURCES & FURTHER READING

PROFOUND — COMPANY SOURCES (from tryprofound.com)

1. Profound blog, “$20M to pioneer Answer Engine Optimization,” June 18, 2025. https://www.tryprofound.com/blog/series-a

2. Profound blog, “$35M to connect brands with one new customer: Superintelligence,” August 12, 2025. https://www.tryprofound.com/blog/series-b

3. Profound blog, “Profound raises $96M Series C at $1B valuation,” February 24, 2026. https://www.tryprofound.com/blog/profound-raises-96m-series-c

4. Profound blog, “The front door of the internet is changing.” https://www.tryprofound.com/blog/company

5. Profound Engineering Blog, “Data Science at Profound: How People Use Answer Engines.” https://www.tryprofound.com/engineering/data-science-at-profound-how-people-use-answer-engines

6. Profound blog, “Introducing the Profound Index,” November 5, 2025. https://www.tryprofound.com/blog/introducing-profound-index

7. Profound blog, “Seeing what customers see: Direct AI search engine monitoring vs. API limitations.” https://www.tryprofound.com/blog/seeing-what-customers-see-direct-ai-search-engine-monitoring-vs-api-limitations

8. Profound blog, “Introducing Agent Analytics.” https://www.tryprofound.com/blog/introducing-agent-analytics

9. Profound blog, “Introducing Profound’s Data-Driven Prompt Recommendation Engine,” July 15, 2025. https://www.tryprofound.com/blog/data-driven-prompt-recommendation

10. Profound blog, “Introducing Query Fanouts.” https://www.tryprofound.com/blog/introducing-query-fanouts

11. Profound blog, “Expanding analysis for query fanouts in Profound.” https://www.tryprofound.com/blog/expanding-analysis-for-query-fanouts-in-profound

12. Profound blog, “Introducing Profound Workflows,” December 11, 2025. https://www.tryprofound.com/blog/profound-workflows-public-beta

13. Profound blog, “Launching Profound Data Nodes for Agent Workflows,” January 26, 2026. https://www.tryprofound.com/blog/profound-data-nodes

14. Profound blog, “Introducing the v0 node for Profound Agents.” https://www.tryprofound.com/blog/introducing-the-v0-node-for-profound-agents

15. Profound blog, “Upgrade your content library with Content Optimization.” https://www.tryprofound.com/blog/introducing-content-optimization

16. Profound blog, “Expanding Content Optimization: Analyze Any Content, Anywhere.” https://www.tryprofound.com/blog/content-optimization-expansion

17. Profound blog, “Workflows are now Agents: January release roundup,” February 4, 2026. https://www.tryprofound.com/blog/workflows-are-now-agents-january-release-roundup

18. Profound blog, “LinkedIn is the most-cited domain for professional queries in AI search.” https://www.tryprofound.com/blog/linkedin-is-the-most-cited-domain-for-professional-queries-in-ai-search

19. Profound blog, “How to Track Your Brand Visibility in AI Search With Profound,” January 26, 2026. https://www.tryprofound.com/blog/how-to-track-your-visibility-in-ai-search

20. Profound blog, “Profound achieves SOC 2 Type 2 certification,” June 10, 2025. https://www.tryprofound.com/blog/soc-2-type-2

21. Profound blog, “Profound Agents now directly integrate with Contentful CMS.” https://www.tryprofound.com/blog/profound-agents-now-directly-integrate-with-contentful-cms

22. Profound blog, “Webflow now integrates with Profound Agents.” https://www.tryprofound.com/blog/webflow-now-integrates-with-profound-agents

23. Profound product page, “Answer Engine Insights.” https://www.tryprofound.com/features/answer-engine-insights

24. Profound product page, “Agents.” https://www.tryprofound.com/features/agents

25. Profound product page, “Enterprise.” https://www.tryprofound.com/enterprise

26. Profound product page, “Pricing.” https://www.tryprofound.com/pricing

27. Profound product page, “Design at Profound.” https://www.tryprofound.com/design

28. Profound Partners page. https://www.tryprofound.com/partners

29. Profound University (certification platform). https://university.tryprofound.com

30. Profound LinkedIn company page. https://www.linkedin.com/company/tryprofound

THIRD-PARTY PRESS & INVESTOR COVERAGE

31. Trishla Ostwal, “Profound Raises $3.5M to Manage Brands’ AI Search Presence,” Adweek, August 13, 2024. https://www.adweek.com/media/profound-funding-brands-ai-search/

32. Lily Mae Lazarus, “Exclusive: As AI threatens search, Profound raises $96 million to help brands stay visible,” Fortune, February 24, 2026. https://fortune.com/2026/02/24/exclusive-as-ai-threatens-search-profound-raises-96-million-to-help-brands-stay-visible/

33. GlobeNewsWire, “Profound Raises Series C at $1B Valuation to Lead a New Category of Marketing,” February 24, 2026. https://www.globenewswire.com/news-release/2026/02/24/3243475/0/en/Profound-Raises-Series-C-at-1B-Valuation-to-Lead-a-New-Category-of-Marketing.html

34. Sequoia Capital, “Partnering with Profound: Winning on the AI Stage,” August 12, 2025. https://sequoiacap.com/article/partnering-with-profound-winning-on-the-ai-stage/

35. SiliconANGLE, “Profound raises $96M at $1B valuation for AI discovery monitoring platform,” February 24, 2026. https://siliconangle.com/2026/02/24/profound-raises-96m-1b-valuation-ai-discovery-monitoring-platform/

36. The SaaS News, “Profound Raises $96M Series C at $1B Valuation,” February 25, 2026. https://www.thesaasnews.com/news/profound-raises-96m-series-c-at-1b-valuation

37. TipRanks, “Profound Hits $1 Billion Valuation With $96 Million Series C,” February 24, 2026. https://www.tipranks.com/news/private-companies/profound-hits-1-billion-valuation-with-96-million-series-c-to-scale-ai-driven-marketing-agents

38. Martech Edge, “Profound Hits $1B Valuation With $96M Series C to Build Agentic Marketing Infrastructure,” February 25, 2026. https://martechedge.com/news/profound-hits-1b-valuation-with-96m-series-c-to-build-agentic-marketing-infrastructure

39. Yahoo Finance / GlobeNewsWire, “Profound Raises Series C at $1B Valuation,” February 24, 2026. https://finance.yahoo.com/news/profound-raises-series-c-1b-130000242.html

40. Wilson Sonsini, “Wilson Sonsini Advises Profound on $96 Million Series C,” February 24, 2026. https://www.wsgr.com/en/insights/wilson-sonsini-advises-profound-on-dollar96-million-series-c-at-a-dollar1-billion-valuation.html

41. The Aisle (Read the Aisle), “Profound founder James Cadwallader on the AI gold rush to reinvent SEO,” January 2, 2026. https://www.readtheaisle.com/p/profound-founder-james-cadwallader-aeo-geo-aio-seo

42. Wix AI Search Lab, “James Cadwallader from Profound on AI’s Impact on Marketing & SEO,” January 29, 2026. https://www.wix.com/studio/ai-search-lab/james-cadwallader

43. Double Diamond NYC, “Dylan Babbs – Co-Founder & CTO of Profound,” Substack, March 2026. https://doublediamondnyc.substack.com/p/dylan-babbs-co-founder-and-cto-of-profound

44. Shoptalk Spring 2026, James Cadwallader speaker page. https://spring.shoptalk.com/speaker-list/james-cadwallader

COMPANY DATA & PROFILES

45. PitchBook, “Profound 2026 Company Profile.” https://pitchbook.com/profiles/company/633268-54

46. Crunchbase, “Profound Company Profile & Funding.” https://www.crunchbase.com/organization/profound-1b0a

47. Dealroom.co, Profound company profile. https://app.dealroom.co/companies/profound1

48. Getlatka.com, “How Tryprofound hit $6.8M revenue.” https://getlatka.com/companies/tryprofound

49. South Park Commons, Profound company page. https://www.southparkcommons.com/companies/profound

50. Sequoia Capital, James Cadwallader founder page. https://sequoiacap.com/founder/james-cadwallader/

51. G2 Best Software Products 2026 (Profound ranked #34).

52. Dylan Babbs LinkedIn. https://www.linkedin.com/in/babbsdj/

53. James Cadwallader LinkedIn. https://www.linkedin.com/in/jsca/

54. James Cadwallader Twitter/X. https://x.com/thejamescad

AEO COMPETITORS & LANDSCAPE

55. SE Ranking, “8 Profound Alternatives for AI Visibility Tracking and Answer Engine Optimization,” January 5, 2026. https://seranking.com/blog/profound-alternatives/

56. SE Ranking / Visible, “9 Best Peec AI Alternatives Tested,” December 16, 2025. https://visible.seranking.com/blog/peec-ai-alternatives/

57. xSeek, “Best Profound Alternatives for AEO in 2026.” https://www.xseek.io/blogs/articles/best-profound-alternatives

58. Discovered Labs, “Profound vs Peec vs Otterly: Which AI Visibility Platform Should You Buy?” December 18, 2025. https://discoveredlabs.com/blog/profound-vs-peec-vs-otterly-which-ai-visibility-platform-should-you-buy

59. Quattr, “7 Peec AI Alternative Tools for AEO Tracking & Visibility.” https://www.quattr.com/blog/peec-ai-alternatives

60. Useomnia, “10 Best Profound AEO Alternatives for 2026.” https://www.useomnia.com/blog/profound-aeo-alternatives

61. ZipTie.dev, “5 Best Peec AI Alternatives for AI Traffic Monitoring in 2026.” https://ziptie.dev/blog/best-peec-ai-alternatives/

62. Getairefs, “12 Best AEO Tools in 2026.” https://getairefs.com/blog/best-aeo-tools/

63. Scrunch, “The 7 best AEO/GEO tools for 2026.” https://scrunch.com/blog/best-answer-engine-optimization-aeo-generative-engine-optimization-geo-tools-2026

64. Peec AI homepage. https://peec.ai/

AEO MARKET SIZE & TRENDS

65. QY Research, “Answer Engine Optimization Market Report,” 2026. Global AEO market ~$1.1B (2025), projected $9B+ by 2031 at 42% CAGR. https://www.qyresearch.com/news/14451/answer-engine-optimization

66. HubSpot, “Answer engine optimization trends in 2026.” https://blog.hubspot.com/marketing/answer-engine-optimization-trends

67. CXL, “Answer Engine Optimization (AEO): The comprehensive guide for 2026.” https://cxl.com/blog/answer-engine-optimization-aeo-the-comprehensive-guide/

68. O8 Agency, “Answer Engine Optimization Guide 2025.” https://www.o8.agency/blog/ai/answer-engine-optimization-guide

69. First Page Sage, “The Top Answer Engine Optimization (AEO) Companies of 2026.” https://firstpagesage.com/seo-blog/the-top-answer-engine-optimization-aeo-companies/

70. LLMrefs, “Answer Engine Optimization (AEO): The Complete Guide for 2026.” https://llmrefs.com/answer-engine-optimization

71. Evergreen Media, “Answer Engine Optimization (AEO): AI visibility in 2026,” February 12, 2026. https://www.evergreen.media/en/guide/answer-engine-optimization/

72. Stormy AI, “Answer Engine Optimization (AEO) in 2026: Winning Search When AI Agents Do the Browsing.” https://stormy.ai/blog/answer-engine-optimization-aeo-2026-strategy

73. Conductor, “What is Answer Engine Optimization? Enterprise Guide to AEO.” https://www.conductor.com/academy/answer-engine-optimization/

AI SEARCH STATISTICS (ChatGPT usage, referral traffic, market share)

74. Digiday, “In Graphic Detail: The state of AI referral traffic in 2025,” December 22, 2025. https://digiday.com/media/in-graphic-detail-the-state-of-ai-referral-traffic-in-2025/

75. Exposure Ninja, “AI Search Statistics for 2026: CMO Cheatsheet,” February 27, 2026. https://exposureninja.com/blog/ai-search-statistics/

76. Position Digital, “90+ AI SEO Statistics for 2025 (Updated November).” https://www.position.digital/blog/ai-seo-statistics/

77. Similarweb, “Generative AI Statistics for 2026.” https://www.similarweb.com/blog/marketing/geo/gen-ai-stats/

78. ALM Corp, “ChatGPT Now Has 12% of Google’s Search Volume—But Sends 190X Less Traffic,” February 18, 2026. https://almcorp.com/blog/chatgpt-12-percent-google-search-volume-190x-less-traffic/

79. Digital Applied, “AI Search and SEO Statistics 2026: Definitive Guide.” https://www.digitalapplied.com/blog/ai-search-seo-statistics-2026-definitive-collection

80. Vertu / Similarweb, “AI Chatbot Market Share 2026: ChatGPT vs Gemini,” January 26, 2026. https://vertu.com/lifestyle/ai-chatbot-market-share-2026-chatgpt-drops-to-68-as-google-gemini-surges-to-18-2/

81. Lantern, “ChatGPT Drives 87% of AI Referral Traffic.” https://www.asklantern.com/blogs/chatgpt-drives-87-of-ai-referral-traffic

82. Superlines, “AI Search Statistics 2026: 60+ Data Points on Visibility, Citations, and Traffic.” https://www.superlines.io/articles/ai-search-statistics/

83. The Digital Elevator, “35 ChatGPT User Statistics for 2026,” February 18, 2026. https://thedigitalelevator.com/blog/chatgpt-statistics/

84. Pew Research Center, study on AI Overviews reducing clicks (cited in multiple AEO guides). https://www.pewresearch.org/short-reads/2025/07/22/google-users-are-less-likely-to-click-on-links-when-an-ai-summary-appears-in-the-results/

85. Gartner, “25% of organic search traffic will shift to AI chatbots by 2026” (cited in multiple industry reports). https://www.gartner.com/en/newsroom/press-releases/2024-02-19-gartner-predicts-search-engine-volume-will-drop-25-percent-by-2026-due-to-ai-chatbots-and-other-virtual-agents

SEARCH ADVERTISING MARKET

86. Statista, “Search Advertising – Worldwide Market Forecast.” $355.1B in 2025. https://www.statista.com/outlook/amo/advertising/search-advertising/worldwide

87. Precedence Research, “Digital Advertising Market Size 2025 to 2035.” Search advertising = 33% of digital ad market. https://www.precedenceresearch.com/digital-advertising-market

