Snorkel AI raises $350M after turning its software into a data service

Why it matters

Snorkel's valuation jump shows investors moving upstream from AI applications into the expert datasets, evaluations and simulated environments required to train increasingly capable agents.

Snorkel AI has raised $350 million at a $3.5 billion valuation, rewarding co-founder and CEO Alex Ratner (@ajratner) for a sharp turn from selling data-labeling software to delivering the finished datasets and training environments that frontier AI labs need.

The financing, reported by Reuters on September 22nd, was led by Insight Partners and S32. Existing investors Addition, Greylock and Wells Fargo also participated. The deal nearly tripled Snorkel's valuation from the $1.3 billion price attached to its $100 million round in May 2025.

Ratner's route to a $3.5 billion company began as a Stanford computer science doctorate. Working under professor and Snorkel co-founder Christopher Re, he developed "data programming," an approach that lets experts produce training data with rules and structured knowledge instead of manually labeling each example. Ratner led the open-source Snorkel project at Stanford before the researchers spun the work into a company in 2019, according to his company biography.

That research produced software for enterprises building machine-learning applications. The business generating Snorkel's current growth looks different: customers increasingly buy completed data products, evaluations and reinforcement-learning environments rather than tools for assembling those materials themselves.

The pivot produced the round

Snorkel unveiled Expert Data-as-a-Service in May 2025, alongside an evaluation product and its previous $100 million financing. Reuters reported that Snorkel launched the data-service business in September 2025.

Snorkel now says its annualized revenue run-rate has passed $350 million, up from roughly $20 million a year earlier. That 17.5-fold increase is a company-reported annualized snapshot, rather than a full year of audited revenue. At the stated figures, investors valued Snorkel at about 10 times its run-rate.

The growth gives context for Insight and S32 resetting the valuation only 16 months after Snorkel's previous round. Frontier labs are spending heavily to improve coding, tool use and expert reasoning, where another large scrape of public internet text offers diminishing returns. Those systems need carefully designed tasks, grading rubrics, edge cases and simulated environments that can test whether an agent completes complicated work correctly.

Snorkel says its platform combines human specialists with thousands of narrower AI models and agents. Experts in fields including coding, law and medicine create scenarios and evaluation standards. Automated systems then help generate, check and refine the resulting data.

ratner told Reuters that valuable training data will continue to require human input, while synthetic and automated methods will be needed to produce it at sufficient scale. That position preserves the role of subject-matter experts while treating automation as the machinery that makes their knowledge commercially useful.

Selling the output instead of the hours

Snorkel draws on a network of tens of thousands of specialists, according to Ratner. Snorkel sells the datasets and environments those experts help produce rather than billing customers directly for units of human labor.

That distinction is central to Ratner's business model. A conventional annotation provider grows by recruiting and managing additional workers. Snorkel is trying to turn expert judgment into a repeatable data product, using software to handle quality assurance and other work that would otherwise expand headcount alongside revenue.

The model also pulls Snorkel closer to the operations of frontier-lab customers. Designing a reinforcement-learning environment for a coding agent requires understanding the behavior a lab wants to produce, the failure modes it needs to catch and the signals used to reward the model. That relationship can be deeper than a software license, though it also creates demanding delivery work and exposes Snorkel to the spending cycles of frontier labs.

Snorkel says it also serves hyperscalers, enterprises and the US federal government. Coding is among its largest areas of demand. Ratner plans to use the round to hire researchers and engineers, expand enterprise and government operations, support third-party model evaluations and enter additional industries and data formats.

AI labs are buying harder tests

The financing lands in a market reshaped by Meta's $14.3 billion purchase of a 49% stake in Scale AI in June 2025. That transaction turned training-data suppliers into strategic infrastructure and pushed customers and investors to look for additional providers. Mercor and Surge AI have also attracted capital as labs compete for expert-generated training and evaluation material.

Snorkel's pitch comes from a different technical lineage than the large labeling marketplaces. Ratner and his collaborators spent years working on programmatic methods for encoding expert knowledge into training data. Snorkel now applies that foundation to the tasks and environments used for post-training, agent development and model evaluation.

The new round gives Ratner enough capital to expand that delivery system while frontier-lab demand remains high. Snorkel says it expects to become profitable in 2026 even as it continues hiring, an assertion that will depend on whether automation keeps the cost of producing specialized datasets from rising as quickly as sales.

Snorkel's transformation also changes what investors are underwriting. The $1.3 billion company funded in May 2025 still presented a unified software platform as the center of its strategy. The $3.5 billion version is being valued for supplying the data products themselves.

Ratner has spent a decade arguing that AI performance would eventually be constrained by data rather than model architecture alone. Frontier models have made that constraint expensive enough to support a $350 million financing. Snorkel's next job is to show that expert knowledge can be manufactured with the consistency of software, without reducing the people providing that knowledge to another interchangeable labor pool.