The AI data market, weekly · 2 Nov 2026

Anthropic's IPO filing puts a number on what a lab spends with data vendors

The first frontier lab to file for a listing showed its spending with data vendors rising fast, while micro1 rewrote its licence so that it owns whatever it builds from the data it buys. Reddit's licensing backlog shrank, and Britain's insolvency filings showed what the code and records of failed software companies fetched.

Filings and courts

  • Anthropic's S-1 (26 October, price range blank) says its vendor costs for data quality and performance evaluation rose by $251 million in the first half of 2026 against the first half of 2025, after rises of $165 million in each of 2024 and 2025. It trains on datasets licensed or commissioned from third parties, naming no vendor and no total. MoreSource: Anthropic, SEC Form S-1
  • Reddit's licensing contracts still to be recognised fell from $92.1 million at the end of June to $62.4 million at the end of September, with no renewal and no new licensee. Its 10-Q says the two partners behind substantially all of that value are approaching the end of their current terms. MoreSource: Reddit, SEC Form 10-Q, third quarter 2026
  • Appen's third-quarter revenue was $73.9 million, up 34% on a year earlier, most of it from Appen China. MoreSource: Appen, ASX quarterly activity report
  • Judge Kaplan denied Runway's motion to dismiss the YouTube creators' class action (S.D.N.Y., 30 October): their claim that Runway circumvented YouTube's technical protections to scrape training video goes ahead. MoreSource: Ace Cam v. Runway AI, opinion (Dkt. 61)

What buyers publish now

  • micro1 rewrote its Data License Master Terms on 30 October: it now owns everything made by or for it from licensed data (datasets, models with their weights, outputs, synthetic data, analyses) and need not delete or retrain any of it. The data itself stays licensed, not sold. MoreSource: micro1, Data License Master Terms
  • Ooak Data raised its online estimator: a 51 to 100 person operating company with four years of records and six tools now shows $111,000 to $157,000, against $85,000 to $121,000 on 11 October. It now gives no figure at all for a company of fewer than 20 people. MoreSource: Ooak Data, partner page estimator
  • Collinear now publishes a list of the records it is buying, by industry, among them healthcare operations, banking and lending, legal, construction, security operations, marketing operations, and engineering design and CAD. MoreSource: Collinear AI, data licensing page
  • Mercor added a 25th listing: chemistry and materials science R&D records (hypotheses, experiments, results and failed work) at $10,000 to $300,000 a company. It is the first of its listings that pays whoever introduces the company nothing. MoreSource: Mercor, company referral listings
  • DataVendor still keeps 20% of what a seller earns, taken at withdrawal, but buyers now pay no transaction fee. MoreSource: DataVendor documentation

What it sold for

  • Norway's National Library pays the copyright collective Kopinor NOK 45 million a year (2026) to train Norwegian language models on the country's newspaper text, under a contract Kopinor has published. MoreSource: Kopinor and the National Library of Norway, the agreement
  • In Korea, the press foundation sold ten years of news from 30 regional outlets to the National Institute of Korean Language for about KRW 960 million in 2019, about $2,400 per outlet per year: the one completed sale of news archives the Korean press cites. Since September 2024 each outlet sets its own price for AI training. MoreSource: Journalists Association of Korea
  • Britain's insolvency filings show what the code, records and IP of failed software companies fetched: GBP 39,668 for Brainnwave's (to its lender), GBP 65,000 for Actual Experience's (to an independent buyer in India) and GBP 110,000 for Massive Analytic's (to a connected company). MoreSource: Companies House, administrators' reports
  • General Motors made over $20 million nationwide selling its drivers' data to Verisk and LexisNexis, Washington's complaint says. It pays $220 million to settle with 36 attorneys general. MoreSource: State of Washington v. General Motors, complaint; New York Attorney General

The AI data market, weekly

Issues go out on Mondays: the prices, deals, buyers and rulings that moved in the market for AI training data, each with its source.

Optional: so we write about what matters to you

Free. Issues go out on Mondays, and every one has a link to leave. Subscribing means you accept our terms and privacy policy.