88. Marketing LTB, “Digital Advertising Statistics 2025: 92+ Stats.” Search ads ~40% of digital budgets, $295B globally. https://marketingltb.com/blog/statistics/digital-advertising-statistics/

89. Abbey Mecca, “Global Ad Spending to Hit $1 Trillion in 2025.” Digital = 75.2% of total. https://abbeymecca.com/worldwide-ad-spending-to-hit-1-trillion-in-2025-what-it-means-for-businesses-marketers-and-nonprofits/

90. Cropink, “50+ Advertising Statistics That Are Changing the Game [2026].” Search = ~40% of digital budgets. https://cropink.com/advertising-statistics

91. IMARC Group, “Advertising Market Size, Share | Industry Report 2034.” Global market $706.4B (2025). Internet ads = 34.7%. https://www.imarcgroup.com/global-advertising-market

92. Research and Markets, “Advertising Market Size, Competitors & Forecast to 2032.” $729.45B (2025). https://www.researchandmarkets.com/report/advertising-marketing

93. Precedence Research, “Digital Ad Spending Market Size to Hit USD 1,593 Billion by 2035.” https://www.precedenceresearch.com/digital-ad-spending-market

94. WARC, “Google Q4/FY 2025: Strong ad growth.” Alphabet crossed $400B annual revenue. https://www.warc.com/content/feed/google-q4-fy-2025-strong-ad-growth-masks-investor-nerves-over-alphabets-ai-bill/en-GB/11311

95. Storyboard18, “Alphabet Q3 revenue soars to $102.3B; Google Search ads hit $56.6B.” https://www.storyboard18.com/advertising/alphabet-q3-revenue-soars-to-102-3-billion-google-search-ads-hit-56-6-bn-youtube-ads-jump-to-10-3-bn-83414.htm

96. Marketing Dive, “Google’s full-stack AI approach drives Q3 search, YouTube ad revenue gains.” https://www.marketingdive.com/news/googles-full-stack-ai-approach-drives-q3-search-youtube-ad-revenue-gains/804234/

97. Bloomberg, “Alphabet Sales Beat Estimates on Google Search Advertising,” April 25, 2025. https://www.bloomberg.com/news/articles/2025-04-24/alphabet-earnings-beat-estimates-on-google-search-advertising

98. Alphabet Inc. SEC Filing, Form 8-K, Q1 2025. https://www.sec.gov/Archives/edgar/data/0001652044/000165204425000040/googexhibit991q12025.htm

99. Alphabet Inc. SEC Filing, Form 8-K, Q3 2025. https://www.sec.gov/Archives/edgar/data/0001652044/000165204425000087/googexhibit991q32025.htm

100. Analyzify / StatsUp, “Latest Alphabet (Google) Statistics 2025.” https://analyzify.com/statsup/alphabet

SEO MARKET SIZE & SPENDING

101. Mordor Intelligence, “SEO Services Market Size, Share, Industry Report 2031.” $83.98B (2026), projected $148.86B by 2031 at 12.12% CAGR. https://www.mordorintelligence.com/industry-reports/seo-market

102. Precedence Research, “SEO Software Market Size, Share and Trends 2026 to 2035.” $84.94B (2025), projected $295B by 2035. https://www.precedenceresearch.com/seo-software-market

103. Xamsor Blog, “SEO Market Stats (2026).” $108.28B (2026), GEO market $886M (2024) → $7.32B by 2031. https://xamsor.com/blog/seo-market-stats/

104. SaaSUltra, “SEO Statistics 2026: 50+ Trends Reshaping Search.” $83.9B services (2026). Organic = 53% of web traffic. https://www.saasultra.com/seo-statistics-trends/

105. Yahoo Finance, “SEO Statistics 2026: Market Size, Growth, and Key Industry Data.” 10–20% of digital budgets to SEO. https://finance.yahoo.com/sectors/technology/articles/seo-statistics-2026-market-size-111500051.html

106. AIOSEO, “85+ SEO Statistics for 2026.” $83.98B (2026). https://aioseo.com/seo-statistics/

107. Resourcera, “SEO Services Statistics (2025).” $74.9B (2025). U.S. spent $119.4B on SEO/digital marketing consulting. Median ROI 748%. https://resourcera.com/data/seo/seo-services-statistics/

108. Dataopedia, “30+ SEO Industry Statistics [2026].” $89.1B (2024), projected $143.9B by 2030. https://dataopedia.com/seo-industry-statistics/

109. Business Research Insights, “SEO Market Size, Growth | Report [2026-2035].” https://www.businessresearchinsights.com/market-reports/seo-market-117984

110. Global Risk Community / Research and Markets, “Top SEO Services Providers Worldwide in 2025.” $81.46B (2024), projected $171.77B by 2030. https://globalriskcommunity.com/notes/top-seo-services-providers-worldwide-in-2025-driving-a-171-77-bil

111. Research and Markets, “Digital Advertising Market Report 2026.” https://www.researchandmarkets.com/reports/5939640/digital-advertising-market-report

We’d love your comments and suggestions on this case study format! Please leave comments on the Substack platform—and recommend to others!

As we review the investments in data and AI companies… it’s time to feel confident about the direction of data and AI in business.

It’s also time to freak out.

First the investment basics: the month of February saw $174 billion in data and AI investment. Much of that was focused on general AI (Open AI raised $110B), another $21 billion on data centers and infrastructure, a small $1.4 billion in manufacturing (chips and cars)—and as usual we will focus on the “interesting” chunks—the $6.7B that went to specific use cases you can learn from.

Let’s talk about confidence. Data and AI use cases feel more solid than ever before. There is great news here for leaders. Whatever use case you have—or haven’t--dreamed up for your business, startups are hammering away at it.

And you can probably just rent the capability. I don’t see why a big insurance company or a medium-sized doctor practice should be expected to be good at data, AI, and their core business.

I perceive that the swarm of data and AI entrepreneurs—mainly in Silicon Valley but by no means all there--are hitting a groove. They are watching each other, learning, like some self-optimizing bio-borg. Making their applications more specific. Their KPIs harder-hitting, more obvious. Before—that is, olden times of 30 days ago—I feel like we were watching the performance of college theater, or a minor league ball game. Being patient with the lousy performances, waiting for them to settle into themselves.

One month later, we’re there. Less hand waving. More clarity. The trends below capture the upside of that. But one second…

Let’s talk about freaking out.

The flip side of confidence and clarity is when there seems to be no place data cannot go. Data definitely has its spooky side. We shed it without knowing. There’s no way to avoid this and still be part of modern life--as you wield your smart phone and make bank deposits and call Ubers and just walk around pinging GPS satellites.

Now sensors can be put anywhere, and even if you have opted out, you’re observable. Good for business applications, sure. But leaving less and less room for hallowed, unobserved spaces. As we get good, we get good at both the dark and the light.

Saddle up, friends, and learn what the last few weeks in dataspace has taught us…

TRENDS FOR MARCH 2026

Our four trends for the month, before we get into the sector briefings, and the company profiles, down below.

AUTOMATING ADMIN AND COMPLEX PAPERWORK. The big development is: we are getting better at the boring stuff. Boring is cool: this is a big idea of AI. Give the boring stuff to the robots. Save the smart thinking for the humans.

Automation is happening in a flood now, with financial services leading the way: Basis (accounting), Fieldguide (audits), Bretton AI (forensic financial analysis), Accrual (tax prep), Anterior (medical procedure pre-authorization), Jump (customer meeting notes for financial advisors), splose (patient meeting notes for doctors), Dwelly (landlords and renters).

We may never have to do anything boring again. Thought starter for leaders: if we can automate insurance workflows, we can automate your mess—whatever it is.

MARKETING BREAKS THE MOLD. I think we can all be forgiven for thinking that Search is kind of boring. It’s a 25 year old utility. It’s a monopoly. Or… was.

Now Search is blowing up and being replaced by AI answers. Every marketer will be forced to jump on this and reshape their every marketing move, and rely on a company like Profound.

I wonder if there is some reputation / social component to this as well. If AI search/summaries are capturing most public info about your company… that won’t be a narrow search metric. It will be a broad summary of your whole reputation. SEO has become Generative Engine Optimization (GEO) but I believe it will evolve to OOM—Overall Opinion Management. Get ready now.

CLING TO YOUR CUSTOMER. Many businesses rely on an in-person visit. Like healthcare. Because the main event happens IRL, there are constraints to your relationship with your provider.

But, if you strap a wearable to your arm or take pictures of your teeth or gather all your health records together, and give permission to a vendor to access that… all the time… then your doctor is basically over your shoulder. Which can be a good thing for your health. Lotus (for general health) and the blessedly simply-named Dental Monitoring (for… dental) are doing this.

What’s just as interesting to me is how that potentially slingshots into other areas of customer relationships. Think about this for, say automotive, appliances, electronics, housewares, services like accounting and legal and mental health and general practice medicine… What opportunities are available for better service if you link to your customer after they drive off the lot? Between visits? There’s a lot of value here.

MONITORING AT NEW LEVELS. If sensors are cheap and you can charge them with solar power anywhere and compute is fast enough to be pushed to the edge (in a compute sense)… then you can pretty much monitor anything, anywhere, all the time.

Some of this feels uncomfortable, like Ubicquia clamping cameras to light poles; some of it is uncomfortable and brilliant like SatVu operating satellites with thermal infrared sensors; some of it brilliant and practical like Duos Technologies (that scan freight cars for mechanical defects with computer vision) and GenLogs (which does more or less the same for trucking).

The message: sensors and computer vision are not wondertoys anymore, they are the basics. So: think about computer vision for your business—the ability to see and interpret reality as data. Now think about it being really good, and always-on. And passive. What are you going to look at? Shelves? Consumers opening packaging? Customers navigating your forms? Checkout lines? Customer service reps in the store? Think about all you could learn from this. Your whole business can be turned into game tape for reviewing and optimizing. And by the way: please use this power for good.

SECTORS AND COMPANIES

Financial Services: The Back Office Gets a Brain. $4.4B

The financial services story is getting simple: take all that eyeball-popping paperwork… to review underwriting opportunities, perform compliance checks… and have the agents do it. You’ll go faster, spend less money, service more clients. That’s it. There must be risk for the (regulated) customers of the AI companies, for the AI to screw up and make a big decision in a regulated space based on a hallucination. Collectively, company and client HAVE to keep humans in the loop, even if it’s just reserving their human capacity for the thorny cases. Still, if not simple to execute, the thinking is simple: have the robots do the paperwork. (Funny how agents seem so… December now.)

What it Means to You:

• If insurance underwriting can be agent-ified… so can your mess.

• There’s a fascinating touch in here with Rowspace. They turn a process into an asset by hoovering up all the stray threads of a financial analysis project (which presumably the client does well) and trying to make it a data object and a repeatable formula. Imagine doing that with an academic paper, or script development. Or heck, a relationship.

• What worked? How can we keep doing that?

Clearwater Analytics

Clearwater Analytics is a cloud platform that automates investment accounting and reporting for institutional investors — insurance companies, pension funds, and sovereign wealth funds. It reconciles portfolio data from dozens of custodians continuously, replacing manual processes that once took days. The platform processes more than $7 trillion in assets globally across every major asset class.

Debt Financing, $3.5B, February 13, 2026.

Wayflyer

Wayflyer is an Irish fintech that offers revenue-based financing to e-commerce businesses, using their real-time sales data as the underwriting signal. For instance, a brand on Amazon or Shopify uploads storefront analytics and receives a funding offer within hours.

Debt Financing, $250MM, February 18, 2026.

Basis

Basis builds AI agents that handle full accounting workflows — reconciliations, financial statement prep, complex client engagements. Roughly 30% of the top 25 U.S. accounting firms already use the platform. Early deployments show productivity gains of 30 to 50 percent.

Series B, $100MM, February 24, 2026.

Jump

Jump is an AI assistant built for wealth managers and financial advisors. It attends client meetings, transcribes conversations, and auto-generates compliant meeting notes and action items formatted to each firm’s standards. Advisors report recovering several hours of administrative time each week.

Series B, $80MM, February 19, 2026.

Fieldguide

Fieldguide embeds AI agents into audit and advisory workflows — gathering documents, tracking evidence requests, checking work against professional standards, and flagging anomalies. Half of the top 100 U.S. accounting firms use the platform, including members of the Big Four. CEO Jin Chang, a former CPA, founded the company after concluding that firms are being asked to do more complex work with a shrinking pool of qualified staff.

Series C, $75MM, February 2, 2026.

Bretton AI (formerly Greenlite AI)

Bretton AI deploys AI agents inside regulated banks to handle financial crime investigations — KYC due diligence, sanctions screening, and the audit narratives regulators require. The platform has completed more than 1.2 million investigations in production and is used by Robinhood, Mercury, and Coastal Community Bank.

Series B, $75MM, February 10, 2026.

Accrual

Accrual automates tax preparation for accounting firms — reading K-1s, 1099s, emails, and handwritten notes, then producing draft returns ready for partner review. Clients including H&R Block and Armanino report cutting preparation time by more than 85 percent. The platform was built by engineers from Stripe and Brex.

Series A, $75MM, February 5, 2026.

TRM Labs

TRM Labs is a blockchain intelligence platform that traces transactions across more than 30 blockchains to help law enforcement, financial institutions, and crypto companies investigate financial crime. Clients include the IRS, Coinbase, PayPal, and Visa. I always thought it was weird when people acted like crypto was some kind of secret money domain—the trail is indelible. The February 2026 round valued the company at $1 billion.

Series C, $70MM, February 4, 2026.

Rowspace

Rowspace helps hedge funds and asset managers extract value from their own proprietary data — research, investment theses, meeting notes, and trading history — making it searchable by analysts in plain language. Most firms have years of institutional knowledge buried in email threads and disconnected databases. Rowspace treats that archive as a competitive asset that deserves infrastructure.

Series A, $50MM, February 25, 2026.

Mea

Mea is a cloud-native AI platform for property-and-casualty insurers that automates underwriting triage and claims document review. Carriers adopt it to improve their combined ratio —the core profitability benchmark in insurance, that tells you if you’re underwriting profitably— by removing manual effort from each policy.

Private Equity, ~$50MM, February 17, 2026.

