Snorkel AI Raises $350M to Build Training Data With Humans and AI

Snorkel AI has raised $350 million in a Series E at a $3.5 billion valuation, backing its promise to build harder training data for frontier AI systems. The company says the next generation of model training will require expert-designed tasks and simulated environments, not simply more labeled examples; investors are funding that thesis at nearly three times Snorkel’s valuation 17 months ago.

Insight Partners and S32 led the round. Existing investors including Addition, Lightspeed, Greylock, GV and Wells Fargo also participated, alongside new backers that included March Capital, Blumberg Capital, Allegis Capital, Frontline, Standard and Third Point Ventures.

Snorkel said its annualized revenue run rate reached $375 million this week, more than 18 times its level over the prior 12 months. That figure is company-reported, as is the growth rate, but it helps explain why a business that began with data-labeling automation is now being valued as a supplier of finished datasets and reinforcement-learning environments.

From labeling work to tougher AI tasks

The company’s pitch is a shift in what AI developers need from data providers. Snorkel uses software and its own models alongside subject-matter experts to produce training data. It has moved from selling software that automates data labeling toward selling completed datasets and synthetic-data-based reinforcement-learning environments, where models can practice and be evaluated on tasks.

Snorkel calls this transition “Data 2.0.” Its argument is that simple labeling was largely a volume and staffing problem, while advanced AI needs precisely targeted tasks, scoring rules and environments that can expose a model’s weaknesses. The company says such work can take human experts hours, days or longer to construct and must be designed to resist a model gaming the task.

Where Snorkel says the money will go

• Expand the capacity of its agentic data factory, which creates data and environments for AI systems.
• Accelerate investment in vertical and enterprise AI.
• Extend its research and technology into new domains and modalities, while increasing open-research work.

The human role is central to the bet

Snorkel is not presenting this as a fully automated data pipeline. It argues that humans need to remain involved because training data defines how systems are evaluated and how their behavior is shaped. Its proposed loop has specialized AI agents assist experts with drafting, quality checks and task routing; human feedback then helps evaluate and improve those agents.

The company reports that, for coding-agent environments, its specialized agents improve quality-control efficiency by more than 50% and review accuracy by more than 15 points compared with a baseline combining human review and an off-the-shelf language model. Those are Snorkel’s own measurements, rather than independently verified results, and they do not establish how its approach performs across customers or domains.

A larger market for model tests

The financing also arrives less than a year after Snorkel launched its expert Data-as-a-Service offering in September 2025. The company says it works with AI labs and enterprises on data used for frontier-model evaluation and training, and plans to expand its Open Benchmarks Grants program for independently developed model tests.

Snorkel was launched commercially in 2019 after research by co-founder and CEO Alex Ratner’s team at a Stanford AI lab. The fresh capital gives it room to pursue a much bigger version of that research-rooted business. The unresolved test is whether its human-and-AI production model can keep delivering data difficult enough to improve increasingly capable systems at the scale its new valuation implies.