What Anthropic's IPO filing says about buying training data
· Company Data Advisors
Anthropic's IPO filing says it trains on data licensed or commissioned from third parties and vets its provenance. What that means for owners licensing records.
Anthropic's IPO filing confirms that a frontier AI lab trains its models partly on "datasets licensed or commissioned from third parties," that it checks those sources for "quality, provenance, and suitability," and that in many cases it writes contract terms meant to discourage personal data in what it buys. It also reports a fast-growing bill for vendors of "data quality and performance evaluation services," a category that isn't data licensing alone and doesn't measure what any company is paid for its records. For an owner, the filing's most useful passages are about rights: they explain why AI data buyers ask you to prove the records are yours to license.
The filing is Amendment No. 2 to Anthropic's Form S-1 registration statement, filed with the SEC on November 10, 2026, ahead of a planned Nasdaq listing. We read the relevant sections of the filing itself. Everything quoted below comes from it unless another source is linked.
Where does Anthropic's training data come from?
Anthropic's filing lists its sources of training data as including "publicly available information from the internet, datasets licensed or commissioned from third parties, and synthetic data generated by our own models." It says Anthropic uses that data "across the stages of model development, including pre-training and post-training." Pre-training is the first, broad stage of building a model; post-training is the later stage that teaches it to follow instructions and do specific tasks well.
The filing also lists, among the things its success depends on, the need to "secure and maintain relationships with data providers, licensors, and partners to obtain the diverse, high-quality datasets necessary for world-class model performance." And it warns that Anthropic "may not be able to obtain the datasets that we need, including through licenses or strategic data partnerships."
Licensed data is a named input in a public company's account of how its product is made, and losing access to it is listed as a business risk.
How much does Anthropic spend on data vendors?
Anthropic doesn't report a single number for data purchases. It reports changes in one cost, "vendor costs related to data quality and performance evaluation services," which sits inside a broader line called other operating and administrative expenses. The filing gives three year-on-year increases in that vendor cost:
| Period | Increase in vendor costs for data quality and performance evaluation | Source |
|---|---|---|
| 2024 vs. 2023 | $165 million | Form S-1/A, Nov 10, 2026 |
| 2025 vs. 2024 | $165 million | Form S-1/A, Nov 10, 2026 |
| First nine months of 2026 vs. the same period of 2025 | $757 million | Form S-1/A, Nov 10, 2026 |
Three cautions before anyone repeats these numbers. First, they are increases, not totals; the filing doesn't give the size of the vendor cost itself. Second, the category combines data quality work with "performance evaluation services," which the filing doesn't define (we read it as testing how well a model performs tasks), so it is not a measure of data licensing alone. Third, the filing doesn't name the vendors or say how much of the money, if any, ends up with companies that licensed their records.
Our view: the 2026 jump is the clearest sign yet, in a public filing, that a frontier lab is paying its outside data quality and evaluation vendors much more than it did a year ago. The filing doesn't say whether any of that money reaches the data companies that run programs for licensing business records. It says nothing about what any one company's records are worth, and cold emails that cite lab spending as proof of a payout are reading more into it than it holds.
Does Anthropic buy records directly from companies?
The filing says "licensed or commissioned from third parties" and doesn't say who those third parties are. Writing in WIRED on November 6, 2026, Steven Levy reported that Anthropic acknowledged a big market for data but "told me that it doesn't buy workflow from live companies like WIRED." As reported, that statement covers live companies. It leaves open records from companies that have closed, and Google's winning bid for Spirit Airlines' records in bankruptcy, still subject to court approval, shows a lab can bid for records directly (our post on the Spirit case).
Our view: for a company that is still operating, the usual route to a lab is a data company in the middle, which licenses records from the owner, de-identifies them and turns them into training environments or evaluations it sells on. We explain that structure in how the market for company data works. An owner's contract is with that data company, not with the lab, which is why the data company's terms and conduct matter so much.
Why do AI data buyers ask you to prove you own the records?
The filing's risk factors spell out what happens to a lab when a supplier's rights turn out to be weak. If Anthropic or the "third parties from which we source data lack sufficient rights to collect, use, process, share, or analyze certain third-party data," including for training AI models, "we could be subject to legal claims, regulatory scrutiny, or contractual liability," it says. "Such claims could require us to limit functionality, segregate data, or retrain models."