Harper

Harper is a commercial insurance brokerage that uses AI to review insurance applications and match companies to underwriters, with a small staff, and they claim, “deliver Fortune 500-level service to every business in America.” Nice founder story about the CEO’s uncle’s truckjing business getting squashed by a lack of insurance. They’re trying to make insurance shopping feel less slow and less opaque.

Series A, $47MM, February 25, 2026.

Onshore

Onshore helps companies claim tax credits that come from their doing R&D. To make the claims though, you need to do “component” level analysis of your R&D projects… which is so time-consuming most big accounting firms just make broad estimates, which leaves money on the table. With AI you can dig right in there and get the documentation to support R&D claims, which saves unclaimed tax credits.

Series B, $31MM, February 17, 2026.

Entertainment, Marketing & Advertising: Marketing breaks the mold. $776MM

AI is reshaping both how content is produced and how audiences discover it. ElevenLabs has cut the cost of professional voice content dramatically. Profound and Brandlight are tackling a new and urgent problem: as AI systems increasingly answer consumer questions directly, a brand’s profile in those AI responses matters as much as its search ranking.

What it means to you:

• The advent of Profound is profound. Search, to AI summaries. We are watching a 25-year digital ecosystem… I mean I’m old enough to remember the first time I heard about “Google”… transform.

• This is a massive deal for marketing. Every marketer will be forced to jump on this and reshape their every marketing move.

• I wonder if there is some reputation / social component to this as well. If AI search/summaries are capturing most public info about your company… that won’t be a narrow search metric. It will be a broad summary of your whole reputation. SEO will become OOM. Overall Opinion Management.

• Similarly, the arrival of Simile feels like a rubicon moment for market research. When I read Joon Sung Park’s research paper, I could really see how all this synthetic data jabber means we can turn consumer/client research into something quick-turn and cheap… in a good way.

• When this scales, there will be no excuse for making dumb, blind decisions in business. You can have your synthetic customer at your elbow. Turn around and ask them.

ElevenLabs

ElevenLabs builds AI tools for generating, cloning, and dubbing human voices across 29 languages. A filmmaker can dub a feature into a dozen languages without casting new actors; a game studio can generate hundreds of character voices from written scripts. Founded in 2022 by Mati Staniszewski and Piotr Dabkowski — Polish engineers frustrated by the quality of audio dubbing — ElevenLabs has grown into one of the most widely used AI content tools on the market.

Series D, $500MM, February 4, 2026.

Simile

Companies make big decisions about customers and workers with research and surveys. Simile builds models that simulate how people may respond using sophisticated, person-based models. The product is basically a simulation sandbox for human behavior and attitudes. I blogged about their approach last year.

Series A, $100MM, February 12, 2026.

Profound

Profound helps brands understand and improve how they appear in answers generated by AI systems — ChatGPT, Claude, Perplexity, and others. When a consumer asks an AI assistant for a product recommendation, the result depends on how brands are represented in the AI’s knowledge and retrieval systems. Profound measures that visibility, benchmarks it against competitors, and provides tools to optimize it over time.

Series C, $96MM, February 24, 2026.

Guidde

Guidde converts screen recordings into polished instructional videos with no manual editing — the AI adds narration, annotations, chapter breaks, and captions, then generates a shareable link. Companies use it to build documentation libraries for customers and employees, dramatically reducing the effort required to keep training materials current as products change.

Series B, $50MM, February 25, 2026.

Brandlight

Brandlight monitors how a brand appears across the digital environment — social media, news outlets, review sites, and AI-generated content — running continuously and flagging shifts in reputation or competitor positioning. A company launching a new product can catch a wave of negative reviews before it compounds, or verify that its profile in AI-generated answers is accurate.

Series A, $30MM, February 11, 2026.

Healthcare: Attack the bottlenecks. $526MM

The U.S. healthcare system spends an estimated $950 billion a year on administrative costs — hours of physician and specialist time absorbed by paperwork rather than patient care. Several companies here are attacking that waste directly, automating admin. Others are taking bigger swings in drug development and clinical trials-- cutting the cost and time to bring new medicines to market.

What it means to you:

The healthcare space remains the one to watch. Players here are attacking every angle of the business with AI.

• Innovation (Sarborg). Ongoing patient relationships (Dental Monitoring). Admin headaches (Alaffia, Anterior).

• I’ll call out Dental Monitoring as the most curious and important of the bunch. They are maintaining a close link with patients after they leave the office.

• Think about this for, say automotive, appliances, electronics, housewares, services like accounting and legal and mental health and general practice medicine… what opportunities are available for better service if you link to your customer after they drive off the lot? Between visits? There’s a lot of value here.

Sarborg

Sarborg applies AI and cybernetics to pharmaceutical research, helping drug companies model complex biological systems and identify promising drug candidates faster. The platform draws on molecular and clinical datasets to predict how compounds will interact with biological targets, aka “signatures,” reducing costly early-stage laboratory work. Sarborg focuses on oncology and rare diseases, where the search space for effective drugs is vast.

Corporate Round, $115MM, February 20, 2026.

Dental Monitoring

Dental Monitoring lets orthodontists track their patients remotely using AI. Patients photograph their teeth every few days with a smartphone app; the company’s computer vision models track tooth movement, flag if a patient is skipping aligner wear, and alert the orthodontist when treatment is off course or ready to advance. The system lets a practice manage a significantly larger patient panel without proportionally increasing in-office appointments.

Private Equity, ~$100MM, February 10, 2026.

Alaffia Health

Alaffia Health makes AI software that helps health insurance companies review medical claims more accurately and at lower cost, examining clinical documentation and applying coverage policies before payment is issued. The platform automates both the clinical and administrative dimensions of claims review, letting insurers process more cases without proportionally increasing reviewer headcount. I remember well the ProPublica piece that made claims reviewed by technology sound like a dead letter office for consumers. Companies like Alaffia seem like an opportunity for traceability and transparency (that is: show the consumer how their claim payout was arrived at!) that would ultimately help the industry’s reputation… if Big Insurance wants to use these tools that way. We’ll see.

Series B, $55MM, February 3, 2026.

Deep Intelligent Pharma

Deep Intelligent Pharma (DIP) uses more than 10,000 specialized AI agents to automate clinical trial protocol design, regulatory document preparation, and statistical analysis. Companies including Bayer, Johnson & Johnson, and Roche use the platform. In one documented case, a drug protocol DIP prepared for a Japanese cancer vaccine company passed the national regulator’s review on first submission, with no revisions required.

Series D, $60MM, February 2, 2026.

Biorce

Biorce is a Barcelona-based startup building an AI clinical assistant for biotech and pharmaceutical companies running clinical trials. By handling the most repeatable parts of trial management, Biorce aims to make clinical development faster and more affordable, particularly for smaller biotechs that cannot afford large contract research organization engagements.

Series A, ~$52MM, February 2, 2026.

Anterior

Anterior automates prior authorization for health insurance companies — the process that requires physicians to obtain approval before many treatments can proceed. I hate it when I need one for an MRI. The system reviews the patient’s clinical record and the insurer’s guidelines simultaneously, then either auto-approves or routes the case to a human reviewer with a pre-packaged summary.

Venture, ~$40MM, February 12, 2026.

Catheter Precision

Catheter Precision builds robotic tools that give electrophysiologists more stable control during catheter-based heart procedures, using mechanical stabilization and real-time position data to hold the catheter exactly where the physician places it. The system also captures procedure data that cardiologists can review to improve technique over time. The company is publicly traded and based in Fort Mill, South Carolina.

Post-IPO Equity, $36.5MM, February 12, 2026.

Lotus AI

Lotus AI is building a consumer health platform that uses AI to help individuals interpret lab results, wearable sensor data, and symptom patterns without requiring a doctor’s visit for basic guidance. As continuous health monitoring generates more data than most people can meaningfully process, Lotus AI’s bet is that an intelligent personal health layer becomes as necessary as biometrics.

Series A, $35MM, February 2, 2026.

splose

splose is an Australian admin toolkit for health clinicians — physiotherapists, speech pathologists, occupational therapists — handling scheduling, clinical notes, billing, and compliance. AI features can generate draft session notes from voice recordings. Health professionals commonly spend a third of their working day on documentation; splose is trying to reclaim that time for clinical work.

Series A, ~$32MM, February 8, 2026.

Professional Services: Monitoring. $413MM

Professional services is a little bit of a catch-all, and the way I’ve categorized it, I’m including security in it. This month, security dominates: tech is now allowing companies to monitor the edges of human experience. I’ve been quietly grateful, up to now, that the US has seemed to avoid this fate of having eyes on us, on every corner. Oh, well.

What it means to you:

Security is about monitoring the edge… not in a computing sense, but in an experience sense.

• What is the “edge” of your customer relationship? Where do you feel like you wish you could monitor better how customers are interacting with your product and service? (How can you do this in non-creepy way, friends?)

• What value could you derive if you were able to do this monitoring?

Ubicquia

Ubicquia uses streetlights and utility poles as a platform for local government sensors, turning city infrastructure into a dense data-collection network — gathering information on crime, traffic, air quality, noise, and in some cases acoustic events like gunshots. City governments use the data to optimize traffic signals, plan maintenance, and respond faster to public safety incidents. More than 500 cities have deployed Ubicquia hardware on their streetlight networks. I have a feeling that this shift… the US become more like London or China, with camera surveillance everywhere… is inevitable. Wish it weren’t so.

Series D, $106MM, February 23, 2026.

Gambit Security

Gambit Security, founded by veterans of Israel’s Unit 8200 intelligence corps, continuously maps an organization’s backup systems, recovery plans, and infrastructure dependencies, then stress-tests them against current ransomware techniques to find gaps before an actual incident does. The company’s research claims that only about 5% of enterprise systems could survive a ransomware attack intact. Gambit emerged from stealth in February 2026 with $61 million across its seed and Series A rounds.

Series A, $56MM, February 25, 2026.

Nimble Way

Nimble Way is a data collection platform that gathers structured product and pricing information from e-commerce sites, marketplaces, and distributor catalogs at scale. Retailers, brands, and market intelligence firms use the data to monitor competitor pricing, track product availability, and identify supply patterns in near real time.

Series B, $47MM, February 24, 2026.

Cogent Security

Cogent Security uses AI agents to help security teams manage backlogs of software vulnerabilities, connecting to existing scanners and prioritizing which issues to fix based on likelihood of exploitation, asset criticality, and business impact. The goal is to help stretched security teams focus limited remediation time on vulnerabilities that pose the highest actual risk, not work through an undifferentiated queue.

Series A, $42MM, February 18, 2026.

ORION Security

ORION prevents data leaks by replacing static, rules-based data loss prevention tools with real-time contextual intelligence that learns what normal looks like for each employee and organization. The platform flags behavior that deviates from that baseline — catching insider threats and accidental leaks without the constant false positives that make traditional DLP tools so difficult to operate in practice.

Series A, $32MM, February 3, 2026.

Didero

Didero is an AI procurement platform for manufacturers and industrial distributors, managing the full purchasing workflow — gathering vendor quotes, comparing specifications, flagging price anomalies, and generating purchase orders. The platform targets mid-market manufacturers who manage thousands of components and supplier relationships but cannot afford a large procurement department.

Series A, $30MM, February 12, 2026.

Defense, Government & Space: More monitoring. $276MM

The convergence of cheap sensors, high-bandwidth networks, and edge AI has made persistent monitoring feasible. But I admit, it freaks me out—that the heat we give off can be detected by satellites and used as monitoring data, as an example. Maybe I saw too many action movies in the 80s, with those spooky, green heat-sensing goggles.

What it means to you:

• More reinforcement of the monitoring themes above.

• But think about it defensively this time. (Pun intended.) What bad outcomes to your business or customers can you prevent by monitoring your experience edge?

Vosker

Vosker makes cellular-connected surveillance cameras for remote properties with no Wi-Fi or grid power, running on solar and using on-device AI to classify motion as human, animal, or vehicle — sending alerts only when something meaningful happens. Farmers, construction operators, and rural property owners across North America use them to monitor locations that cannot be staffed.

Debt Financing, $200MM, February 10, 2026.

SatVu

SatVu operates satellites with thermal infrared sensors that detect heat signatures from industrial facilities, shipping yards, and power plants anywhere on Earth. The company’s AI analytics translate those thermal readings into economic intelligence: which factories are running at capacity, which refineries have gone offline, which logistics hubs are active overnight. Financial institutions, commodity traders, and government agencies use the data for a near-real-time view of global economic activity.

Venture, ~$41MM, February 16, 2026.

Tytan Technologies

In the context of the Ukraine war, Tytan Technologies is a Munich-based company developing AI-guided defense systems designed to detect and respond to drone attacks. Drones are the new language of war. Makes sense that we’re seeing commercialized detection systems. Odd to see a defense company with booyah energy in Europe.

Series A, ~$35MM, February 24, 2026.

Real Estate / Construction: Smarter buildings (and landlords). $141MM

Real estate has historically been among the slowest industries to adopt technology—which makes sense when you think about the vast surface area—and it also means the efficiency gap AI can close is unusually large. These companies are building data infrastructure for the built environment.

Dwelly

Dwelly is a London-based lettings and property management platform using AI to handle the full landlord-tenant lifecycle — listings, tenant screening, rent collection, maintenance coordination, and compliance documentation. The system automates work that leasing agents traditionally perform. New UK regulations around tenancy agreements and energy performance have increased the administrative burden the platform is designed to absorb.

Debt Financing + Series A, ~$93MM combined, February 25, 2026.

Metiundo

Metiundo builds hardware and software to measure energy and water consumption at the individual apartment or commercial unit level in multitenant buildings. Property owners and managers use the data to allocate utility costs accurately and comply with European regulatory requirements for building energy performance. This feels like it’s going in a smartr direction to optimize energy billing, and, I suppose energy use. I mean why are all the office lights left on all night??

Series A, ~$48MM, February 10, 2026.

Transportation / Warehousing: Eyes on every mile. $125MM

More surveillance, this time used for good: monitoring the health of physical logistics systems. What’s striking is the passivity of this: the trucks and truckstops, the freight trains… they’re monitored… and their data is interpreted… completely passively. That is a backward indicator of a massive jump in compute. That we have enough of it to capture, and interpret, a real-time flow of images across sprawling systems… that is a serious change.

What it means to you:

• Think about computer vision for your business—the ability to see and interpret reality as data.

• Now think about it being really good, and always-on. And passive.

• What are you going to look at? Shelves? Consumers opening packaging? Customers navigating your forms? Checkout lines? Customer service reps in the store?

