Snorkel and micro1 Show What Human Data Is Now Worth

The quiet business behind frontier AI had a loud Tuesday. On September 22, Snorkel AI announced a $350 million round at a $3.5 billion valuation. The same day, Forbes reported that micro1 had raised more than $100 million at a $4 billion valuation. Neither company trains models. Both sell the thing models still cannot make for themselves: carefully made data, much of it produced by paid human experts.

The two rounds are worth reading together, because they show the same market from two different business models and with two different ways of counting revenue.

Snorkel: from software to finished data

Snorkel started in 2019 out of the Stanford AI lab as a software company that helped enterprises label data programmatically. The new round, a Series E co-led by Insight Partners and S32, prices a different company. Reuters, which broke the story, reported that Snorkel now supplies training data and reinforcement learning environments to AI developers, and that the valuation is nearly triple the $1.3 billion it reached when it raised $100 million in May 2025. Addition, Greylock, Lightspeed, GV and Wells Fargo are among the other backers named across the press release and coverage.

The turn came in September 2025, when Snorkel launched what it calls Expert Data-as-a-Service. Instead of selling a labeling tool, it now delivers datasets, benchmarks, evaluations and RL environments, such as simulated developer workstations used to train coding models. SiliconANGLE reported that the work draws on tens of thousands of human experts.

The revenue numbers need a small asterisk. Reuters put Snorkel's annualized run rate at $350 million, up from roughly $20 million a year earlier. TechCrunch and SiliconANGLE quoted CEO Alex Ratner saying the company had grown more than 18 times since the data service launched and "this week crossed an annualized revenue run rate of $375 million." Either way, these are company-reported figures, not audited results. Reuters also reported that Snorkel expects to be profitable this year, which is unusual for a company growing that fast.

Ratner's thesis, as he put it to Reuters, is that "100% of the data that labs will get value out of will have some human input in the foreseeable future." Synthetic data is part of Snorkel's pipeline, but the pitch is that humans remain in the loop.

micro1: from recruiting to the labs

micro1's path is stranger and faster. Forbes reporter Anna Tong wrote that at the start of 2025, founder Ali Ansari, now 25, was running an AI recruiting business doing about $7 million a year. The change started when another data-labeling company asked micro1 to help it recruit hundreds of engineers. Ansari concluded the bigger business was supplying the labs directly.

The new round of more than $100 million values micro1 eight times higher than the $500 million valuation it raised at in September 2025. Forbes reported that customers include frontier labs, Microsoft, Amazon and the robotics company 1X, and that two frontier labs and two xAI cofounders participated. No lead investor was named in the coverage we read.

The product mix goes beyond text. Forbes reported that micro1 builds RL "gym" environments that simulate workplaces, and that defunct companies' Slack and email threads are especially useful raw material. It pays $50 to $90 an hour for people to annotate videos of robots doing tasks. It even bid $12.5 million for a portion of bankrupt Spirit Airlines' operational data, a late bid against Google's winning $10 million offer.

Gross, net and the number on the slide

Here is where the two companies stop being comparable. In August, TechCrunch reported that micro1 had reached a $500 million gross annual run rate. That figure includes money that passes straight through to the contract experts. TechCrunch put micro1's net annual run rate at $150 million to $200 million, and noted that expert-marketplace companies like it pay out roughly 60% to 70% of top-line income to the specialists doing the work.

The larger marketplaces report the same way. TechCrunch said Mercor hit $2 billion in gross annualized revenue this summer and Handshake reached $1 billion earlier this year.

Snorkel frames its number differently. The company told TechCrunch that because it sells finished datasets and environments rather than labor, expert payments sit in its cost of goods sold. That is a real distinction in how the business is structured, though it does not mean the experts cost less. It means a buyer is paying for a product with a margin attached, not renting people by the hour. Readers comparing a $4 billion valuation on $500 million gross against a $3.5 billion valuation on $350 million to $375 million should keep that in mind. On a net basis, Snorkel is arguably the larger business.

Why the labs keep paying

The demand side explains the prices. As models move from simple chat into coding, law, medicine and long-running agent work, the data that improves them has to be written or judged by people who actually know those fields. Reuters framed the boom as AI developers moving beyond simple labeling to higher-stakes datasets. RL environments add a second layer: labs need realistic sandboxes where an agent can try a task and be graded, and building those takes both engineers and domain experts.

It is also a crowded field. Reuters listed Scale AI, Mercor and Surge AI as Snorkel competitors, and Meta's $14.3 billion investment in Scale in June 2025 pushed several labs to diversify suppliers. micro1's Ansari has been pitching on trust as well, telling TechCrunch in August that some human data companies work with foreign adversaries.

The risk in the model

The obvious risk is concentration. A small number of frontier labs buy most of this data, and each one can change suppliers, bring work in-house, or lean harder on synthetic data. Valuations built on run rates that grew 18 times in a year assume that spending keeps rising.

The less obvious risk is quality. The whole premise is that human judgment is worth paying for. That only holds if the humans are actually doing the judging, a question that surfaced in the same news cycle, when contractors on OpenAI rating projects were reported fired for using AI to do the work. For Snorkel and micro1, verifying that their experts are real and working unaided is not an operations detail. It is the product.

For now, the market has answered the question these rounds were testing. Frontier AI still runs on human expertise, and the companies that can supply it at scale are being priced like infrastructure.