Snorkel AI, the Stanford spinout that builds the training data and testing environments behind frontier AI models, has raised $350 million at a $3.5 billion valuation — nearly triple the price tag it carried when it last raised money sixteen months ago. The round, announced September 22 and co-led by Insight Partners and S32, ranks among the largest of the year for a company whose product consumers will never see: the datasets and simulated worlds that teach models how to think.

The eye-catching number is not the valuation. It is the revenue. Reuters reported that Snorkel's annualized revenue run-rate has crossed $350 million, up from roughly $20 million a year earlier — a seventeenfold jump driven by a pivot the company made last September, from selling data-labeling software to selling finished, expert-grade data as a service. In an industry where Meta paid $14.3 billion for a 49% stake in Scale AI last year, Snorkel is arguing that the next leg of the AI boom belongs to the companies that manufacture the hardest data.

The deal in brief#

The financing was co-led by Insight Partners and S32. New money came from March Capital, Blumberg Capital, Allegis Capital, Frontline, Standard, and Third Point Ventures; existing backers including Addition, Greylock, Lightspeed, GV, Factory, Prosperity7, Walden Catalyst, and Wells Fargo participated. The $3.5 billion valuation is nearly three times the $1.3 billion figure attached to Snorkel's $100 million round in May 2025, according to Reuters.

Snorkel plans to spend the capital expanding the capacity of what it calls its agentic data factory, pushing deeper into vertical and enterprise AI, extending its research into new domains and modalities, and funding open research through initiatives like its Open Benchmarks Grants program. The company also says it expects to reach profitability this year — a rare claim for a company growing at this pace, and one worth watching against.

From labeling software to data factory#

Snorkel started in 2019 as a spinout of the Stanford AI lab, founded by a team including chief executive Alex Ratner that had pioneered data-centric AI — the idea that model progress comes as much from better data as from better architectures. The group has since published more than 250 peer-reviewed papers, and its original product was software that automated data labeling with programs instead of hand-labeling every example.

The business generating today's growth looks different. Last September Snorkel launched a data-as-a-service offering, and it now describes itself as an agentic data development platform: tens of thousands of human specialists in coding, law, medicine, and other fields devise task scenarios and grading rubrics, while thousands of AI models and agents automate the labor-intensive quality-assurance work around them. The company sells the finished data products — datasets, benchmarks, evaluation suites — rather than billing for hours of human work, a structure it says lets it pay experts more generously while protecting margins.

The company's own framing is that AI data has moved from a 1.0 era to a 2.0 era. In the first, building AI was a volume problem solved with headcount and simple labeling tasks. In the second, frontier and agentic systems need expert-crafted tasks, simulated environments, and rubrics that can take even the most qualified humans hours or days to construct. Designing that material well, the argument goes, is research work — and research work commands a research-grade price.

The data arms race#

Snorkel is not raising into a vacuum. Frontier labs have an effectively unlimited appetite for harder training and evaluation data as models move into post-training and reinforcement-learning environments — precisely the kind of data that is difficult to scrape and expensive to synthesize. TechCrunch reports peers are scaling on the same tailwind: Mercor has climbed to around $2 billion in gross annualized revenue, Handshake crossed $1 billion earlier this year, and Micro1 has reached $500 million.

Those headline figures come with a caveat that matters. Most of these companies pay out roughly 60 to 70% of their top line directly to the specialists doing the work, so their real net revenue is substantially smaller than the gross numbers suggest. Snorkel's differentiator, per the company, is that because it sells completed data products and reinforcement-learning environments rather than hours of expert labor, payments to its human experts sit in cost of goods sold rather than inflating the revenue figure — which would make its $350 million-plus run-rate a cleaner measure of the actual business. That accounting nuance is worth remembering whenever anyone in this sector throws a number at you.

What to watch#

Whether the profitability promise holds. Fast-growing AI infrastructure companies rarely claim a path to profit this early; hitting it would validate the product-not-labor model.

Quality at scale. The hybrid of human experts plus AI agents for quality assurance is Snorkel's core claim. It has to hold up as volumes keep multiplying — rubrics and measurement are the moat, and they are only as good as their weakest graders.

The competitive scrum. Scale AI now sits partly inside Meta's orbit, while Mercor, Handshake, Surge AI, and the frontier labs' own internal data teams all chase the same frontier-model budgets. Snorkel's research pedigree is its edge; scale is its risk.

The deeper signal in this raise: training data has gone from a cost center to a $3.5 billion company in sixteen months. The scarce resource in AI right now is not tokens or chips — it is qualified human judgment, packaged well enough to train machines on.