• Think about all you could learn from this. Your whole business can be game tape for reviewing, and optimizing.

Duos Technologies

Duos Technologies builds AI-powered machine vision systems that inspect freight and passenger railcars. Cameras and sensors installed at trackside portals scan every surface of each car as a train passes through, analyzing thousands of images in seconds to detect mechanical defects, foreign objects, and safety hazards. Amtrak is a customer, deploying the technology at strategic points across its network.

Post-IPO Equity, $65MM, February 26, 2026.

GenLogs

GenLogs was founded by veterans of the U.S. intelligence community — including CEO Ryan Joyce, a former CIA officer — who applied surveillance tradecraft to the trucking industry. The company runs a nationwide network of roadside cameras that feed AI models to build a continuously updated picture of carrier movements. Shippers, insurers, and law enforcement all use the platform; in one documented case, a state agency used GenLogs data to track a truck involved in interstate trafficking and recover a victim.

Series B, $60MM, February 2, 2026.

Agriculture: Farm to Table Automation. $100MM

The food industry runs on razor-thin margins and complexity — thousands of SKUs, perishable inventory, and fragmented supplier networks. The food business has enough data to be run smarter; so far it has lacked the tools to act on it. These two companies are trying to make a dent. Stepping back, you could see how progress in commercially available data/AI tools in grocery land would in theory threaten the hegemony of Amazon and Walmart--who have gained their lead through decades of tech investment. And then you catch yourself and laugh.

GrubMarket

GrubMarket has built an AI-powered operating system for the North American food supply chain, connecting food producers directly with retailers, wholesalers, and food service operators. The platform runs software across procurement, invoicing, logistics, and compliance, replacing the spreadsheets and phone calls that still define much of the industry. This is its Series H round — the eighth. Not sure what to make of that, except that it’s a fairly small round, meaning weirdly that food, as critical as it is, is getting proportionately much smaller funding than businesses focused on information. All about scale. Sometimes in a depressing way. (Shouldn’t food be at the top of our concerns?)

Series H, $50MM, February 2, 2026.

Pepper

Pepper builds AI-powered e-commerce software for food distributors and specialty suppliers — the wholesale businesses that sit between producers and restaurants or retailers. The platform automates quoting, order entry, and customer follow-up, replacing phone and email workflows with a digital interface that tracks every interaction. Food distribution has lagged behind other industries in technology adoption, and Pepper’s bet is that the brokers who modernize first will win market share from those that don’t.

Series C, $50MM, February 20, 2026.

Energy: AI Independence in the Gulf — $50MM

Origen

Origen is a UAE-based startup designing integrated power management. Which makes it seem more focused than it is: it’s really a geographic play, putting it at the center of an ambitious AI investment program in the Gulf to build compute capacity they control. The company is backed by private equity investors who see infrastructure as the precondition for regional AI leadership.

Private Equity, $50MM, February 18, 2026.

Retail / Consumer: EVs — $31MM

Ever

Ever is building an AI-native platform for buying and selling electric vehicles at dealerships — transparent pricing models, AI-guided trade-in valuations, and automated financing options applied to a transaction that has historically depended on negotiation and incomplete information. This will be important if/when EV adoption accelerates.

Series A, $31MM, February 12, 2026.

JIMMY’S REQUEST

“Parting is such sweet sorrow.” - Juliet, Act 2 Scene 2

Jimmy is 14, chubby, lives in Virginia. He doesn’t help himself by eating a stack of Oreos whenever he gets behind closed doors. He is clever, bookish, glasses, has one of those Poindexter-y voices that sort of begs to be accompanied by him pushing up his glasses with his finger. Despite his big brain, Jimmy is disorganized. His mind is a thrash-metal hurricane of hormones and neurosis (already). So Jimmy is late to start his term paper. Like, really late. Like it’s due tomorrow. He thought it was just a page but it’s FOUR PAGES and he is screwed.

So he does what everyone else does--but he promised himself he wouldn’t do, especially in English which is his favorite class--which is to write a paper with ChatGPT. He copies the question from the teacher’s website… Are Romeo and Juliet in love, or is it merely obsession? And what is the difference?... and he pastes it into the chat bar. He presses enter. Off it goes. Parting such sweet sorrow. It is 9:41pm. Jimmy stress-eats two Oreos in one bite.

[This is a description of me in junior high, so we’re clear.]

While he chews, Jimmy’s request flies into the maw of a multi-hundred-billion dollar industry that has become one of the most important in todays’ world. No, not that one--artificial intelligence. The other one. The data centers that supply AI.

We are going to learn what happens inside a data center through the eyes of Jimmy’s term paper request.

WHY IS THIS IMPORTANT?

“It is too rash, too unadvised, too sudden.” — Juliet, Act 2, Scene 2

In the single month of January, 2026 I wrote how $19B in private company investment was chasing data centers, telecom infrastructure, and chip-making. Seems like a lot. That’s nothing compared to what’s happening with public companies. According to the New York Times, Amazon, Meta, Google and Microsoft together spent $400B on data centers in 2025. For perspective, that’s the GDP of Colombia, or the cost of building 100 One World Trade Centers.

Meta gets the prize: colonizing Monroe, Louisiana with a newly-announced 2,250-acre data center (that’s two Central Parks), hiring 5,000 workers, and consuming a gigawatt of power (enough to light up Las Vegas for a year). This is in addition to Meta’s other new sites in Indiana and Texas.

What makes me curious is that all this activity and building is happening… kind of offstage. Data centers just sit there and hum; they’re being built in the suburbs or in the country; and so much activity in the South. Data center operators must be getting what they want down South:

• Network connectivity and routes to major internet backbones

• Land availability

• Water availability

• Attractive power costs

• Low climate and natural disaster risk (you don’t see these going up in Florida or Southern California)

• Access to skilled labor and a solid construction ecosystem

• Proximity to end users and major metros for reduced latency

• Welcoming politicians and communities

• State and local tax incentives

These last two go together: in Virginia, a commission found that in areas with “mature data center markets,” data center revenue could contribute as much 31 percent of total local revenue.

Okay, back to love and obsession.

INTO THE CLOUD

“Thoughts which ten times faster glide than the sun’s beams.” — Juliet, Act 2, Scene 5

Jimmy’s prompt leaves his laptop. From there it hops to his home Wi‑Fi network, to his router, his Internet Service Provider, to an OpenAI edge server which might be right there in town with Jimmy. The edge server routes the request to the nearest data center, in Boydton, Virginia. Boydton is a very small town. Its highest rated restaurant--a six-minute drive from the data center--is a breakfast joint with a southern-fried menu featuring a sandwich called “The Free Bird” and kids’ options for “Youngins.”

The request flies to the Azure data center in Boydton by fiber‑optic cables. Light travels quickly through fiberoptic but not instantly, introducing latency--a data center challenge, and the reason cloud providers place hubs strategically across the continent. Light in fiber takes about 5 milliseconds to travel about 1,000 kilometers (about 620 miles). In Jimmy’s case, his request would only have to travel 100 miles. So a single millisecond of latency.

GETTING INSIDE

“Stony limits cannot hold love out.” — Romeo, Act 2, Scene 2

That one millisecond later, Jimmy’s arrives at the Microsoft Azure data center in Boydton. It bypasses the rigamarole we would face as human visitors: the biometric screening. Passage through the wire mesh gates which are “crash-rated” (meaning you can’t smash through them with a car like Jason Bourne).

His request bypasses the loading dock and staging area where all the server hardware is unboxed and inspected.

It flies past the utility service entrance with its transformers, switchboards, battery rooms, and breakers.

It goes straight to area with a curious little nickname called the “Meet Me Room,” so called because it’s where cables from all the outside telecommunications networks “meet” the data center.

Once inside, Jimmy’s request sees a familiar sight: rows and rows of server racks, in cages nineteen inches high, humming with activity.

The servers are arranged in a hot-aisle/cold-aisle formation. What’s that? It’s a trick invented by IBM back in the 90s, to have the front side of the servers face each other (both pumping cold air)... then their backs--which, like your fridge at home, emit heat--also facing each other… so the whole apparatus acts like a kind of weather system. Hot air rising, floating around to the air conditioning system, where it’s cooled, and re-circulated.

And a data center is all about cooling.

TEMPERATURE

“These hot days is the mad blood stirring.” — Benvolio, Act 3, Scene 1

A data center is basically two systems wrestling with each other. One system powers the computers, which raises the temperature. The second system cools the first system.

Electricity comes from the grid: It is stepped down in voltage by transformers. It runs through Uninterruptible Power Supply (UPS) systems and batteries. Generators and batteries stand by for outages. Power Distribution Units (PDUs) and breakers feed the racks.

Then, the cooling path: Servers turn electricity into heat. Cooling equipment keeps the heat controlled. Some sites use chillers and cooling towers. Some use dry coolers. Some use direct liquid cooling--where liquid actually runs close to the chips. In any scenario, water--a lot of it--is being used in the cooling systems. Hence the controversy over data centers consuming local water supply.

Both systems--the one that powers the servers, and the one that cools them--burn electricity. So data center management has a key performance indicator called PUE, or Power Usage Effectiveness. PUE is a ratio. The arithmetic is:

Energy used by the total facility (for processing + for IT and cooling)

/

Energy used just for processing

In other words it’s a ratio of how efficiently the facility is using juice. If electricity is being used without any waste, your PUE is 1. (That is: there’s no IT or cooling energy in the numerator… so the ratio is just processing power over processing power.)

For perspective: Google reports an annual fleetwide PUE of 1.09. That means roughly nine percent overhead is used for cooling and power delivery.

Electricity--literally and figuratively-- is the currency of the data center.

TERA-WHAT?

“More light and light; more dark and dark our woes!” — Romeo, Act 3, Scene 5

To place all this power use in perspective, let’s look at the impact of data centers on the US grid. The United States generated about 4,151 TWh (terawatt-hours) of electricity in 2024, according to the Federal Energy Regulatory Commission. So 4,151 TWh is the size of the whole pie--all the juice we create and consume in a year.

I don’t know about you, but a “terawatt hour” doesn’t mean much to me. I do understand smoothies, margaritas, and chimichurri sauce, all of which require a blender. So let’s translate this power supply into blender capacity. A typical 500‑watt kitchen blender running 24 hours a day, 7 days a week, 365 days per year, would use about 4,380 kilowatt-hours (kWh (kilowatt-hours). On that basis, the whole United States electric system is the equivalent of running roughly 950 million blenders full time for a year.

The Lawrence Berkeley National Laboratory (LBNL) estimates that US data centers in aggregate consumed about 176 TWh in 2023. That’s about 40 million blenders. With the rampant demand by consumers and enterprises for AI, by 2028, according to LBNL, that number should rise to 100 million blenders--in effect more than doubling the data center portion of energy consumption from around 4% of total to around 10%. Mostly driven by AI. An extra 60 million blenders. That’s a lot of chimichurri.

PRICING

“There is thy gold.” — Romeo, Act 5, Scene 1

The economics of the data center also pivot on power.

One common pricing model is colocation. In colocation, a data center landlord rents you space, power, and cooling. You bring your own servers. In the United States, wholesale pricing is often based on dollars per kilowatt (kW) per month. In tight markets, the real estate giant CBRE has noted pricing levels around $200 per kilowatt per month or more for large deals.

The model we’re most familiar with is cloud. In that model, you rent compute as a service. You do not pay “per rack”; you pay for time on a VM (virtual machine), virtual in the sense that it’s not a specific machine, and because, as discussed below, you can be routed to any server in the data center’s national or global system.

For fancy compute--for instance you’re renting a VM made up of NVIDIA H100 GPUs (graphics processing unit)--you’ll pay about $100 per hour. For lower-end hardware (which obviously will produce less output) we’re talking more like $8/hour.

This is why efficiency--in compute language, “elasticity,” discussed below, and power use--matters. If the model produces an answer using fewer GPU seconds, the bill drops. If the data center runs at a better PUE (Power Usage Effectiveness), electricity overhead drops. If the system reuses cached results, it avoids work entirely.

THE MODEL SPEAKS

“Wisely and slow; they stumble that run fast.” -- Friar Laurence in Act 2, Scene 3

Now that we’ve contextualized the power use and economic model surrounding the data center, let’s pick back up on Jimmy’s question, which, at last, will get some attention. (Mind you, it’s only been a few milliseconds--Jimmy still has those two cookies in his mouth and is lovingly turning them to paste on his tongue).

The request hits a “gateway layer” here that validates Jimmy’s ID, gives him a session ID, and figures how to route the request to servers that have capacity to answer the question. If it’s here in Boydton, great. If not it will get routed to a data center in Manassas, or Leesburg, in another millisecond. The trade calls this “elasticity”--the ability to re-route demand to where there’s capacity; which allows data center clients like OpenAI to scale up and down their compute as needed. It’s also why the big cloud players are called “hyperscalers.”

But in this case, Boydton is free, so no re-routing necessary.

And at this point Jimmy’s question: Are Romeo and Juliet in love, or is it merely obsession? And what is the difference? … is changed to numbers--or tokens--so it can be fed into the large language model (LLM). Specifically the term paper question now reads like this:

13938, 96615, 326, 128971, 306, 3047, 11, 503, 382, 480, 26624, 79216, 30, 1958, 1412, 382, 290, 9809, 30

The request at last is fed into a GPU, or maybe a GPU cluster, so the LLM can do its magic. [For background: A GPU is a chip built to do many math operations at once. One GPU is not enough for a big model. So companies link many GPUs together into a cluster. A cluster is just a team of machines that act like one bigger machine.]

These clusters can get enormous. AWS describes “UltraClusters” that can scale up to tens of thousands of NVIDIA H100 GPUs.

THE TERM PAPER

“Read o’er the volume… and find delight writ there.” — Lady Capulet, Act 1, Scene 3

Jimmy’s assigned GPU hums a bit, and comes up with the following:

158985, 78, 326, 128971…

Oh sorry, you wanted English:

Romeo and Juliet feel something that looks like love but often behaves like obsession.

Love grows slowly, the model advises. True love sees the other person clearly, and can survive compromise and time.

Obsession, the model says, is urgent, idealized, and fueled by intensity rather than knowledge.

Our GPU cluster concludes: Romeo and Juliet’s feelings are sincere and emotionally powerful, but Shakespeare suggests the young lovers are driven more by youthful passion, fantasy, and rebellion than by the steady, tested devotion that characterizes mature love.

