Confirmed in a primary source 22 Sept 2026 Capital

The company that makes training data for the frontier labs tripled its valuation in seventeen months.

On September 22 Snorkel AI announced a 350 million dollar Series E led by Insight Partners and S32, valuing it at 3.5 billion dollars against 1.3 billion seventeen months earlier. TechCrunch reports its annualised revenue run rate at 375 million dollars, an eighteen-fold rise over twelve months, driven by a shift from selling labelling software to selling finished datasets built by subject matter experts.

The caveat that matters is that Snorkel's own announcement discloses no revenue figure at all, so the growth number is reported rather than published by the company, and an annualised run rate is one month multiplied by twelve.

Until now

Expert training data was a cost line inside the labs, bought as tooling and staffed internally. The suppliers sold software.

Now

Treat expert data as a priced input with its own supply market and its own valuations, and watch whether the labs keep paying for it while cutting the price of the models it feeds.

Snorkel spent six years selling a tool. The research it came from, four years in Stanford's AI lab, was about labelling data automatically so that companies would not need armies of annotators, and the product was software an engineering team ran themselves. What it sells now is the opposite: finished datasets, built by paid subject matter experts with synthetic generation underneath, delivered as a service to the labs. That is a different business with different margins and a different ceiling, and the market repriced it accordingly, from 1.3 billion dollars to 3.5 billion in seventeen months, with eighteen investors named in the release and two leads.

Put it next to the other thing that happened on September 22. Two frontier labs cut the price of their models by 40 and 50 percent on the same afternoon, which is the price of intelligence falling. On the same day the price of the expert data that trains intelligence went up by a multiple, and the comparable companies named alongside Snorkel, Mercor near 2 billion dollars of gross annualised revenue, Handshake past 1 billion, Micro1 near 500 million, say it is not one company's round. Models are getting cheaper to sell while the scarce human input behind them gets more expensive to buy, and only one of those two lines can be extrapolated comfortably. Treat expert data as a priced input with its own supply market and its own valuations, and watch whether the labs keep paying for it while cutting the price of the models it feeds.

How it got here

2019
Snorkel commercialises four years of Stanford AI lab research as data labelling automation software.
Apr 2025
Snorkel raises at a valuation of 1.3 billion dollars, selling tooling rather than datasets.
Sep 22
Snorkel closes 350 million dollars at 3.5 billion, led by Insight Partners and S32, on a data as a service model.
Sep 22
OpenAI and Anthropic cut model prices by 40 and 50 percent on the same day. Carded this morning.
Who has to react
The frontier labs that buy this data and have just cut what they charge for the models it trains; Mercor, Handshake, Micro1 and Scale, whose own valuations now have a public comparison; the subject matter experts themselves, doctors, lawyers and engineers being paid by the hour to write training material, whose rates are the actual cost line here; Wells Fargo, an existing investor and a regulated institution with a stake in how expert financial data is produced; the copyright plaintiffs in the OpenAI, Suno and publisher cases, since licensed and commissioned data is the alternative being priced; and the enterprise buyers who will be told that vertical accuracy is what the data spend buys.

The evidence

ObservedTuesday September 22, 2026. Snorkel AI announced a Series E of 350 million dollars at a post-money valuation of 3.5 billion dollars, led by Insight Partners and S32.

The whole observation, as recorded

New investors named: March Capital, Blumberg Capital, Allegis Capital, Frontline, Standard and Third Point Ventures. Existing investors participating: Addition, Greylock, Lightspeed, GV, Factory, Prosperity7, Walden Catalyst and Wells Fargo. The prior valuation was 1.3 billion dollars seventeen months earlier. Stated use of funds, in three parts: expanding what the company calls its agentic data factory, investing in vertical and enterprise AI applications, and deepening research into new domains and modalities, with additional funding for open research. Alex Ratner, co-founder and chief executive: The teams pushing the frontier want a research data partner who pioneers the science of data development. The business model change is the substance: Snorkel commercialised in 2019 selling software that automated data labelling, and now sells the completed datasets themselves, combining synthetic generation with paid subject matter experts, a model the industry calls data as a service. Revenue figures do not appear in the company's release; TechCrunch reports an annualised run rate of 375 million dollars and eighteen-fold growth over twelve months. Comparable companies cited in the same reporting: Mercor at about 2 billion dollars of gross annualised revenue, Handshake past 1 billion, Micro1 at about 500 million.

Established

Snorkel AI announced a 350 million dollar Series E on September 22, 2026, led by Insight Partners and S32, at a post-money valuation of 3.5 billion dollars.
The prior valuation was 1.3 billion dollars seventeen months earlier, and eighteen named investors participated.
Snorkel has shifted from selling data labelling software to selling finished datasets built with subject matter experts and synthetic generation.
Snorkel's own announcement discloses no revenue figure; the 375 million dollar annualised run rate and eighteen-fold growth are reported by TechCrunch, not published by the company.

Not established

That the revenue figure is right. An annualised run rate is one month multiplied by twelve, it is unaudited, and the company chose not to state it.
That the labs are paying more for data overall. Nothing published shows total lab spending on expert data, only what one supplier is valued at.
That this is durable demand. Frontier labs consolidate suppliers and bring functions in house regularly, and no contract term, customer name or renewal rate was disclosed.
That the price of data and the price of models are causally linked. Both moved on the same day, which is a coincidence until somebody publishes a cost breakdown.

Why it weighs this much

Signal strength

CostMedium

CommitmentMedium

ReversibilityMedium

PersistenceYears

We are scoring this honestly rather than generously, because a funding round is the weakest kind of action this brief cards. Cost is medium: 350 million dollars is real money and eighteen investors did due diligence, but it is their money at risk, not Snorkel's, and a venture cheque is a bet rather than a commitment to deliver anything. Commitment is medium for the same reason, lifted slightly by the fact that the valuation is public and sets a mark the next round has to clear. Reversibility is medium: valuations are marked down routinely and a down round costs the company nothing legally. Persistence is years rather than months because the underlying change, labs buying finished expert data instead of building it, is a supply chain decision that takes years to unwind, and that is the part worth watching rather than the price paid.

What to watch

strengthens Snorkel or a named frontier lab discloses a contract value or a multi-year expert data agreement, or another expert data company raises above a 3 billion dollar valuation, by June 30, 2027.
weakens Snorkel raises a down round, is acquired below 3.5 billion dollars, or no third party publishes revenue supporting the reported run rate, by June 30, 2027.

The words

Series E: The fifth round of venture funding. By this stage a company is normally priced on revenue rather than on a promise.
Annualised run rate: The most recent month's revenue multiplied by twelve. It flatters fast-growing companies, is not audited, and is not the same as a year's revenue.
Post-money valuation: What a company is worth on paper immediately after the new money lands, set by the price the incoming investors paid.
Data as a service: Selling the finished dataset rather than the software to build it. The customer buys the output rather than the tool, which is the change Snorkel made.

Your call

Which explanation do you believe?

A call is final. Other people's calls show after yours.