Snorkel AI Raises $350M at $3.5B Valuation as Training-Data Revenue Soars 17x

Snorkel AI, the Stanford spinout that spent its first five years selling software to help companies label their own data, has raised $350 million in a Series E round at a $3.5 billion valuation, a bet by investors that the company's abrupt turn into a supplier of finished training data for frontier AI labs is more than a passing boom. The valuation is nearly triple the $1.3 billion Snorkel commanded when it raised $100 million in May 2025, and it arrives on the back of a revenue curve that few enterprise software companies ever see.

CEO and co-founder Alex Ratner told Reuters that Snorkel's annualized revenue run-rate has crossed $350 million, up from roughly $20 million a year earlier, a roughly 17-fold jump. In a company blog post, Ratner put the figure even higher. "Since launching our new data-as-a-service offering nearly a year ago, we've grown over 18 times, and this week crossed an annualized revenue run rate of $375 million," he wrote. The company also said it expects to reach profitability this year, though growth remains the priority.

The round was co-led by Insight Partners and S32, with participation from existing backers including Addition, Greylock, Wells Fargo, Lightspeed, Alphabet's GV, Factory, Prosperity7 and Walden Catalyst. New investors March Capital, Blumberg Capital, Allegis Capital, Frontline, Standard and Third Point Ventures also joined, according to the company's announcement.

From software vendor to data factory

Founded in 2019 by researchers from the Stanford AI Lab, Snorkel built its early reputation on Snorkel Flow, a platform that used statistical techniques developed by its founders to automate the laborious work of labeling datasets for supervised learning. In September 2025 the company changed its business model, shifting from selling tools that help developers build training data to delivering the finished product itself: ready-to-use datasets, reinforcement-learning environments and evaluation rubrics sold directly to frontier labs, hyperscalers, large enterprises and the US federal government.

That pivot matters because the kind of data labs want has changed. Reinforcement learning, now central to how reasoning and agentic models are trained, requires tasks without supplied answers, simulated environments in which a model can act, and detailed grading criteria that determine whether its output is actually right. For a coding model, SiliconANGLE noted, that can mean a sandboxed developer workstation and evaluation guidance spanning several pages of security and performance requirements. Coding data is one of Snorkel's largest areas of demand, Reuters reported.

Snorkel describes its current offering as an agentic data development platform, pairing a network of tens of thousands of specialists in fields such as software, law and medicine with thousands of specialized AI models and agents. Experts design scenarios, tasks and rubrics; automation handles much of the quality-assurance work. Ratner said Snorkel sells the resulting data products rather than billing for human labor, a structure he argues lets the company pay experts more while protecting margins.

"Our strong view is that 100% of the data that labs will get value out of will have some human input in the foreseeable future," Ratner told Reuters. "But 100% of that data will have to use synthetic and automated approaches to keep up with this complexity."

The investors are making a scarcity argument. "Data is becoming more rare, more specialized, more difficult to find," said Andy Harrison, the S32 partner who co-led the deal. "If you want to train the most frontier, complex and capable models, now you need superior data."

Snorkel said it will use the capital to hire researchers and engineers, expand its enterprise and government operations, support third-party model evaluations and open-source benchmarks, and push into new industry verticals and data modalities.

Why It Matters

Snorkel's raise is the latest evidence that the humble data-labeling business has become one of the most richly valued layers of the AI stack. The market was reset in June 2025, when Meta paid $14.3 billion for a 49% stake in Scale AI, a deal that pushed several labs to diversify away from Scale and opened room for competitors. Mercor and Surge AI have since drawn heavy investor interest on the strength of rapid revenue growth, and on the same day as Snorkel's announcement, Forbes reported that Micro1 raised more than $100 million at a $4 billion valuation, an eightfold jump from its $500 million valuation a year earlier. Micro1 said in August that it had reached a $500 million gross run rate.

What distinguishes Snorkel in that crowd is its pedigree in programmatic data development. Many rivals grew out of staffing and expert-marketplace models, where revenue scales roughly with headcount. Snorkel is pitching the opposite: that research-driven automation, layered over human expertise, can produce harder data at better margins. If that holds, a valuation of roughly 10 times run-rate revenue looks less like froth and more like a claim on a durable input to model training.

The risk is concentration. A business built on selling to a handful of frontier labs is exposed to their budgets, their in-house data efforts and their appetite for synthetic data. Revenue that grew 17-fold in a year can also shrink quickly if a few large contracts move elsewhere.

What to Watch

The first test is whether Snorkel hits its profitability target this year while still expanding headcount, which would support its argument that selling data products rather than labor is structurally more efficient. Watch, too, for how much revenue comes from the federal government and large enterprises outside the frontier labs, since diversification is the clearest hedge against customer concentration.

More broadly, the rapid run of rounds for Snorkel, Micro1, Mercor and Surge suggests investors now view RL environments and expert evaluation as core infrastructure. The next signals will be whether labs keep outsourcing that work or bring more of it in-house, and whether rising synthetic-data capabilities erode the premium on human expertise that these valuations assume.