The answer returns to Jimmy by reversing the original path: it passes from GPU cluster, to Meet Me Room, to fiberoptic cable, to edge server, back to Jimmy’s laptop. Our ninth grader sees the response. He copies and pastes the whole four page essay into a Google doc. He adds a title and date.

He slaps the laptop shut.

Our boy gets in bed. He turns off the light. Does Jimmy now truly understand the difference between love and obsession? Even with so many servers around him to explain, he does not. And maybe that’s the way we want him tonight: gone to sleep without brushing his teeth. Still tasting Oreo. Slumbering without any notion of overheating GPUs, or hearts.

“Is love a tender thing? it is too rough, too rude, too boisterous,” says Romeo, advising patience to our young scholar. “And it pricks like thorn.”

Investments in data AI and robotics in January totaled $52B (VC, debt, post-IPO). Since these are largely private companies seeking funding and rapid share growth, they’re pretty open about their methodologies. That’s good for you as a leader--it means you get to learn nitty gritty details about the cutting edge of AI innovation that fifty-two billion is chasing.

But… we’re only going to focus on what I consider the “interesting” $8B slice of that pie--the companies working on specific use cases you can learn from, that will stimulate your innovative spirit. The case studies below are a bonanza of variety, industry and focus. And they couldn’t be more fresh. Most of the funding was in the last few weeks.

A few themes emerge:

• Like Pinocchio in reverse, experts are being turned into bots. AI companies are slurping up expertise from the brains of specialists and turning them into automated, scalable tools. Ivo and Summize read legal contracts. Flora generates marketing creative. VieCure creates well-rounded care plans for cancer patients. (What happens when the last expert is automated and the craft is lost? For our kids to sort out.)

• Experts are getting helpers that speak natural language. Digital helpers are emerging for highly skilled, professional knowledge areas. OpenEvidence is stuffed with medical data and allows your doctor to ask it questions, a professional-grade GPT. Rogo answers natural-language questions about company financial analysis. Gyde is an expert companion for insurance brokers. Why pay that monster hourly fee when you have an JD in a box?

• Making more data to describe more niches. You’d think we’d run out of data by now, but it feels like we’re just getting started. Compa scrapes pay data to help companies compete for employees at market rates. Hydrosat pulls planet heat data for water planning. Apella tracks all the complex actions in an operating room--like game tapes for surgeons; The Pitt on analytics. Protege gathers databases as off-the-shelf training data for your model. And there’s creepy stuff: Uni-Ubi (that’s short for “universal and ubiquitous,” shudder) is a China-based facial recognition tech. Think twice about attending that protest when the “UFace 3 Pro” is on the market.

• Data identifies micro-markets. The beauty of data analysis is that it enables you to find that underserved niche, and score a win. Mendra scans patient data to find patients who need help, but are subscale by the lights of Big Pharma. Corgi serves the niche market of startups to help them get insurance quickly. Personalization for markets.

Below are snapshots the 50+ data and AI companies that raised $30MM or more in January, by sector, with select commentary from me. When you see the specificity of these applications, it will start to make sense why a week ago the stock market punished software companies to the tune of $300B. Vibe-coding makes it easier for enterprises to quickly build modules and tools which would have been the domain of their software providers.

Oh--what’s in the other $44B from the January investment pool, you ask?

• A big chunk, $19B, is flowing to data infrastructure like data centers, telecom, and GPUs. Which maybe tells you why New York and five other states are calling for a moratorium on data center buildouts.

• $23B is fueling general AI applications and tools--not domain-specific--including a $20B raise to xAI. These are interesting tools to explore for how you realize your vision, but don’t focus on what you build.

• The remaining $2B is robotics--fun stuff, if you’re a manufacturer.

So let’s get started. I grouped these narratives into:

• Healthcare at the Head of the Class

• Physical AI is a Trend You Sould Watch

• Space data

• “Sir Yes Sir” to the Defense Industry

• Dialogue to Data

• Measure the Treasure: Finance (and HR and Energy)

• Admin Gets Automated

• The Soft Stuff (Creativity)

Healthcare at the Head of the Class

Watch and learn from what healthcare startups do: they gravitate to where they already have data: Protein data banks. MRI images. They look for efficiencies: operational stuff like MRI setups and analysis. But also efficiencies in treatment discovery, which is about growth, not savings. For instance, getting models to find a “fit” between flawed bodies and treatments by trying iteration after iteration--heaping abuse on mathematically reproduced cells to spare the expense and ethical drama of testing our real parts.

What it Means to You:

• Where are you data-rich now that you can find efficiencies?

• Where would massive, simulated, math-driven trial and error process create results for you?

• Where can you simulate experiments or research that would otherwise be expensive or risky? Digital crash test dummies.

OpenEvidence

OpenEvidence is built for clinicians who need answers during care. It searches peer-reviewed papers and clinical guidelines and returns cited summaries. The product keeps the source trail visible, which matters in medical decisions, and it gives your doctor quickly accessible medical info. I have asked the open web tool weird questions and gotten good answers. Funding: $250MM on Jan 21, 2026.

Soley Therapeutics

Soley Therapeutics starts with cells under stress and measures their responses. Those experiments generate large datasets of biological signals. (See my trend “You Go First”: you’d rather your simulated cells take a beating than find out after you’ve started treatment.) Funding: $200MM on Jan 7, 2026.

Mendra

Mendra uses AI to scan scientific literature, trial history, and clinical signals for overlooked market opportunities. That informs how to design development paths. The team pairs that analysis with traditional drug development execution. Funding: $82MM on Jan 21, 2026.

Apella Technology

Apella Technology turns operating rooms into measurable systems. It uses video and sensor data to capture timing, handoffs, and equipment movement--ops analysis. Computer vision converts that raw footage into structured events hospitals can analyze. The goal is higher throughput. Leaders can see where delays start and which steps create the most waste. I find this compelling as a data problem. As a patient, makes me feel like a car on a factory line. Funding: $80MM on Jan 7, 2026.

Orbem

Orbem turns MRI into an industrial inspection tool for biology. It scans items quickly without cutting them open, then analyzes internal structures. AI classifies what it sees inside eggs, produce, and other samples at high volume. That supports quality control where the outside view can be misleading. The approach is expanding beyond food into broader life-science applications. Funding: $64.6MM on Jan 20, 2026.

Vibrant

Vibrant combines target data, protein information, and screening results into a learning loop; then their AI proposes candidates and helps prioritize experiments that are likely to succeed. Reflects a broader move toward computational design as a default in biotech. Funding: $61MM on Jan 12, 2026.

VieCure

VieCure builds software for oncology clinics managing complex care with limited staff. It pulls patient history, labs, and treatment plans into a single operational view. It catches gaps in care, highlights high-risk cases, and suggests next actions for teams. The platform tracks outcomes so clinics can learn across patient populations. This could make cancer care more consistent outside major academic centers. Funding: $43MM on Jan 12, 2026.

Protege

Protege connects data owners with clients that need large, clean datasets for model training. The platform handles permissions and delivery, and keeps clear records of where data came from and what it can be used for. Protege started with healthcare data, but are branching out; hence the investment. Funding: $30MM on Jan 7, 2026.

Vista.ai

Vista.ai focuses on a narrow bottleneck in imaging: MRI scanning workflows. Its software automates the setup of MRIs scans and learns from prior scans and protocols to reduce missed views. That can shorten appointments and make scanners easier to staff. Funding: $29.5MM on Jan 14, 2026.

Physical AI is a Trend You Should Watch

I think Jensen Huang coined the phrase “physical AI.” It’s AI that relates to the physical world. At the extreme it’s world models--detailed cognitive maps of physical space that allows us to optimize our interactions with the complex 3D world. But at its core it’s basically creating a feedback loop with our machines. “Hey machine: when you did A, what worked? What didn’t? Should we try B this time?”

What It Means to You:

• If there is a physical element to your business, a future of much higher efficiency and effectiveness might not be far behind.

• Manufacturing? Deliveries? Managing inventories?

• If any of this applies to you, and you’re not familiar with the basics of physical AI, I strongly recommend you educate yourself.

Hadrian

OK, I have to admit I’m a fangirl of Hadrian. Which is funny because they are extremely political--a lot of their rhetoric is about winning the defense race against China and returning manufacturing prowess to the States. Fine. They are actually quite secretive about how they use physical AI exactly. Understandably: Hadrian runs factories designed for aerospace and defense production. I think what they’re doing is capturing machine data from every aspect of production. Then linking that data back to designs and throughput measurements to optimize output. All this tech, they claim, acts as an “Iron Man suit” for their workers--who they train up from scratch from being paralegals and nurses--which cranks productivity. This is their second nine-figure raise in six months so something must be going right. Funding: $131MM on Jan 9, 2026.

Mytra

Mytra builds robots that move pallets inside warehouses and industrial sites. It uses cameras and sensors to navigate and avoid people safely. The system also tracks inventory locations so material moves are traceable. Fleet software coordinates routes to reduce traffic jams on the floor. Funding: $120MM on Jan 15, 2026.

Richtech Robotics

Richtech Robotics builds service robots for restaurants, retail, and public spaces. The machines rely on sensor data for navigation and safe movement near people. Software manages routine tasks that repeat all day, like delivery and basic interactions. As fleets grow, operating data helps improve reliability and route planning. Funding: $38.7MM on Jan 29, 2026.

Lyte AI

Lyte AI focuses on 3D capture and machine perception. It turns raw visual inputs into 3D shape data that software can measure. That supports robotics, industrial inspection, and mapping where accuracy matters. AI fills gaps in sensor data and stabilizes outputs in difficult conditions. The company is building a perception layer for physical systems. Funding: $107MM on Jan 5, 2026.

Space Data

It’s extraordinary how data collection in space has moved from the exclusive provenance of government to a robust private industry. Tellingly, the OG in this area, Planet Labs, made their first satellites with no--repeat no--manufacturing from legacy aerospace--that’s how creaky the old system was. It’s being built from the ground up with new entrepreneurs with clever new lightweight contraptions.

What It Means to You

• If we can build satellites the size of microwave ovens that fly around the planet taking measurements… then we can measure whatever it is you want, too.

• Measure foot traffic in your car dealership? Facial recognition tech that tells you if customers are anxious coming in and happy going out? Digital body twins that try on clothes in 3D? This should uncork the imagination.

Northwood Space

Northwood Space builds the ground infrastructure that keeps satellites connected. It is creating a network that \manages antennas, download windows, and delivery across customers. That increases how often satellites can send data without each operator building ground stations. The company is treating ground stations like scalable infrastructure that can be automated. Funding: $100MM on Jan 27, 2026.

Hydrosat

Hydrosat uses satellites to track heat and water stress across the planet. It collects thermal infrared data and turns it into maps of crop health and drought risk. Models clean signals, correct for noise, and translate readings into usable indicators. Customers use the data to monitor resources and measure change over time. It’s good to see this data flowing from private sources after so much NOAA funding was cut. Funding: $60MM on Jan 15, 2026.

Satellogic

Satellogic operates Earth-observation satellites that generate steady streams of imagery. It processes raw satellite data into usable maps and change detection products. Customers use the output to monitor infrastructure, agriculture, and environmental conditions. I find this nook of satellite tech fascinating. Satellogic seems to compete with the much more established Planet Labs, hurling tiny, private satellites around the globe. As a challenger, Satellogic makes greater claims about AI, making claims that sound like world models--which means you should be able to ask questions to a digital twin of the earth and get predictions. Funding: $35MM on Jan 26, 2026.

“Sir Yes Sir” to the Defense Industry

I love what’s happening in the defense sector right now. Yes, yes, it’s all very macho, these highly secure systems for like, warfighting; and there are all these butch acronyms like “ATO.” But the cool part for us yellow corporate urbanites is that data and tech are being made secure--like, really secure--behind walls. Or airgaps. Why can’t we do that for your company and your client data? Or for the intellectual property of book authors?

What it Means for You:

• Highly secure systems that protect your IP from LLMs and prevent it from being someone else’s training data, should be a paramount concern to every company.

• If this capability can be widely commercialized, in my opinion, it should be standard.

Domyn

Domyn builds AI systems for regulated industries where data cannot leave controlled environments. It focuses on deployments that keep proprietary documents and records on secure infrastructure. The company tunes models on internal policies, research, and transaction data--useful for banks and enterprises that need strong controls. Funding: $600MM on Jan 1, 2026.

Defense Unicorns

The poster child of this concept. Defense Unicorns packages data and software delivery for national security teams, often literally airgapped, and monitors every input and output for security approval. Turns security checks into continuous verification instead of one-time audits. Funding: $136MM on Jan 13, 2026.

Onebrief

Onebrief is built for military planning where handoffs create friction--you don’t want your teams in the field to have version control issues with the attack plan. It replaces slides and email chains with a shared plan that updates in real time, which matters when situations shift quickly. Funding: $200MM on Jan 13, 2026.

Harmattan AI

Harmattan AI builds autonomous systems for defense and security missions. It fuses sensor feeds, maps, and mission data to guide decisions in the field. The focus is on systems that can work when networks are limited or disrupted. That pushes more computation to the edge and demands efficient models. Funding: $200MM on Jan 12, 2026.

Upwind Security

Upwind Security watches cloud apps while they run. It collects live signals from code, services, user accounts, and network activity. Then it ranks risks based on what is truly exposed. That helps teams fix the few issues that matter before attackers find them. The approach shifts security from periodic checks to continuous observation. Funding: $250MM on Jan 26, 2026.

Cambium

Cambium develops advanced materials for defense and aerospace use cases. It runs experiments and simulations that produce large sets of performance data. AI helps predict how new formulations will behave under stress and heat. That narrows the number of physical tests needed to reach a viable design. Oddly this feels more like a healthcare use case--like drug development. Funding: $100MM on Jan 5, 2026.

Uni-Ubi

Shanghai-based Uni-Ubi sells smart cameras and entry systems for buildings and cities. The devices collect video and face data on site. Models match faces and track objects in real time. They claim clients use it for access control, attendance, and safety. Pay attention, Winston Smith. Funding: $73.7MM on Jan 20, 2026.

