Snorkel AI has raised $350 million at a $3.5 billion valuation. It is the clearest sign yet that the market for AI training and evaluation data has become a major business of its own. Insight Partners and S32 co-led the round, announced on 23 September 2026. The company says the money will go into expanding its "agentic data factory", which produces the specialised datasets and test environments frontier labs use to train and evaluate their most capable systems.
The round comes in a month dominated by far larger checks: Mistral's €3 billion Series D and Cognition's reported raise at around $47 billion. Snorkel's round is smaller, but it shows where investors now see lasting value in the AI stack.
The Deal in Detail
Besides the two co-leads, new investors include March Capital, Blumberg Capital, Allegis Capital, Frontline, Standard and Third Point Ventures. Existing backer Addition participated significantly. Returning investors also include Greylock, Lightspeed, GV, Factory, Prosperity7, Walden Catalyst and Wells Fargo.
Co-founder and CEO Alex Ratner said the teams at the frontier want a research data partner that advances the science of data development. Insight Partners' Lonne Jaffe said Snorkel's research-grade approach to data is becoming more important for building capable AI systems.
The company lists four uses for the new money:
- Scaling the agentic data factory to meet frontier-lab demand
- Investing more in vertical and enterprise AI, especially healthcare, law and software engineering
- Taking its research into new domains and modalities
- Expanding its open research and Open Benchmarks Grants programme
From Labelling Tool to Frontier Supplier
Snorkel grew out of a Stanford research project on programmatic labelling, which uses code instead of manual annotation to create training labels at scale. Its main customers used to be enterprises trying to build task-specific models from their own data.
The company's current pitch reflects a change in how frontier models are built. Reinforcement learning, reasoning training and agent development now depend on:
- Expert-written tasks with verifiable answers
- Simulated environments where agents can take actions and be graded
- Evaluation suites difficult enough to separate one frontier model from the next
The internet cannot supply these, and they are expensive to produce well. A reasoning model learning to review legal contracts, for example, needs thousands of realistic contracts, carefully graded answers and a scoring method that cannot be gamed. An agent learning to fix software needs sandboxed codebases with tests that pass only when the fix is correct. Snorkel calls its approach a data factory because it combines domain experts, automated generation and quality control in a repeatable pipeline instead of one-off labelling contracts.
Why It Matters
For much of the past two years, AI investment has concentrated on two layers: model builders and compute. The Snorkel round is part of a growing case for a third layer, data and environments. As labs exhaust freely available text and move towards agents that act in software, the constraint shifts to high-quality, verifiable training signal.
The raise also shows how competitive this segment has become. Specialist data vendors, expert-network marketplaces and in-house lab teams are all bidding for the same pool of doctors, lawyers and senior engineers who can write and grade hard tasks. A $3.5 billion valuation implies investors think Snorkel's research background can win frontier-lab contracts on quality, not only on the size of its workforce.
The wider funding context adds to that view. On the same day, Crunchbase reported at least 114 Series A rounds of $100 million or more globally so far in 2026. Separately, Yahoo Finance cited Goldman Sachs estimates that Alphabet, Amazon, Meta and Microsoft will spend about $800 billion on capital expenditure this year. With that much going into compute, the data that makes the compute useful becomes more valuable.
What to Watch
The main risk for data suppliers is concentration. A few frontier labs account for most spending, and any of them could move more of the work in-house. Snorkel's plans for enterprise and vertical markets in healthcare, law and software are partly a hedge against that. Its Open Benchmarks effort is also a way to earn trust in a market where the credibility of evaluations is itself a selling point.
The next useful signals will be customer disclosures, any revenue figures, and whether Snorkel uses its new capital for acquisitions in a crowded AI data market.
