Snorkel AI at $3.5B: Experts Weigh the Training-Data Trade

Snorkel AI now carries a $3.5 billion price tag on roughly $375 million in annualized revenue. Whether that is a sensible number depends almost entirely on which revenue you think you are looking at, and on who will be doing the expensive human work behind it three years from now.

The basics: on September 22, Snorkel closed a $350 million Series E co-led by Insight Partners and S32, up from a $1.3 billion valuation after its $100 million Series D in May 2025. Its run rate grew about 18x in a year, from around $20 million, after the company stopped leaning on labeling software and started selling finished datasets, benchmarks, evaluations and reinforcement-learning environments out of what it calls an agentic data factory. Our funding tracker shows how it stacks up against the rest of this year's AI rounds.

RelatedAnthropic IPO Filing Shows $42B Loss, $518B AI Bet

We asked people who look at this market from three different angles what the round actually says: an investor, an engineer who builds on frontier models in healthcare, and a founder who pays for model output by the token.

Aggressive, but how aggressive?

Veni Dhir is Director of Corporate Venture Capital at ADP and a founding member of its venture platform, which has deployed more than $50 million across AI, HR tech and compliance automation. She spoke to us as an AI investor, and her views are her own rather than ADP's. She did the arithmetic first. "Roughly 9-10x run-rate revenue for a company growing ~18x year over year with net-revenue quality is aggressive but not absurd in this market," she said. "The question is whether the growth is durable or a one-time lab-spending surge."

Varun Gazala, an AI Healthcare Network Innovation Engineer at Kno2 who builds multi-agent systems and clinical benchmarking frameworks, was less generous with the adjective. "Snorkel's $3.5 billion valuation on a $375 million run rate is definitely aggressive," he said, before making the case for why it might still hold. The company has "pivoted from just selling labeling software to selling finished datasets and reinforcement learning environments. If they treat human expertise as a cost of goods sold for a premium, proprietary product rather than just brokering freelance labor, that high valuation multiple starts to make more sense."

So the two agree on the direction and differ on the degree. Neither calls it a bubble. Both tie the multiple to the same thing: what kind of revenue Snorkel is booking.

Gross billings are not the same dollar

The headline comparison makes Snorkel look small. TechCrunch put Mercor at about $2 billion in gross annualized revenue, Handshake past $1 billion and Micro1 around $500 million gross. But those rivals typically pass 60 to 70 percent of that money straight through to the contractors doing the work. Snorkel books expert pay as cost of goods sold, so its figure is already close to net.

Dhir was blunt about which number investors should use. "Investors should absolutely compare on net revenue," she said. "The Mercor/Handshake gross-billings headlines are real dollars moving through the platform, but 60-70% passes through to contractors: that's a marketplace take rate, not SaaS revenue. Valuing gross billings at software multiples is how bubbles get built."

Gazala made the same point from the operating side. Many data companies "operate essentially as marketplaces," he said. "This makes a massive gross run rate look impressive, but the net margins are actually quite thin."

Why buyers still pay for humans

The whole trade rests on a bet that expert-made data stays scarce while synthetic data gets cheap. All three sources think that bet is sound, with caveats.

"It's not either/or, it's a ladder," Dhir said. "Synthetic data is getting cheaper for the middle of the distribution, but frontier capability gains now come from the tails: expert evaluations, RL environments, benchmarks that measure what actually matters. The bottleneck isn't 'data' in bulk, it's judgment-dense data." In her reading, that is why the pivot matters more than the round: "the margin moved up the stack from tools to outcomes."

Gazala works with frontier models' limits in clinical settings. "Scraping the public web worked for the initial models, but those easy gains have flattened out," he said. "Synthetic data is useful for scaling up basic patterns, but you cannot synthesize the underlying ground truth of advanced human reasoning without actual human anchors."

Mykyta Chernenko, founder of the Oslo-based book writing platform AIWriteBook, offered the view from the customer end. "I'm not in the data business. We're an application company that buys model output by the token," he said. For long, specific work, he sees expert data as the real bottleneck. "Models got very good at a paragraph a while ago. What still separates them is whether they hold a plot, a character or a fact steady across three hundred pages." Synthetic data struggles there, he said, "because a model grading its own long-form writing tends to miss the same mistakes it made."

RelatedSnorkel AI Hits $3.5B Valuation in $350M Series E

The people inside the data factory

Here is where Dhir and Gazala, without comparing notes, landed on the same risk, and neither thinks the market is pricing it.

"It's the least-discussed risk in the whole trade," Dhir said of the labeling workforce. "The 'expert' in expert-made data is often contract labor doing cognitively brutal work at piece rates. If that workforce churns or organizes, the cost structure of the entire data layer reprices. Nobody's modeling that."

Gazala went further, calling it a bigger long-term threat than the labs themselves. "As models get smarter, the tasks require doctors, senior developers, and legal experts to evaluate the outputs. The cost to source and retain that level of specialized talent will eventually squeeze margins across the board." His prediction for who survives: "the ones building the best internal tooling to make their high-end experts significantly more efficient," not the ones with the biggest contractor pool.

That lands awkwardly on Snorkel's accounting advantage. Booking expert pay as cost of goods sold makes the revenue cleaner, but it also puts every rise in expert wages directly on Snorkel's own margin.

Will the labs just do it themselves?

Dhir calls insourcing "the real valuation question." If frontier labs internalize data work, she said, "third-party data companies get squeezed to the commodity layer. Snorkel's defense is moving up-stack into evaluations and RL environments, which are harder to insource."

Chernenko expects a split rather than a wipeout. "I'd expect them to bring in the generic work and keep buying the narrow, expert stuff," he said. "The narrow stuff is expensive to staff and doesn't reuse well, which is the same reason we build on top of the big models instead of training our own."

Our take

Put the three views together and the $3.5 billion number looks less like a bet on data volume and more like a bet on a specific slice: narrow, judgment-heavy work that labs would rather buy than staff. That slice is real. It is also the slice most exposed to the cost of the people who produce it. If Snorkel's 18x year repeats, the multiple will look cheap. If expert pay climbs faster than prices, the clean net-revenue story becomes a thin-margin one, and the rivals' gross numbers will not be the only ones worth rereading.

Quotes gathered directly by GENZ TECH from sources who volunteered to comment on this story, with full attribution. Veni Dhir's comments are her own views as an investor.