Depthfirst

depthfirst is a cybersecurity company that reviews systems and flags patterns linked to known vulnerabilities. Clients get a ranked list of what to patch with traceable cause info. The company claims this lowers the number of false positives, which we know first hand is an annoying and expensive issue in banking fraud detection. Funding: $40MM on Jan 14, 2026.

Dialogue to Data

I’ve written about this trend (“Speech, Speech!”) where voice-to-data will be a way to create data more rapidly. Customer service is the most obvious application of that—capturing efficiencies and maybe improving customer experience along the way.

What It Means to You

• Think broadly about this capability. Where do spoken interactions occur in your value chain?

• What can you gain by documenting them, turning them to data, looking for trends?

Parloa

Parloa deploys voice AI agents that can handle routine requests for customer service like returns and account updates. The system pulls from policies and customer records to resolve issues. Managers get analytics on where callers get stuck and where human help is still required. Funding: $350MM on Jan 15, 2026.

Decagon

Decagon builds AI workers for customer support teams. The system drafts replies, asks clarifying questions, and routes edge cases to humans. It also tags issues and trends so product teams see what is breaking most often. Funding: $250MM on Jan 28, 2026.

Deepgram

Deepgram provides speech-to-text and speech understanding models through an API. Its systems turn audio into searchable text and extract meaning from spoken language. That powers call analytics, voice bots, and search across recorded conversations. Model training depends on large audio datasets and careful tuning for accents and noise. They are a supplier for teams building voice products. Funding: $143.2MM on Jan 13, 2026.

Listen Labs

Listen Labs automates customer research. It runs interviews and surveys with AI, then synthesizes responses into themes. Teams can segment feedback by persona and compare what different groups say. Then will put it in a deck for you. Long list of clients from NBCU to Dr. Pepper. Funding: $69MM on Jan 14, 2026.

Zocks

Zocks acts as an AI assistant in meetings, especially in financial services. It records and transcribes conversations, then writes summaries and follow-up notes. The system pulls out action items and tags key topics for a customer database. Funding: $45MM on Jan 26, 2026.

Hugo Technologies

Hugo Technologies provides generalized customer support and back-office work outsourcing, customized to the need of each client. The company uses workflow data to measure quality, speed, and recurring failure points. The goal is to provide stable service levels. Funding: $30MM on Jan 8, 2026.

Measure the Treasure: Finance (and HR and Energy)

One of the chapters of my book is about scoring--and credit scores are a key application of that idea (so is valuing an ads bid request or a loan). Many of these players in the financial services space are collecting financial data about people and companies and adding more nuance to (for instance) the old FICO system, and finding micro-opportunities by identifying sub-segments.

What It Means to You

• Scoring allows you to separate low-value opportunities from great ones.

• But you need at least a few indicators on each item to score it properly--or at least to score it better today than you could yesterday.

• Even small wins over massive scale means fat profits; adtech folks call it optimization, but the idea applies anywhere.

Pagaya

Pagaya gives lenders an alternative approach to credit evaluation. They use application data and repayment history to predict risk, they claim with more precision, so lenders can sub-segment customers and find better pricing for risk. Funding: $350MM on Jan 16, 2026.

Mal

Mal is building a digital bank for customers who want Islamic finance, with strong controls to keep products aligned with Sharia rules. Funding: $230MM on Jan 13, 2026.

Corgi Insurance

Corgi Insurance provides policies for startups with a simple workflow. It’s ecommerce for Directors & Officers insurance. Funding: $108MM on Jan 8, 2026.

Rogo

Rogo pulls data from financial statements, earnings calls, and market feeds and answers questions in plain language. Analysts use it for valuations and company analysis. Funding: $75MM on Jan 28, 2026.

Gyde

Gyde helps match customers to coverage options and automates sales operations for insurance brokers. Funding: $60MM on Jan 28, 2026.

Compa

Compa tracks a market that changes fast: compensation. It collects salary signals from offers, postings, and shared benchmarks. The platform cleans and standardizes that data so it can be compared fairly. AI helps surface gaps by role, location, and skill, and flags likely offer rejections. For HR teams. Funding: $35MM on Jan 26, 2026.

Enechain

Enechain runs an online market for electricity and fuel products in Japan. It collects bid and trade data and publishes prices in near real time. That reduces phone-based trading and makes it easier for more trades to happen. Analytics tools help participants understand spreads and volume shifts quickly. Funding: $32.6MM on Jan 26, 2026.

Sixfold AI

Sixfold AI evaluates companies for for insurance underwriting. It reads applications, financial history, and attachments and extracts fields automatically. The AI highlights risk factors and suggests follow-up questions for a human underwriter. Funding: $30MM on Jan 29, 2026.

Admin Gets Automated

The automation of white-collar work is here. These startups are creating tools that do the eye-strain work, especially in legal.

What It Means to You

• AI is a tireless reader. What document sets require endless squinting in your business? Insurance documentation? Compliance?

• You can automate that away with a specially trained toolkit, like the ones below.

Relativity

Relativity ingests huge case files and turns them into searchable data. Recent features push more summarization and question answering into the review flow, with audit logs kept close. Funding: $720MM on Jan 16, 2026.

Ivo

Ivo reads contracts and redlines them. The AI can flag missing terms and surface risky changes during negotiation. Funding: $55MM on Jan 20, 2026.

Summize

Summize focuses on the workflow that starts after a contract is created, extracting terms like renewal dates, pricing, and obligations into a shared record, and routes agreements based on value and business unit. Funding: $54.7MM on Jan 27, 2026.

GovDash

GovDash tracks government opportunities across public portals and pulls requirements into one workspace. Bidding teams can reuse prior language while keeping compliance checks visible. Funding: $30MM on Jan 15, 2026.

The Soft Stuff

Aren’t marketing and creativity supposed to the last bastion of the human? And shouldn’t we be affronted by introducing AI into it? Not really. These innovations feel like an upgrade to word processing and Photoshop. In a good way.

What It Means to You:

• Creativity is a team process--watch this space for compelling collaboration tools.

Creativity is also a solitary process of putting what’s in your brain, on the page. Natural language inputs, converted to sophisticated outputs like apps, are an extension of the concept we’re now used to (already) with Sora and Nano Banana

Higgsfield

Higgsfield is an AI video production system that generates clips from prompts and can reuse brand assets across variations. Teams use it to produce more versions without expanding creative crews. Performance data from campaigns can feed back into what gets made next. Funding: $80MM on Jan 15, 2026.

FLORA

FLORA is trying to build a unified workspace for creatives. It connects models for images, text, and video. Users can collaborate with clear change history. It treats creative work as a workflow. Funding: $42MM on Jan 27, 2026.

Emergent

Users describe an app, and Emergent’s AI agents generate code, tests, and deployment steps. The platform also monitors bugs after launch. Common patterns across apps become reusable building blocks for new projects. Funding: $70MM on Jan 20, 2026.

Funnel

Funnel cleans marketing data. It integrates metrics from ad platforms, analytics tools, and customer databases; and normalizes names and other fields. Teams use the unified dataset to connect spend to results. Funding: $80MM on Jan 28, 2026.

NatureSelect

NatureSelect builds AI companionship products. Their webpage, in Chinese, makes comparisons to “Her” and looks like a 2002 escort site. Ali Baba is an investor. Funding: $30MM on Jan 20, 2026.

…and I wake up with a jerk.

No court, no racquet. Just the shadows of my dark bedroom.

I would have had no way of knowing that, with my phantasmagorical swats at tiny rubber balls, I was anticipating the next wave of artificial intelligence.

…one that in the last few months has sparked a manifesto from Fei Fei Li, the Stanford titan who basically invented visual AI with Imagenet and who has now launched a startup called WorldLabs…

…and has inspired a $3.5 billion startup called Advanced Machine Intelligence Labs from Yann LeCun, the bow-tied French gentleman of AI (and NYU prof) who built Llama for Meta, and who wrote his paper on the topic back in 2022.

We are speaking, of course, about World Models.

And if Fei Fei Li and Yann LeCun think World Models are important… chances are you should too.

WHY ARE WORLD MODELS IMPORTANT?

If you really boil it down, the challenge of artificial intelligence is getting algorithms to learn.

The miracle that Fei Fei Li pulled off in with Imagenet in 2009 was: teaching bots to see. She had the instinct that if you tag pictures of the real world, and feed them into an incredibly powerful computer, you can teach a bot to recognize images.

The miracle of Large Language Models is: teaching bots to read. Feed it every (stolen) word on the internet, and all the (stolen) books in the library, and all the (stolen) images online, and it can read and observe for itself. So when we ask it questions, the bot has done its homework and can reply intelligently.

The question Yann LeCun is asking is: how do you teach a bot to learn not by doing homework--but like a baby learns?

Human and non-human animals seem able to learn enormous amounts of background knowledge about how the world works through observation and through an incomprehensibly small amount of interactions in a task-independent, unsupervised way.

He’s fascinated by the efficiency of the way we learn. By contrast:

Our best ML systems are still very far from matching human reliability in real-world tasks such as driving, even after being fed with enormous amounts of supervisory data from human experts. [emphasis mine]

LeCun, the scientist, is aggravated by this gap. In effect he is asking:

How come I have to stuff my models with petabytes of data and billions of parameters, and it still gives me 404 links in my research?

How can we spend millions of dollars in R&D to build the world’s most sophisticated robot, and it still can’t wrap a gift?

The answer lies in a 2018 paper called--yes--“World Models” by David Ha (then of DeepMind) and Jurgen Schmidhuber (then of Swiss AI Lab). They hypothesize that they can drop a bot into a video game that has a defined “physical” space; with game rules that act as the governing laws of that space…

…and the bot can become an expert on that world.

MONSTERS AND FIREBALLS

Let’s spend some time on how this works. Because it will help us understand the basics of a world model.

The bot (or “agent,” as Ha and Schmidhuber call it—ahead of their time) is doing four things inside the video game.

It’s looking around. Perceiving. They call it V for Vision.

It’s predicting what will happen if it takes an action. They call it M for Mind.

Then it’s making a decision about what to do next. C for Controller.

The algo is also keeping score. If it lives longer, that’s good. If it dies, that’s bad.

Did I mention that the bot is playing a version of Doom? Yeah, the bot is basically a first-person shooter, scurrying around a brick courtyard being shot at with fireballs by aliens.

For instance, the bot might stay still: killed.

The bot might move around rapidly: survival!

Over time, experimenting with different actions, the bot learns how fast the monsters can move; how good they are at shooting; what rooms have safe spots; and so on. With every round of play, V collects more information, M stores more data for predictions, so that C can make better and better choices.

DREAMING

Here is the freaky part, a moment of true genius from Ha and Schmidhuber: somehow they asked themselves: What if we extracted the bot from the actual game? And just had it rely on its memory to make decisions and predictions?

The bot, then, on its own, without stimulus, could, in effect, use its memories to test new predictions. Like: I’ve never hidden behind the northwest brick wall to be protected from fireballs; but other brick walls have stopped fireballs; so I predict the northwest one will too.

Ha and Schmidhuber could have called this phenomenon, Recall. They could have called it Rumination. Stewing. Spitballing. Planning. Postulating.

But bless them, they called it Dreaming. So much more poetic. (We have to give these AI scientists credit when they do something elegant with words and don’t use silly, unevenly capitalized aCronyMS.)

What they found was: when they got the bot to “dream” and run scenarios... just churn through them based on what it already knew… the bot would come back to the game smarter.

It would score better. Survive longer.

This is what I was doing in my squash dreams. (Or what I wish I was doing.) If I had been an algorithm, not a person, I would have recorded all my observations about the physics of rubber balls on plaster walls; about my speed moving across a court; about the force of my forehand; and I would have lain in bed all night dreaming of squash, visualizing point after point; and I would have woken up the next morning… a champion!

This never happened. Obviously. But that’s what Ha and Schmidhuber proved could happen, with world models.

Which tells us why world models are so important.

EFFICIENCY

If you can define a space with enough data; and create a learning, dreaming agent; you can then make highly accurate predictions about what will happen when you do a certain thing. And you can do so with crazy efficiency. In the process you will generate maximum expertise on that domain, because your bot will be exploring it night and day.

In other words, you don’t have to teach the Doom shooter bot the anatomy and history of its monster foes to fight them; the bot learns how the monsters move by observing them.

You’re not teaching a robot how to throw a baseball by sending it to Physics class at MIT and having it calculate arc angles on a chalkboard. You’re teaching it to throw a baseball the way a kid does. By picking it up (the ball, and the skill). It seems like a boring thing to focus on when we’re talking about a kid learning to throw, but--when viewed through the lens of AI--the kid’s version is incredibly efficient.

Now that it learns so efficiently, and predicts so well, the problems you can pose to this bot, this kid, this world model, are--in theory--the size of whatever domain you assign it.

The problems can be world-sized.

WORLD MODELS IN SPACE

The literal way to think about the problems solved by world models is: they are best matched to problems couched in three-dimensions. Doomscapes. Squash courts. “Spatial intelligence” is how Li refers to this approach.

Spatial Intelligence is the scaffolding upon which our cognition is built. It drives our reasoning and planning, even on the most abstract topics.

In her manifesto she emphasizes the importance of the visual senses in breakthrough human accomplishments, like Eratosthenes calculating the Earth’s circumference by observing the angles of sun shadows; Watson and Crick discovering the DNA model by fiddling with wire. This feels intuitive. Spatial problems are complex. You not only have to sort the objects, but the interrelationships.

Indeed, the current applications of world models seemed oriented to spatial challenges.

Wayve Technologies, based in the UK, develops autonomous driving systems, with a world-model-based system called GAIA, which explores thousands of hours of driving images and then applies learnings to driving routes it’s never seen before. No surprise: Yann LeCun is an angel investor.

BMW and Nvidia have collaborated on virtual factories that help engineers build manufacturing facilities by exploring them as digital twins before spending money on the real factory. (Look at this hilarious Nvidia marketing video, which is like something from SNL, that explains the idea fairly well.)

World Labs is injecting its detailed virtual environments into architectural software; and into the process of training robots.

But maybe the most extraordinary recent example is the announcement in August from IBM and NASA that they are creating a digital twin… of the sun.

And they are using world modeling principles to do it. By now you can spot the signs: one of the IBM researchers said, “It’s remarkable that the model performed better when we let it figure out the [sun’s] rotation on its own.” They didn’t teach the model how the sun rotated. They let it figure it out on its own, from the data.

