Snorkel AI Raises $350M Series E at $3.5B Valuation to Build the Frontier Lab for AI Data

Snorkel AI announced a $350 million Series E financing at a $3.5 billion valuation on September 22, 2026, in a company blog post written by co-founder and CEO Alex Ratner. Insight and S32 led the round.

Participation also came from Third Point, March, Blumberg, Allegis, Standard VC, and Frontline, alongside existing investors Addition, Lightspeed, Greylock, GV, P7, Wells Fargo, Walden Catalyst Ventures, and Factory.

Ratner said the company’s data-as-a-service offering, launched nearly a year earlier, had grown over 18x and crossed a $375 million annualized revenue run rate in the week of the announcement. He said Snorkel now partners with frontier labs, hyperscalers, neolabs, vertical AI leaders, enterprises, and U.S. government agencies.

Prior Funding and Company Background

The Series E follows Snorkel’s $100 million Series D at a $1.3 billion valuation, announced in a company release dated May 29, 2025. Addition led that round, with participation from Prosperity 7 Ventures, existing investors Greylock and Lightspeed, and strategic investors including BNY and QBE Ventures. The 2025 release said the round brought Snorkel’s total funding to $237 million since the company’s 2019 founding.

That release, datelined Redwood City, California, also announced general availability of Snorkel Evaluate and Snorkel Expert Data-as-a-Service, two new offerings on the company’s AI Data Development Platform for evaluating and tuning specialized AI systems. It said the Series D capital would support expansion of the company’s engineering, research, and go-to-market efforts.

Data 2.0 and the Agentic Data Platform

Ratner’s post frames the new financing around what he calls Data 2.0. Snorkel began as a research project at Stanford roughly a decade ago, he wrote, on the thesis that AI progress would become increasingly data-centric and that data development should be studied as a research and technology problem rather than simply a staffing and crowdsourcing exercise.

The post describes a first mile of AI data development, labeled Data 1.0, as driven by volumes of simpler data and largely a staffing and logistics problem, and a last mile, labeled Data 2.0, as driven by the quality of more complex data and primarily a research and technology problem. Using coding data as an example, it says frontier reinforcement-learning datasets and environments must now approximate software problems a senior engineer might struggle with for days or weeks, capture nuanced reward signals over product-scale outputs, target specific model error modes, resist attempts by advanced models to hack or cheat, and pass several hundred quality-control checks.

To produce that data, the post describes an internal Agentic Data Platform in which human experts are supported by specialized AI models and agents, sometimes hundreds per task type. Those agents expand expert sketches into full environments and data instances, guide human effort toward distributional targets and model error modes, provide live feedback and post-submission quality control, and route subcomponents to the right experts for review. Human feedback at scale is then used as weak supervision to measure and improve the agents, a compounding loop Ratner calls an RSI, or recursive self-improvement, engine for data.

Ratner reported that in coding data work, specialized agents running quality control alongside human expert review currently accelerate QC efficiency by 50%+ and improve review accuracy by 15+ accuracy points compared with a baseline of human reviewers using only off-the-shelf LLMs. He said the scaled human feedback applied as weak supervision then yields a current 2x+ accuracy improvement over a non-specialized frontier LLM baseline, and that figures like these continuously improve as the loop runs.

Open Benchmarks and Research Plans

Snorkel also said it will significantly expand its Open Benchmarks Grants program, which funds open, independent AI benchmarks. The program’s page states the effort is backed by a $3 million commitment to fund open-source datasets, benchmarks, and evaluation artifacts, and that applications are reviewed on a rolling basis. Proposals are reviewed with input from a steering committee of academic and industry researchers, and the page states Snorkel does not direct the committee’s decisions.

Benchmarks listed on the page include the Terminal-Bench line, OSWorld 2.0 with the University of Hong Kong’s XLANG Lab, Agent’s Last Exam with UC Berkeley’s RDI, Continual Learning Bench with UC Berkeley’s SkyLab and UW-Madison, SlopCode Bench with UW-Madison, DARPA, and the National Science Foundation, and Senior SWE-bench with the University of Wisconsin and Princeton University.

Beyond benchmarks, Ratner said Snorkel plans to invest heavily in support of specialized AI models, writing that the company expects an increasingly large share of the world’s data development to focus on specialized AI over the next several years. He also said a significant portion of Snorkel’s research and development in the years ahead will focus on data and environment development for AI alignment and safety.