Snorkel AI Triples Valuation to $3.5 Billion as Training-Data Demand Explodes

The bottleneck moved. Capital followed it.Snorkel AI closed a $350 million Series E at a $3.5 billion valuation, co-led by Insight Partners and S32. Addition came in heavy among existing holders. Greylock, Lightspeed, GV, Wells Fargo, Factory, Prosperity7, and Walden Catalyst stayed in. New checks included March Capital, Blumberg, Allegis, Frontline, Standard, and Third Point.Seventeen months ago the same company raised $100 million at $1.3 billion. The new price is almost a triple. CEO and co-founder Alex Ratner told markets the annualized revenue run-rate has crossed about $350 million and, in the company’s own post, $375 million — from roughly $20 million a year earlier. That is an 18-times jump after Snorkel stopped selling mostly labeling software and started selling finished data.Ratner says the firm expects to reach profitability this year. Headcount is still modest relative to the run-rate. The check goes to more researchers and engineers, a bigger “agentic data factory,” enterprise and U.S. government coverage, safety work, and open evaluation benchmarks.What Snorkel Actually Sells NowFounded in 2019 by Stanford AI Lab researchers, Snorkel first productized programmatic labeling — software that cut the grind of supervised datasets. Last September it launched data-as-a-service: deliver the dataset, not the tool.The mix is hybrid. Subject-matter experts sit beside models that generate synthetic examples. Customers get training sets, reinforcement-learning environments — a fake developer workstation for a coding model, for example — and rubrics that score whether an agent did the job. Coding is a large demand spike. Law and medicine are the other named verticals. Frontier labs, hyperscalers, “neolabs,” enterprises, and government agencies are on the roster.Ratner’s phrase is Data 2.0. Data 1.0 was cheap labels and headcount. Data 2.0 is expert tasks that take hours or days, plus environments that teach models to act, not just classify. Labs that already scraped the public web now pay for work that does not exist on the internet.Why the Multiple MovedEvery frontier lab is starving for the next increment of hard data. Marketplace peers have printed huge gross numbers: Mercor in the billions of annualized gross, Handshake across a billion, Micro1 in the hundreds of millions. Those figures often include payouts to experts. Snorkel’s $375 million is closer to net software-plus-service revenue. Compare with care.The valuation still assumes the hunger lasts. If labs slow new model generations, the data factory idles. If open synthetic pipelines get good enough, the expert premium shrinks. If a lab brings data in-house, the contract walks. Eighteen-times growth is not a law of physics. It is a twelve-month panic buy.Snorkel’s pitch against that fade is a loop: models help experts produce data faster; experts grade the models; the factory compounds. That is also how every data vendor describes itself in 2026.Where the Money GoesCapacity first. Then healthcare, legal, and software-engineering workflows that look more like products than one-off packs. Open benchmarks are cheaper insurance than a safety scandal. Government work pays and constrains. Hiring researchers is how a Stanford lab stays a lab instead of a staffing firm.Total capital across rounds is now well past $500 million.The Category, Not Just the CompanyPicks-and-shovels for AI have migrated from GPUs everyone can name to datasets almost nobody can inspect. That is good for gross margins if quality holds. It is bad for transparency if the only audit is the vendor’s rubric.Investors just paid 3.5 billion dollars for a factory that makes the invisible input. The next test is whether $375 million of run-rate becomes a durable software business or a cyclical services spike with a research logo.