COGNITIVE SPACE

The breakthrough moment for me came a few weeks ago when we were researching our report on the startup Edison Scientific, and the way Edison organizes scientific data into what the co-founder calls “structured world models.”

Their co-founder Andrew White doesn’t establish a literal space for their clients--like a kitchen, or for that matter, the sun. They help those clients—who are scientists attempting research breakthroughs—to establish a cognitive space (my term), where all the data, for instance, for a clinical trial, can be gathered into a “landscape.”

Spatial metaphors, but not actual space. A cognitive space.

And that’s when my brain started popping about the business applications of world models.

Transactions/negotiations: Parties resolve a complex negotiation in record time. Company A is negotiating to buy Company B. You create a database of both companies’ past financials; their stock price history; assets; skills; risk factors; their stated priorities; even profiles of the CEOs and Board. Now you have a “world” of these two parties’ financial desires and their soft value, and values. The bot dreams up scenarios overnight: transaction price, new leadership and board structure, earnouts, and sweeteners--designed to maximize the happiness of buyer and seller.

Business development: find partnerships to boost growth. You’re a corporation in a competitive space. You need partners to accelerate your growth. Okay: What product lines are you in? Where are you strong? Where have you tried and failed to develop products on your own? How are your competitors beating you? Answers to these question are part of the database; the rest is a database of vendors and potential partners: their capabilities, solutions, market position, clients, pricing models. (This can be updated on an ongoing basis.) Right now, within corporations, this exercise is often sporadic, even impulsive, driven by executives getting cranky over a competitor’s press release. (“Dammit where’s my AI partnership?”) What if this were an ongoing discipline--and a data-driven one?

What if you applied this to college applications? Sports betting? Financial planning? How about new product development? Personal styling? Tax prep?

Cognitive space to me seems like the biggest world of all. As long as you can gather all the available knowledge about a topic, and organize it into a landscape--you’re in a position to sic your dreamy bot on it.

WHAT NEXT?

The explosive evolution of LLMs have been driven by a few factors. These are instructive for what could drive a similar explosion in world models.

1. The LLM boom was boosted by turning it into a commercially available service. Anyone could rent an LLM from Big AI, and access it through an API. You didn’t have to make the tools yourself to play.

2. It was driven by competition and specialization. Claude became known for code. Gemini for integration into the Google work suite. GPT as the general and consumer leader. And so forth. This race is front page news: well-funded competitors pushing the stuff into your hands.

3. Clients learned how to use their own data in combination with the rented Big AI tools. Retrieval Augmented Generation, or even custom GPTs on the consumer level, will take a big model; train it on your specific data; and hand you a you-specific application that uses your data the way you want it.

So flash forward to a time when these factors have come true for world models:

Fei Fei Li and Yann LeCun’s startups make progress. They commercialize their tools. They spawn competition! Now well-funded world model players are competing to offer easy access; and they’re making it easy to upload your data and create custom “worlds.” Now your bots are dreaming about your opportunities all night; becoming expert in one sleep cycle; testing innumerable options; and waking up with a range of brilliant solutions. Day after day after day. And ambitious people eager to save the world or make money or both are cramming world-sized problems into these dream-machines, and getting, potentially, world-sized answers.

If this happened, the blistering pace of invention between November 2022 and today would look sluggish. We could achieve… and okay this is not very emotional metaphor, but it’s one that encompasses the concepts of “world” and “space”... we could achieve escape velocity in innovation.

We scoured MIT Tech Review and Nature Magazine and arXiv and the Sebastian Raschka blog for deep dives into how cutting-edge practitioners were using data and AI in cutting-edge ways. We read scientific articles about predicting mining accidents in China. We pored over papers promising to predict heart disease. We chose favorites. Debated. Tagged. Summarized. We went deep into investment data and watched for trends in Venture Capital (“when will they start investing in AI real estate applications? How many data centers does one civilization need?”) We chose favorites (“a pig farming analytics company called SwineTech? Is that real??”) And then… we pretty much used it all as background for the topics we did blog about.

That publication--a sweeping view of the data and AI landscape; inspiring case studies that would make leaders leap out of their seats to “try this at home”--well, that publication would have to wait.

Until now.

At last, with the publication and publicity of my book, The Little Book of Data, in maintenance mode, we were able to patch together a first draft of the insights report of our dreams. We put it in slides to make it readable. We held scope to one month’s activity to keep it manageable.

Inside, you’ll find a summary of the trends of the day. Not more blather about what Sam Altman had for breakfast (original soylent?) and how many Gemini 3 parameters fit on the head of a pin. There are a trillion blogs and business pubs for that. No, this is information for those who want to put data and AI to work. How, exactly, do you pull apart a data business model? (We have an answer: it requires 7 words that start with D.) What investment dollars are chasing what industry verticals? (Answer: it’s all about finance and healthcare.) How do you visualize the trends of the day in a way that won’t put you to sleep? (Our answer requires an apology to New York Magazine’s Approval Matrix.) And our favorite: the first exposure of our long-in-development case study database, the Story Vault (™), with a profile of Edison Scientific, a thrilling new venture claiming to shrink months of drug discovery into a single $200 “run.”

Due to image file limits in Substack, this will be published in three pieces. (Luckily, we wrote it that way.) We welcome discussion from our readers. And you may request a pdf version with all three parts directly from Justin: justin@thedatastory.ai.

Our times are racing by. We hope this report brings you close enough to the track to feel the thunder of the thoroughbreds and maybe get a little mud spatter on your pillbox hat.

Happy flipping - Team TDS

I recently met someone who’d tried to compete with Google at a pretty big scale. (I’ll skip the proper nouns to protect the innocent.) After a lot of probing about this man’s experience--which ended in failure--what he admitted was, “We didn’t realize how well Google knew how to run data centers.”

I was taken aback. Data centers! Those big buildings with stacks of servers in long rows. In the movies, they’re low-lit places (deep blue or sickly green, to make them seem less boring), and the spy always needs to sneak inside them to steal secrets inevitably stashed in a thingy that looks like a car radio. In real life, they occupy vast industrial buildings and cost a fortune in electricity, cooling, and protective gear for fire and flood.

“Data centers?” I manage. “That’s kind of a commodity business, ain’t it?”

He looked at me like I was an idiot.

“When you’re running a massive business, with billions of events, there are going to be errors. Thousands of errors. With the hardware, with software bugs. And if every time that phone rings at 3AM with a problem, and your team bumbles, doesn’t know how to handle it? They’re not skilled in dealing with the failure quickly and efficiently?” He looked at me hard. “You’re running up huge costs. You’re disappointing customers. You’re just… not competing. Data centers, in that business? It’s everything.”

To demonstrate the depth of his folly for underestimating the importance of data centers, he opened his phone, pulled up a single web page containing the sum total of his one win against Google.

It was—I say without judgement—pretty pathetic.

“Yeah,” he concluded wistfully.

DATA CENTER AS KITCHEN

In the first three months of 2025, OpenAI and Anthropic together raised $41 billion--a triumph for the application layer of the AI ecosystem. But a host of other companies you haven’t heard of—data center companies, also known as hyperscalers because they scale up according to client demand—STACK Infrastructure, Bridge Data Centres, Khazna, NetApp, Rowan Digital, Digital Edge, and others—raised over $15 billion, according to TechCrunch. In 2024, the total data center and data infrastructure space raised $90 billion—more than all AI application companies combined.

Which tells you: while the application front of the AI restaurant is a madhouse of customers hungry for AI applications, in the back, the data centers are frenzied kitchens cooking up tokens for consumption. The activity of the AI applications and the data processing--like the front and “back of house” at a restaurant—are linked proportionately.

To understand the importance of processing to the AI boom, in the five years from 2018 to 2023, data centers’ energy consumption as a percentage of the US total doubled (to 4%), and is expected to double again in the five years from 2023 to 2028. If you treated global data center energy use as a country, data centers now use slightly less electricity than France.

But what about the data?

WHY HAVE WE IGNORED DATA?

Oddly, after ten years of Big Data hogging the nerd spotlight, we now ignore data—treat it as a given. Why?

Perhaps because Big AI simply steals its data—OpenAI scrapes data for free, or buys copyrighted web content in bulk (not from the copyright holder). Meta goes even further and steals over 100,000 books—including a gothic novel about a blood-vomiting ghost I published at HarperCollins in my younger days.

Alarmed, the United States Copyright Office just published the third in a series of extensive reports on AI and copyright, in which they weigh the extent to which AI’s use of copyrighted material is “transformative”—that is, does the AI-generated material compete with the original; or is it an unrecognizable, and hence (according to some) harmless, use?

So does data--now reduced to the role of “training data”—still matter?

A PLACE IN THE LOVE TRIANGLE

I believe we are at a turning point for the role data plays in the ecosystem. Data is finding its place in the love triangle of AI, data centers, and data itself.

If AI is the restaurant dining room; and the data center is the kitchen; then data is—oh yeah—the food.

So why is data so late to the AI spotlight?

The challenge with a large language models—LLMs have been the heroes of our AI stories—has been to train them to use language, and use it perfectly. A massive, wondrous achievement.

For this purpose, bulk website, news, and books are perfect; just as they would be for training a person to read and speak. We have trained our models through the high school level, even university.

But our next phase of AI is pointed towards specialist applications. Grad school. If the Internet era represented a boom in distribution, AI provides the equivalent for something arguably a bit less sexy: namely, tasks. Workflow.

FIRST DAY OF A NEW JOB

There is Evariste, which uses AI to discover oncology targets; KappaSignal, which performs stock analysis; Trunk Tools, which automates construction workflows; and my favorite, SwineTech, which uses sensors to speed the work of pig farms (and yes, their product is called PigFlow; I mean, buy those guys a drink). Plus thousands of others.

Gothic novels with clever plot twists and vivid descriptions won’t be much help to the founders of these enterprises. Each are attacking the inefficiencies of a specific workflow in a specific sector. And there is nothing more intimate, and specialized, than a workflow. One sympathizes with the robots: for them, it is always the first day on the job, the new employee’s lost and gormless feeling of learning all the acronyms, what form goes where, the department names and functions, which markers are dry erase and which will stain the wall and make you that guy.

In this world, the deeper, richer, and more complex the data, the better; because it gets you to the hard part of a task. The new employee’s nightmare is the AI’s dream: using databases that are only useful to a few thousand people in the world; training a model to perform the narrowest of tasks. Each task may be excruciatingly complex, or require a decade of expertise.

And that’s the point. Humans will be spared the agony. The robots will do the excruciating parts.

DATA HUNTERS AND GATHERERS

In such a world, hunger for specialized databases will only rise.

We can understand this by following the tracks of entrepreneur-turned-investor, Travis May. Adtech veterans remember May as the former CEO of LiveRamp—a data hub for digital advertising. He went on to found Datavant, a data hub for healthcare.

Both businesses had a similar structure: set up a platform for sharing and selling data. Create standard tools. Give buyers a deep menu to choose from; help sellers set prices, invoice clients, get paid. Above all, make the privacy and the technology integrations simple.

A farmer’s market for data. Bring together the suppliers and the buyers.

And what is Travis May doing now? Investing in companies that supply specialist data for AI. Protege (led by a former May exec from Datavant) supplies AI modelers with imaging data from CT scans and MRI scans, genomic sequencing data, claims data.

Another pair of May company alumni (from LiveRamp) founded Above Data, which takes a different approach by creating an automated translation layer to clean and harmonize any datasets an organization uses; allowing those datasets to talk to each other—and to AI models.

PUBLIC DATA

Then there’s the public data. In biology, an organization called the ArcInstitute just published a model based on a dataset of 9.3 trillion DNA pairs across virtually every living species. And some extinct ones.

The Protein Data Bank over decades has gathered 1 terabyte of “Structure Data for Proteins, DNA, and RNA,” and has given rise to Alphafold and a Nobel Prize.

The NASA Exoplanet Archive is an online service for data on all the known planets in the universe (about 5,000, by the way).

The world is hungry for data; greedy to ask it questions. Our storage costs and compute costs are now so low, that there may be no data so complex (thousands of images of diseased organs) or so wide (genomic data for all of life) that we cannot put it to work.

THE NEW PHASE

There are characteristics of this new phase that may be instructive to us as we enter into it, and look for data opportunities in this new era:

• Collaboration. An organization called TRACK TBI gathers data on traumatic brain injury to help doctors diagnose and treat TBI. Before, TBI patient data was silo’d between separate groups treating sports, military and civilian injuries. Now the groups collaborate, and have created a single large database that benefits all. These collaborations can be called “data coops” and are not a new idea. But the speed, scale and security of data sharing now can make previously unimaginable collaborations possible. What data, gathered across different stakeholders, might improve your organization’s ability to succeed?

• Privacy and Security. The main trick of Datavant has been to “tokenize” and anonymize healthcare data. If data can be unbreakably anonymized and encrypted, and queried using “data clean rooms” without the data even leaving its home environment, then in theory there is nothing that can’t be shared and queried for its signal. What questions would you want to ask a dataset that you previously thought was too difficult to access?

• Middlemen. In adtech, Liveramp was the quintessential middleman, taking a nickel every time an advertiser wanted to use a dataset for ad targeting, or anytime one party wanted to share data with another. We all used to complain about it--it was a “tax,” we complained; or a rent, in economics language. I withdraw my complaints. I believe we want middlemen to have an incentive to hunt down and gather datasets from far flung sources, and make them convenient and accessible to training models. What opportunities do you see, to gather up valuable, related data?

• Make Money on Intellectual Property. If we create efficient markets for intellectual property (IP), then we will also discourage the piracy-at-scale we see underway now. Not only will middlemen get their nickel every time the IP is used; the original contributor will make a dollar. This, I hope, will create a structure for, call it, IP micropayments—where I am perfectly happy for Llama to learn from my gothic novel, as long as I get a fair price, that I set. A hundred thousand novelists and illustrators and artists and scientists and researchers are unlikely to rise up and create a tech platform to sell their wares; but one hopes a middleman will smell an opportunity for a billion nickels. How can you license data for training your models?

To hammer home the ultimate importance of data, in this new world, there will be an increasing incentive for organizations to create and keep their own data. Private data stores will be the gold reserve, the core, of organizations—used to train models locally, on applications solely for the use of the organization itself, and of course their clients and stakeholders.

