Snorkel AI Raises $350M Series E for Frontier AI Data

A useful AI training example once looked like a label. At the frontier, it can look like a simulated workplace with tools, documents, policies, hidden constraints, and a grader capable of catching a model that found a shortcut instead of completing the job.

Snorkel AI has raised a $350M Series E at a $3.5B valuation to scale the systems behind those examples. Insight Partners and S32 co-led the September 22, 2026 round, which brings Snorkel's disclosed cumulative funding to approximately $585M.

The financing matters because AI data is moving away from high-volume labeling and toward engineered environments, expert-designed tasks, evaluation rubrics, and adversarial quality controls. Snorkel is betting that this work can become a research and software system rather than a commodity labor market.

What Happened

Snorkel's Series E included significant participation from returning investor Addition. New investors Third Point Ventures, March Capital, Blumberg Capital, Allegis Capital, Frontline, and Standard VC joined returning backers Lightspeed Venture Partners, Greylock, GV, Prosperity7, Wells Fargo, Walden Catalyst Ventures, and Factory.

The round follows a $100M Series D in May 2025 that valued Snorkel at $1.3B. Snorkel had previously raised $135M through its 2021 Series C, so the new financing lifts its disclosed total to about $585M. The latest valuation is nearly 2.7 times the Series D figure in roughly 16 months.

Snorkel says it will use the capital to expand the capacity of its agentic data factory, hire researchers and engineers, deepen enterprise and government delivery, support third-party model evaluations, and extend its technology into more industries and data modalities. Those plans are forward-looking; the financing gives the company room to pursue them, not proof that every expansion will work.

From Programmatic Labels to Agentic Environments

Snorkel began as a Stanford AI Lab research project in 2015 and spun out as a company in 2019. Co-founders Alex Ratner, Chris Re, Paroma Varma, Braden Hancock, and Henry Ehrenberg built the early work around programmatic labeling, a way to let experts encode rules and create training data without manually labeling every example.

That approach addressed the economics of the earlier machine-learning era. Companies needed large volumes of labeled text, images, and documents, and manual annotation made each new model slow and expensive to build. Snorkel Flow turned expert knowledge into reusable labeling logic and gave enterprise teams a system for iterating on their data.

Agentic AI changes the unit of work. A coding agent may need a repository, development tools, product requirements, hidden tests, and a scoring system that distinguishes a clean solution from one that merely games the benchmark. A legal or medical task may require domain rules, realistic documents, privacy controls, and expert judgment across a long chain of actions. The environment itself becomes part of the data.

Snorkel's current agentic data factory combines human experts with specialized AI models and agents to build datasets, reinforcement-learning environments, benchmarks, rubrics, and evaluation systems. The company says these systems serve frontier AI labs, hyperscalers, enterprises, and U.S. government agencies, although it has not publicly named the customers behind the latest growth.

Why Investors Are Paying for Complexity

The Insight Partners investment thesis centers on a constraint that becomes more important as models improve. Simple tasks can be generated, labeled, or evaluated more cheaply, while the remaining model failures concentrate in specialized work that is harder to describe and grade. Better models do not remove the need for human judgment; they push that judgment toward the cases where mistakes carry more operational value.

Snorkel reports that its data-as-a-service business grew more than 18 times after launching in September 2025 and crossed a $375M annualized revenue run rate. Reuters reported a lower rounded figure of more than $350M, up from roughly $20M a year earlier. Both figures come from the company and are not audited disclosures, so they are evidence of reported demand rather than proof of durable margins or profitability.

The size of the round reflects the speed of that reported demand and the strategic position investors believe data providers can occupy in the AI supply chain. It also creates a higher standard. Snorkel must show that a hybrid system of experts, research, software, and automated quality control can scale without becoming a conventional services business with a more expensive technical vocabulary.

The People Behind the Shift

Alex Ratner, Snorkel's co-founder and CEO, has spent more than a decade developing the data-centric thesis from academic research into a commercial platform. Paroma Varma, co-founder and Head of Research, works on systems that let domain experts shape machine-learning data without requiring large hand-labeled datasets. Co-founder Chris Re remains a Stanford computer science professor whose research helped establish the company's technical foundation.

The current scaling assignment also reaches CTO Jimmy Hillis. Snorkel has to turn research-heavy, expert-dependent production into a repeatable engineering system while preserving the judgment that makes its output useful. Automating away too much expertise would weaken the product; adding experts without enough software leverage would weaken the economics.

What This Signals for Frontier AI

The competitive market for AI training data includes companies built around expert marketplaces, large annotation workforces, synthetic-data systems, evaluations, and managed research services. Snorkel is positioning itself around the integration of those pieces. Its claim is that complex data development should be treated as a technical discipline with tools, feedback loops, measurement, and continual improvement.

That framing matters for model builders and enterprise buyers. Procurement is moving beyond the number of annotators or examples a supplier can deliver. Buyers increasingly need to understand whether an environment resembles the real workflow, whether its rubric rewards the right behavior, whether the task exposes model weaknesses, and whether expert review can keep pace with model capability.

Snorkel's Series E supplies capital for that race, but its most important test will happen inside customer work that remains mostly private. The company now has to convert reported growth into dependable environments across domains where an incorrect grader, unrealistic tool, or missing constraint can teach a model the wrong lesson at scale.