Snorkel’s $350M raise at $3.5B: the ARR line buyers should not treat as audited

The story is dated 22 September 2026 in company and trade press, and it was read again on 3 October 2026. Snorkel AI said it raised $350 million at a $3.5 billion valuation, with Insight Partners and S32 in the lead. TechCrunch called the round a Series E and described it as a near-tripling of a $1.3 billion valuation attached to a $100 million Series D.

The annualized-revenue figure on a page that actually returned article text is $375 million, a company-stated run-rate via TechCrunch. That number is not an Altovar book. It is also not the same object as a lower run-rate line that a 23 September internal pulse attributed to Reuters. The Reuters URL did not return article text on this reading.

For an MSP that buys training-data and evaluation vendors, the point is who sells datasets, evals, and reinforcement-learning environments into labs and enterprises, and how a headline valuation and a headline run-rate outrun anything a customer can confirm from a wire. The dollars below stay press or company figures, not figures this desk audited.

Two live pages, both fetched on 3 October 2026, agree on the size of the round and on the valuation, and both agreements are press and company figures. TechCrunch, by Marina Temkin, time-stamped 2:56 PM PDT on 22 September 2026 (23:56 Europe/Rome), states a $350 million raise at a $3.5 billion valuation. The PR Newswire release, 22 September 2026 at 18:15 ET (00:15 Europe/Rome on 23 September), states $350 million at a $3.5 billion valuation. Where both pages say the same dollar line, the article may say that both said it. That is still a press and company figure. It is not a bank confirmation and not an Altovar measurement.

The lead investors are the same two names and not the same verb. TechCrunch says the round was led by Insight Partners and S32. The wire says it was co-led by Insight Partners and S32. Those are two formulations. They are not a cap-table fact, and this piece does not upgrade “led” or “co-led” into ownership, control, or a board outcome. No percentage appears on either fetched page, so none is stated here.

The product shift is what the pages actually describe, not a review written from a console. TechCrunch describes a move from labeling software toward a hybrid of software, synthetic data, and subject-matter experts, with completed datasets and simulated or agentic environments in the commercial story. It says Snorkel books expert cost in cost of goods sold, according to the company, because it sells reinforcement-learning environments and datasets rather than labor. The wire contrasts “Data 1.0,” volume labeling, with “Data 2.0,” expert agentic tasks, environments, and rubrics, and it says data-as-a-service launched in September 2025. That launch month is a company line. Stanford AI Lab roots, a commercial launch in 2019, and chief executive Alex Ratner are the origin facts the pages support.

Use of proceeds is only what the wire states: capacity for the data factory, vertical and enterprise AI, research and modalities, and open research including Open Benchmarks Grants. Nothing in that list is a residency promise, a service-level agreement, or a capacity commitment a buyer can point at in an order form.

The wire carries remarks from Alex Ratner, from Lonne Jaffe at Insight Partners, and from Andy Harrison at S32. Those remarks are company and investor comment. They are not audited operating metrics, and this article does not paste them as if they were. Harrison’s “highest frontier model success rate” is a marketing claim. A superlative in a funding wire is a superlative in a funding wire.

The investor names are two lists. They are not a union this article will sell as the truth. “Both” is reserved for a firm that appears on the TechCrunch page and on the wire.

Both pages name Insight Partners and S32 as the lead pair, with the verb difference already noted. Both pages also name Addition, Lightspeed, Greylock, GV, and Wells Fargo. TechCrunch lists those five as existing investors. The wire says Addition’s participation was “significant,” which is the wire’s adjective, not a share count.

Named on the wire and not in the TechCrunch list this desk could read: March Capital, Blumberg Capital, Allegis Capital, Frontline, Standard, and Third Point Ventures, presented there as new names, plus Factory, Prosperity7, and Walden Catalyst among existing names the TechCrunch text did not list. Named by TechCrunch inside the existing set already counted above, and not granted any extra role the wire did not also support. No one on either list receives an ownership percentage, a pro-rata claim, or a “lead” label beyond the two firms already discussed.

A buyer who pastes the lists together will produce a longer list than either page printed. That longer list is a spreadsheet habit, not a source. If a later filing shows a different set, the filing wins. Until then, the honest sentence is the split sentence: both, named on the wire, or named by TechCrunch.

