From $20M to $375M in a Year: Inside Snorkel AI's $350M Bet That Data Is the New Frontier

In a year when AI funding rounds have blurred together, Snorkel AI’s Series E stands out for one line in the fine print: the company says its annualized revenue run-rate crossed $375 million this week, up from roughly $20 million a year earlier. That is an 18x jump in twelve months, and it landed the Stanford-spinout a $350 million round at a $3.5 billion valuation — nearly triple the $1.3 billion it was worth after its May 2025 Series D.

The round, announced Tuesday, was led by Insight Partners and S32, with participation from Third Point, March, Blumberg, Allegis, Standard VC, Frontline, and existing investors Addition, Lightspeed, Greylock, GV, P7, Wells Fargo, Walden Catalyst Ventures, and Factory. Reuters reported the story first as an exclusive interview with CEO Alex Ratner.

What Snorkel actually sells

Snorkel began a decade ago as a Stanford research project in Christopher Ré’s lab, where Ratner completed his PhD. Its founding thesis was that AI progress would become increasingly “data-centric” — and that data development should therefore be studied as a research and technology problem, not a staffing and crowdsourcing one. The company commercialized that idea first as data-labeling software, then in September 2025 pivoted to selling finished expert datasets and reinforcement-learning environments directly to frontier labs, hyperscalers, so-called “neolabs,” vertical AI companies, enterprises, and U.S. government agencies.

That pivot is what the valuation is pricing in. In the reinforcement-learning era, the binding constraint on frontier model progress is no longer just GPUs — it is hard, expert-level training tasks and environments that a strong model cannot already solve. Snorkel’s argument, laid out in a blog post titled “Data 2.0 and the research era of AI data,” is that the industry has moved from “Data 1.0” (volume: scrape, label, repeat) to “Data 2.0” (curriculum: a small number of extremely complex, precisely targeted data points and environments that move the needle on an already-expert model).

The coding example is the clearest. A few years ago, useful coding data meant large volumes of raw code for pretraining and simple problems for SFT and RLHF. Today, a frontier coding dataset or RL environment has to approximate problems a senior engineer would struggle with for days or weeks, capture nuanced goals and reward signals over product-scale outputs, target specific model error modes like a calibrated curriculum, resist the model’s attempts to hack or cheat the reward, and often pass several hundred quality-control checks before it is viable. Human hours alone cannot supply that at scale.

The RSI engine, and why humans stay in the loop

The most conceptually interesting part of the announcement is what Ratner calls the “RSI engine for data” — a recursive self-improvement loop in which specialized AI models accelerate human expert output, and scaled human supervision is in turn used to continuously evaluate and improve those models. Snorkel’s internal Agentic Data Platform is built around this loop. The company shared concrete numbers: for coding-agent environments it uses hundreds of specialized QC agents alongside human review, which it says accelerates quality-control efficiency by more than 50% and improves review accuracy by over 15 points versus a human-plus-off-the-shelf-LLM baseline; scaled human feedback is then used as weak supervision, yielding a current 2x accuracy improvement over a non-specialized frontier model baseline.

The insistence on humans-in-the-loop is not just marketing. Ratner’s argument has two prongs. Mechanically, purely synthetic data is highly correlated with what a model already knows — not what it still needs to learn — and training on it (outside distillation) invites mode collapse. Normatively, data is how we measure and align AI; data developed with no humans involved “is tantamount to abdicating human oversight and alignment entirely.”

Open benchmarks as strategy

Snorkel is also expanding its Open Benchmarks Grants program, which funds open, independent benchmarks. It has supported Terminal Bench, OSWorld 2.0, and Agents’ Last Exam — the same leaderboards that frontier labs now joust over. The company’s answer to “benchmaxxing” and Goodhart’s law is not fewer benchmarks but more: a diverse ecosystem of robust, independently created, public and private benchmarks, with a majority of public ones developed in the open.

That positions Snorkel squarely in the middle of a running fight over evaluation credibility. Just this month, benchmark-editing controversies — including disputes over OpenAI’s Astra scores — put revision practices under scrutiny. A vendor whose business is building benchmark-grade data has an obvious interest in an evaluation ecosystem that is trusted, and an obvious credibility stake in keeping it independent.

Why it matters

The round is the clearest market signal yet that “data and environments” has become its own layer of the AI stack, with pricing power to match. Snorkel is an OpenAI-listed partner, and its customer list reads like the industry’s org chart — frontier labs, hyperscalers, and the U.S. government. Its revenue trajectory (from ~$20M to $375M in a year) suggests frontier labs are spending on expert data at a scale that resembles what they once spent on labeling for chatbots, but at expert task complexity and expert prices.

There are open questions. A $3.5B valuation on a ~$375M run-rate is roughly 9.3x revenue — aggressive for a services-heavy business, defensible only if the data-as-a-service contracts are durable and the RSI loop compounds as advertised. And the model-collapse debate over synthetic data is not settled; if pure synthetic pipelines do improve, the human-in-the-loop premium narrows. But for now, the market has rendered its verdict on where a piece of the frontier’s scarcity lives: in the datasets and environments that teach models what they do not yet know.