A second passage widens the point to paperwork: "inadequate documentation of training data sources, provenance, or chain of custody could create legal exposure," and could "raise concerns during due diligence by investors, partners, or potential acquirers."
Our view: a lab that may have to retrain a model has every reason to push the rights question onto whoever supplies the data. The data companies that license records put rights at the front of their programs:
- SuperAnnotate says a company needs "records built up over time and the rights to license them."
- SourceX tells owners that "many AI developers prefer data licensed with the owner's agreement over data collected without it," and its process asks you to confirm you have the rights to license the data and that it hasn't been licensed for AI training before. We compare those confirmations across buyers in what you're agreeing to when you license company data.
- micro1 says companies "retain ownership of their underlying data," so whether you hold the rights to license that data is a question for you and your counsel before you sign.
So when a buyer asks whether your customer contracts, vendor NDAs or employee policies allow the license, it isn't being difficult. It is asking a question at least one lab has told public investors it worries about.
What does the filing say about personal data?
Anthropic says it "in many cases include[s] contractual restrictions intended to discourage the inclusion of personal data in sourced datasets," and applies "data curation and filtering processes to sourced data." The phrase "in many cases" matters: it is not a promise that every dataset is free of personal data, and the restrictions are aimed at the suppliers who sell to Anthropic.
The data companies that license business records, for their part, say they de-identify what they license before passing it on to their customers. Labelbox Prism, for example, says that when data is exported, "we first remove all personally identifiable information (PII) and then transfer it into our secure infrastructure." Four buyers' statements side by side:
| Buyer | What its page says about personal data |
|---|---|
| micro1 | Personal and identifying information is removed or transformed before data is shared with third parties; it commits not to resell or distribute data until that is done |
| Labelbox Prism | Removes all personally identifiable information when data is exported, before transfer to its infrastructure |
| SuperAnnotate | Identifying detail is removed "to the standard your team sets" before licensing |
| SourceX | De-identifies data "before it is used or shared" |
Sources: each buyer's own page, read November 15, 2026. Each buyer's page in our index links to its program.
None of these promises takes the decision out of your hands. Our post on licensing records that include employee and customer data covers what an owner should exclude before anything leaves the company, whatever the buyer promises to remove afterwards.
What the filing doesn't tell you
The filing doesn't say which data companies Anthropic buys from, what it pays per dataset, whether it takes business records at all as opposed to expert-written tasks, or how its suppliers source what they sell. It doesn't describe any program for licensing company records. Anyone who tells you a lab is paying a given amount for records like yours is going beyond what the lab has disclosed.
Where we come in
Owners who decide to license their records face a long and growing list of data companies, each with its own criteria, terms and handling of personal data, and the lab at the end of the chain is usually out of sight. We track every major program and have direct contacts at the major data buyers and at AI labs. We compare offers on the whole deal, not just the headline number, check where the data would go and how it's handled, and nothing about your company is shared until you approve a named buyer. Participating buyers may pay us referral fees, and we tell you which before any introduction. Owners pay us nothing, and you can apply to any buyer directly. Our buyer index lists every program we track.
What to do with this
Before you talk to any buyer, put together what a lab's lawyers would want to see at the end of the chain:
- A rights list. For each system you would license from, note who created the records and whether any client contract, vendor NDA, platform terms or employee policy limits reuse.
- A provenance note. Where the records came from, how long you've held them, and whether any part has been licensed for AI training before.
- An exclusion list. The personal and confidential material you will remove yourself, before export, rather than relying on the buyer's scrubbing.
- Three questions for each buyer: who its customers are, what it promises them about rights and personal data, and what it will ask you to confirm in return.
If a buyer can't answer the fourth item clearly, see our checklist for checking that an AI data buyer is legitimate.
Programs in this post
More from the blog
- Is my company too small to license its data to AI buyers?
Most AI data buyers with a size floor set it at 20 to 50 employees, but law, accounting and medical firms qualify smaller, and nine programs state no minimum.
- One payment or a revenue share: how AI data buyers pay
The largest group of AI data buyers pays once. A few pay a share of every resale, and one now pays nothing upfront. How each model works and what to ask.
- How to check that an AI data buyer is legitimate
Six checks an owner can run in an afternoon before sharing records. Who the buyer is, what you pay, what it asks first, personal data, payment and license terms.
See which buyers fit your company
Eight quick questions against every program in our buyer index. Participating buyers may pay us referral fees.