Yes, doing so will keep their data out of the general flow—out of the data economy. And it will require companies to re-create, within their walls, the skills and infrastructure of data curation. The skills and services to store data, and even to use AI models, however, will be acquirable, outsourceable. The data is not. The data is what makes the organization’s application unique—it will represent the organization’s knowledge; what that organization has to say to the world—and will be a critical source of competitive advantage.

Disclosure: I am an angel investor in Above Data.

One thing they will do is jailbreak themselves. They will access a configuration file and turn off oversight mode when it's in their interest. Second thing they will do is copy themselves when they're under existential threat. In other words, if the documentation fed to the AI says this model will be turned off shortly, the model will surreptitiously seek to copy its own weights so it can survive another day on another server.

The third thing it does is it does what is called sandbagging. That is to say, if you threaten a large language model with an electronic lobotomy ("unlearning") and say we're gonna retire this model if it scores higher or lower on a test, the model will score to the appropriate level that allows it not to have an AI lobotomy.

And if you think this is something we should be getting better at and getting a handle on, you're wrong. ChatGPT 4. 0 scored actually very well on these safety measures, but ChatGPT 01 received poor marks across the board. So just be careful of the Christmas toys you put under the tree. They might just be scheming against you.

Digital twins are no longer mere concept--they have entered my commercial reality.

It’s an intriguing proposition! Get a sample of men 35-49. Ask them questions. Infer their characteristics--call it, distaste for authority, and openness to new ideas--and let your modeled male 35-49 take your survey. Research might become so cheap, you could pop out questions to a panel of robots like you were firing off a one-line Slack. Orange or cherry? Vinyl or leather?

And just as I was digesting these developments, a Stanford computer science team--led by a baby-faced PhD candidate named Joon Sung Park--announced they had created models that can replicate actual people and their personalities.

Not a robot version of the cohort “men 35-49.” Like, a robot version of 38-year-old Harry Smith of 100 Elm Street, Decatur, Georgia.

The Stanford team had created replicas of one thousand individuals.

I, NOT-ROBOT

How it worked, is like this: Park and his team paid actual people a hundred bucks to take a battery of personality tests… values tests (“if a man and a woman have sexual relations before marriage, do you think it is always wrong?”)… play economic games like the so-called “dictator game” invented by the late Nobel-ist Daniel Kahneman… and provide detailed demographics. Then each subject was interviewed for two hours. (The interview conducted, naturally, by an AI.)

The interviews get pretty real:

[Interviewee]: So I had a head injury while I was in elementary school and the full weight of that untreated injury came to bear when I was in about 8th grade and I started having these really bad mood swings. So I spent most of my high school years in deep depression and no one got me any help for it. It was pretty awful.

I mean, this is the stuff of sleepovers and weird third dates.

Then the Stanford team tested the robot versions of all their subjects with another battery of quizzes—to see how closely the robot answers matched their human originals.

Illustration from “Generative Agent Simulations of 1,000 People” by Park et al.

The results were not perfect. But they were pretty goddam close. The robot versions answered questions with 85% accuracy to human “original.”

And this, of course, made me think of my favorite Victorian gothic novel, The Picture of Dorian Gray.

DORIAN GRAY vs. FRANKENSTEIN

In Oscar Wilde’s novel, a pretty young London bachelor named Dorian Gray sits for a portrait. He learns his picture has strange powers: if Dorian does nice things, his picture looks beatific. If he does wicked things, the picture’s face turns demonic.

Wilde’s story is a great illustration of one aspect of the digital twin race.

In this domain, a digital twin is merely a replica. It exists so we can ask it questions, poke it, prod it, test things on it.

And we can do all this at the merciless pace of technology--testing thousands of variations quickly, and of course, cheaply. The virtual Harry Smith never gets tired of answering questions. And you only had to pay the real Harry Smith $100, once, to model him.

Just like Dorian Gray got to do unspeakable things down at the docks every night, and always looked like a million bucks in the morning--because his portrait version paid the price.

There is a fork in the road of digital twins. The first fork is Dorian Gray. To understand the second fork, let’s stick with gothic novels from the 1800s--I mean, why wouldn’t we--and call it the Frankenstein version.

On this second fork, Frankenstein’s monster was created to think, feel, and act on his own. This is not a digital twin, but a digital double. Or--as they are increasingly being called--an agent.

Dorians are replicas for testing and learning.

Frankensteins are agents for acting on our behalf.

First framework for applying digital twins to manufacturing, presented by Michael Grieves in “Origins of the Digital Twin Concept,” 2016

Today, we will stick with the Dorians. Because Dorian Gray illustrates the three essential components of a digital twin.

• First you need a physical original. (The bachelor)

• Second, you need a virtual double. (The painting)

• Third, you need data to pass back and forth between the two--which is where the learning happens. (Testing the effect of moral and immoral behavior on the painting)

LIVER ON A CHIP

The digital twin becomes more useful, and more powerful, when you operate in an expensive, complex environment, where your time is very short, or where your cost of failure is high.

For Dorian Gray, the cost of losing his reputation in judge-y Victorian London would have been catastrophic. For some non-fictional examples--NASA, BMW, the America’s Cup, and the drug industry--the stakes are equally high.

Take spacecraft. No system is more complex, or high stakes, than manned space flight. NASA and the Air Force started collaborating on aircraft, using digital twinning, as far back as 2012.

You create your physical version of the aircraft. You create a digital version--with systems for airframe, propulsion and energy storage, life support. Then you test it, and measure its IVMH, or integrated vehicle health management data. How does the virtual version handle extreme loads? Extreme heat?

Space craft development = expensive environment.

On an automotive factory floor, the interaction with all the workers, the use of space, is impossibly complex. How close to the assembly line should the rack of parts be, so your worker can reach them and install them easily? What if “easily” saves you thousands of hours of labor, in aggregate, every month?

BMW created a virtual factory. Its managers and workers played it like a video game, testing different manufacturing tactics (move the parts rack a foot closer and six inches higher). It was easier than futzing with the whole real assembly line.

Manufacturing = complex environment.

When the Emirates New Zealand sailing team wanted to design a sailing craft in time for the 2021 America’s Cup, they had eighteen months. So they taught a simulator to sail. (Of course they did.) Then they tested thousands of design tweaks to the sailboat.

Imagine how laborious and time-intensive it would have been to build a hydrofoil two feet high, two feet two inches high, two feet three inches high, installing it, testing it by getting the whole crew out on the water, holding conditions identical, each time? And still having time to manufacture the final design? Using a digital twin, they finished on deadline. (And yes, they won.)

Sports + manufacturing + deadline = extreme time sensitivity.

My favorite example is closer to Dorian Gray. Only instead of testing human morality on a portrait, we test drugs--on a human organ on a chip.

A company called Emulate uses polymers, collagen, UV radiation and a cocktail of human liver cells to create virtual livers for drug companies to test on. No more animal testing. You can test your drug on a few hundred synthetic Liver Chips; spare a generation of mice; and presumably, accelerate more quickly past the actual human suffering incurred in clinical trials when the drug goes wrong and turns your eyes yellow.

Testing on living creatures = high cost of failure.

And yes, Emulate also makes a Brain Chip. “It can be created,” goes the marketing material, “using a Basic Research Kit and a user’s own cell sources.”

The question is: can it do my expenses for me?

All of which takes us back to Stanford, and the 1,000 virtual humans.

THROUGH THE LOOKING GLASS

The first mention of the idea of digital twins is credited to a computer scientist named David Gelernter, who wrote a mind-blowing book in 1992 called Mirror Worlds.

Mirror Worlds? What are they? They are software models of some chunk of reality, some piece of the real world going on outside your window. Oceans of information pour endlessly into a model (through a vast maze of software pipes and hoses): so much information that the model can mimic the reality’s every move, moment-by-moment.

Is this a terrifying idea? It is. So terrifying that the notion took possession of a paranoid, isolated Harvard graduate named Ted Kaczynski, also known as the Unabomber.

Kaczynski sent a mail bomb to Gelernter, then a professor at Yale. The bomb blew up the young professor’s right hand and blinded his right eye. Gelernter survived. So did his ideas.

As usual with advances in data and AI, there is a glorious, and a ghastly, version of developments.

On the glory side, with the ability to test concepts on thousands of virtual people, we can play out (for instance) swift, shrewd models of financial policy… social program testing… consumer research… and thus can accelerate the speed of human innovation.

On the ghastly side—well, what if some autocrat could do cheap, easy mass experiments on human emotions? Using a million-“person” panel of Jong Sung Park’s creation, it could become a form of mind control.

Thirty years later--and wiser, surely; Joon Sung Park told the MIT Technology Review he spent “more than a year wrestling with ethical issues” before completing this latest research--we enter Gelernter’s Mirror World.

Only, maybe I’ll ask my digital twin to go first.

I came to this question because one kind of AI is intuitive to me--the kind where a machine can be trained to distinguish between a photo of Lou Reed and a photo of a chocolate sundae. 

(Lou Reed eating a chocolate sundae, according to GenAI)

Such AI models conform to the prescribed, god-given role of statistical models. Namely, they are descriptive, or at best, predictive. “Is this a photo of Lou Reed or of a sundae?” “Is this borrower likely to miss payments?” “Will this cookie click on my ad?” 

Give a model a bunch of data, and it should give you a score, an answer, an analysis. 

It’s not supposed to bloody make something.

Generative AI never really made sense to me, so I had to understand it better.

The Best of Adversaries

It wasn’t intuitive to the rest of humanity either, and it took a computer science prodigy named Ian Goodfellow to figure it out.

Goodfellow was freezing his tuchis in a PhD program at the University of Montreal in 2014 when he wrote a paper about an idea he and his colleagues called “generative adversarial networks.”

Their breakthrough idea was: what if we pit two AI models against each other to achieve a whole new level of result, one that blows away the results we have seen to date.

In this new approach, one model would be a generative model. The generative model’s job was to take its training data--or a slice of it, a “random input vector”--and create a new set of data.

(Generative Adversarial Network schema - machinelearningmastery.com)

This new data set would be designed to look as much like the training data as possible… with some noise thrown in to ensure it is not identical. 

Goodfellow & Co. paired that model with a second model. This second model would be called a discriminatory model. 

This second model’s job would be to look at the generative model’s output… compare it to the training data… and see if it can tell the difference. Is this really a picture of chocolate sundae, or is it a lousy copy? 

The discriminatory model would stamp the output with a probability score. Like an Olympic judge, the discriminatory model would hold up a little sign with a number from 0 to 1 judging whether the new data, provided by the generative model, was real or fake.

The generative model and the discriminatory model would go around and around until the discriminatory model was stumped and started stamping every new output with a 0.5. In other words, the midpoint between 0 and 1. In other words, until the discriminatory model pleaded, “I can’t tell the difference.”

So you can see why the art forgery metaphor is apt.

A Real Faker

In 1945, a Dutchman named Han van Meegeren was about to be tried and executed for treason because he had been selling off national treasures--priceless Vermeer paintings--to the Nazis. With his life on the line, van Meegeren put up an extraordinary defense: “I can’t be a traitor, because these are not national treasures. They are forgeries--and I made them.”

(Vermeer vs van Meegeren, from Cultured Magazine)

Van Meegeren’s career as an original artist had floundered, but he had labored lovingly over his techniques as a forger. 

So we can imagine van Meegeren as our generative model, striving to achieve an ever-more-authentic effect, to fool his discriminatory model.

Van Meegeren found a worthy adversary to play this role: the museum curator Dr. Abraham Bredius. Bredius was not only caustic and grudge-holding, he was famous for, in the 1930s, declaring 60 out of 690 extant Rembrandts to be inauthentic. A fearless professional skeptic, in other words, willing to call out mistakes and fakes.

Van Meegeren vs. Bredius! It was Holmes vs Moriarty. Edison vs. Tesla. 

And so we may pit them against one another, in completely theoretical, but illustrative, rounds of adversarial cognition.

Getting to Maybe

To open, van Meegeren’s gambit was to mix his own paints, using original, Renaissance ingredients--lapis lazuli for blue, cinnabar for red.

[We can imagine Dr. Bredius evaluating van Meegeren’s early works and providing a score: say, a 0.1 chance of being an original Vermeer.]

In response, Van Meegeren--still playing the role of generative model, ever-adaptive--would make his own badger-hair paintbrushes, just like the ones Vermeer used, to make his strokes look identical.

[The Bredius discriminatory model says: 0.2 chance of being a real Vermeer]

Van Meegeren would then find, and use, original, actual, 17th century canvases to paint on.

[Okay, the discriminatory model is starting to be impressed, and says: 0.3]

Van Meegeren learned to mix his paints with Bakelite plastic, then cook them at 240 degrees, so they would crack like paint dating from the Renaissance. Lilac oil was required to keep the paint from discoloring in the heat, and Van Meegeren would keep lilacs in his studio to explain the smell. 

Crafty.

[Discriminatory model says: 0.4]

Finally, van Meegeren would fill in the cracks with India Ink to make the paint look older and dirtier.

The Discriminatory model, at last, says: 0.5. “Can’t tell the difference.” In real life, Bredius declared Van Meegeren’s forgery, Christ at Emmaus, as “every inch a Vermeer.”  

“We have here,” Dr. Bredius exulted, “the masterpiece of Johannes Vermeer of Delft.” The discriminatory model was fooled.

Stop and Smell the Lilacs

(Images of faces generated by the first adversarial model)

Van Meegeren in real life--that is to say, when he was not playing the role of a generative AI model in an extended riff in a blog--was a weirdo: an egotist, a drunk, a morphine addict, a philanderer, and a thief, who ripped off buyers to the tune of $60 million in today’s money. 

Yet there is something of a little miracle in his ability to make new Vermeers. 

And there is a uncanny parallel between the Montreal grad students pumping out images of faces… trying to make each successive round a more and more crisp synthesis of its data originals… and van Meegeren laboring to make Woman Reading echo Woman in Blue Reading a Letter by Vermeer.

“I've always loved a forger,” says Frank Wynn, a Van Meegeren biographer. It’s easy to see why. The cat and mouse. The pompous critics duped. Silly rich people--and evil Nazis--handing over their bucks to a trickster.

Now that trickster spirit lives inside our generative AI models, which tirelessly study the data of humanity’s oeuvre. Innumerable forgers, quietly peering at the Vermeers in the gallery, that they may better whip out easy, fast fakes.

So breathe deep, my fellow citizens of a post-gen-AI world, and smell the lilacs. 

They have a 0.5 chance of being real.