The run-rate subsection uses one number. TechCrunch reports a company-stated annualized revenue run-rate of $375 million, and it reports the company’s “eighteenfold” description over twelve months. Label: company figure, via TechCrunch. A run-rate is not GAAP revenue. It is a pace someone annualized, on a definition this article does not have in an audit note, and it does not reveal margin, cash collection, or how much of the figure is contracted versus indicated. This piece will not invent a margin to make the headline easier to compare. The cost-of-goods remark is TechCrunch’s, including the clause “according to the company”: expert payouts in human-expert marketplaces were put by that story in a 60 to 70 percent range, while Snorkel’s accounting, as told to the reporter, puts expert cost in cost of goods sold because the sale is environments and datasets rather than labor. That is an explanation of a headline. It is not a gross-margin schedule.

The same TechCrunch page uses peer headlines as context the reporter chose: Mercor at $2 billion gross annualized revenue, Handshake at $1 billion, Micro1 at $500 million, each a press figure. They are not a market census, not an Altovar share table, and not a ranking. This article does not compute a Snorkel-versus-Mercor order, and it does not treat three reporter comparisons as the set of firms an MSP might buy.

The Reuters conflict is mandatory and unresolved. A prior desk note attributed roughly $350 million of annualized revenue to a Reuters quotation, and the URL returned a 401 bot-check on 3 October 2026, so that quote is unconfirmed. The page fetched was not a missing-document response. It was an access wall: enable JavaScript, disable an ad blocker, and no article body. The fetch did not show the words that would confirm a round, a valuation, or a run-rate. A prior-year base stored beside that unread telling is not repeated here, and neither is a calendar month for the Series D that TechCrunch did not print. Series D timing, as TechCrunch wrote it, is 17 months before the 22 September 2026 story. The two run-rate figures are not averaged, and neither is crowned. Until a person can read the Reuters page, the unread quotation stays unread. The URL is the Reuters link that did not yield article text. The readable pages remain TechCrunch and PR Newswire.

This is not a product review of Snorkel. The questions follow from the pages, not from a security audit nobody performed.

What is being bought: a labeling seat, a finished dataset, an eval or benchmark, or a reinforcement-learning environment? The wire and TechCrunch describe a shift away from volume labeling toward data-as-a-service and agentic environments. The contract should name which of those objects the order actually contains. A valuation paragraph does not name it.

Who are the humans in the loop, in which jurisdiction, and what does the processing agreement say about customer content that might be used as seed material? The fetched pages do not quote Snorkel’s processing terms. This article will not invent them. The ask is the question. If the answer is “see the trust center,” the next ask is the schedule, not the adjective on the trust center.

Which number lands on the order form? The $3.5 billion valuation is not a service-level agreement. The $375 million run-rate is not a capacity commitment. A procurement memo that copies the valuation into a risk paragraph has copied a press line. It has not copied a duty the vendor owes on a Tuesday when a dataset is late.

The papers-and-citations line, more than 250 peer-reviewed papers and more than 25,000 citations, stays a company claim from the wire. It is not a literature search this desk repeated. It does not, by itself, say anything about how a customer’s own material would be handled.

No sentence in this pack treats the raise as proof of data residency, and none borrows a sovereign-cloud slogan to decorate a funding story. A data lab’s round is evidence that investors wrote checks. It is not evidence about where a customer’s bytes will sit.

Procurement teams will paste the $3.5 billion line into a risk memo and forget that the only ARR figure readable today was a company run-rate in a trade story.Speculation — finance desk, not in the wire

Four lines are usable. Snorkel AI said it raised $350 million. Both TechCrunch and the company wire stated a $3.5 billion valuation. Both name Insight Partners and S32 in the lead, with “led” on one page and “co-led” on the other. The only annualized run-rate this stamp will carry is $375 million, a company figure via TechCrunch. “Series E” is TechCrunch’s label; the fetched wire did not print those words.

The Reuters page stayed unreadable on 3 October 2026, a 401 bot-check rather than article text, so a quotation an earlier note attributed to that page stays unconfirmed. An application-isolation check is not evidence about a third-party data lab, and this pack does not borrow one. There is no gateway chart and no comparison with Altovar revenue.