Snorkel AI raises $350M at $3.5B and discloses its QC agent figures

Snorkel AI raised a $350 million Series E at a $3.5 billion valuation on 22 September. Its own announcement post also carries three efficiency figures for the AI agents that run its quality control. But neither trade write-up printed them.

Insight and S32 led the round, but the investor list is where the post is fullest. The post names six participating backers, Third Point, March, Blumberg, Allegis, Standard VC and Frontline, plus eight it labels existing investors. That’s a longer roster than the coverage carried: TechCrunch listed five existing investors, and SiliconANGLE named just GV among more than a half-dozen other backers.

TechCrunch reported the round values the company at nearly triple the $1.3 billion it carried after a $100 million Series D 17 months ago. By our calculation that’s a 2.7x step up. Snorkel says its annualized revenue run rate crossed $375 million this week, up more than 18x since it launched a data-as-a-service offering. Ratner’s post puts that launch at nearly a year ago, while TechCrunch describes the growth as an eighteenfold increase over the last 12 months.

What the announcement discloses and the coverage skipped

The figures sit in a passage about coding data, so they arrive as an aside rather than a headline number. There Ratner puts numbers on how much of Snorkel’s review work its own models now do, and against what baseline.

When building coding agent environments and data, we use hundreds of specialized agents to do quality control in addition to human expert review, which currently accelerates our QC efficiency by 50%+ and improves accuracy of review by 15+ accuracy points compared to a human + off-the-shelf-only LLM review baseline.

Alex Ratner, co-founder and CEO, via Snorkel AI

Measure	Disclosed figure	Baseline
QC efficiency	50%+ faster	Human plus off-the-shelf LLM review
Review accuracy	15+ accuracy points higher	Human plus off-the-shelf LLM review
Specialized agent accuracy	2x+ improvement	Non-specialized frontier LLM
Annualized revenue run rate	$375M	Over 18x growth since launch

That matters because it cuts against the way the story got framed elsewhere. SiliconANGLE wrote that Snorkel relies on tens of thousands of human experts to generate reinforcement learning tasks, a figure absent from the company’s own post. Ratner describes something narrower: human experts supported by a stable of specialized models, sometimes hundreds per task type. If you want the mechanics of paying humans to score model output, we covered how RLHF works and what it costs.

Insight’s own account of the deal, though, lands in much the same place. Its investment note calls Snorkel an agentic data factory that pairs AI tooling with a managed expert network, with automated quality checks at every step before anything ships. Even so, Ratner argues that the humans can’t be removed. Purely synthetic data correlates with what a model already knows rather than what it needs to learn, he writes, which is the failure mode we covered in synthetic training data, where it works and where it fails.

The run rate that isn’t comparable

TechCrunch made the accounting point the announcement doesn’t. Rival data firms pay roughly 60% to 70% of top-line income straight to the specialists doing the work, so their gross figures overstate net revenue. It put Mercor’s gross annualized revenue at $2 billion, Handshake at $1 billion and Micro1 at $500 million. Snorkel told TechCrunch its expert payments sit in cost of goods sold instead, which is exactly the sort of gap a round announcement leaves out.

The founding date doesn’t agree across the three accounts either. SiliconANGLE dates it to 2019, Ratner says Snorkel began as a Stanford research project a decade ago, and TechCrunch splits the difference with a 2019 commercial launch after four years of research.

What’s worth watching is the spending commitment attached to the round. Ratner says Snorkel is significantly expanding its Open Benchmarks Grants program, which has already supported Terminal Bench, OSWorld 2.0 and Agents’ Last Exam, and promises more news soon without naming a dollar amount. A data vendor bankrolling the tests its own customers get graded on is a conflict worth tracking, and it sits next to the older question of why AI benchmarks keep lying to you.