The AI training data market runs on numbers that are scattered across press releases, court filings, IPO disclosures, and analyst reports. This page collects the ones that matter, each sourced and dated, and is updated monthly as new deals, rulings, and funding rounds land. Last updated August 2026.
The market
The global AI training dataset market is projected to reach 16.3 billion dollars by 2033, growing at 22.6 percent annually from 2026, according to Grand View Research.
Image and video data is the largest segment of that market, holding a 41.9 percent revenue share in 2025 (Grand View Research).
North America accounts for the largest regional share at 35.1 percent of global revenue in 2025 (Grand View Research).
Research firm Epoch AI has projected that the stock of high-quality public text data could be effectively exhausted for training purposes between 2026 and 2032, one reason labs increasingly license non-public data.
What licensing deals actually pay
News Corp and OpenAI signed the largest disclosed content licensing deal to date: more than 250 million dollars over five years, announced May 2024 (Quartz).
Google reportedly pays Reddit about 60 million dollars per year for data access (2024 reports).
OpenAI's Reddit arrangement has been reported at roughly 70 million dollars per year.
Reddit disclosed roughly 203 million dollars in aggregate data licensing contract value around its 2024 IPO.
Shutterstock generated 104 million dollars in AI licensing revenue in 2023 from deals with OpenAI, Meta, Google, Amazon, and Apple, one of the few audited figures in the market.
Amazon's deal with The New York Times has been reported at 20 to 25 million dollars per year (2025).
Meta's arrangement with News Corp has been reported at up to 50 million dollars per year.
OpenAI's deals with Dotdash Meredith (reported around 16 million dollars per year) and the Financial Times (reported 5 to 10 million dollars per year) show the mid-market range.
Perplexity committed 42.5 million dollars to its publisher revenue-sharing pool (2024).
The major music labels converted their lawsuits against Suno and Udio into licensing agreements in late 2025, establishing that licensed catalogs, not scraped audio, are the going-forward standard for music training data.
Most training data contracts remain private; disclosed figures are the visible edge of a larger market.
Copyright and the courts
Anthropic's 1.5 billion dollar settlement with authors received final court approval in July 2026, the largest copyright settlement in United States history (Authors Guild, TechCrunch).
That settlement works out to roughly 3,000 dollars per book across approximately 500,000 covered works.
The February 2025 ruling in Thomson Reuters v. Ross Intelligence was the first major United States decision to reject a fair use defense for AI training.
Mid-2025 rulings in cases against Anthropic and Meta drew a sharp line between training itself and how the data was acquired: acquisition without authorization is where liability concentrated, which is why provenance became a purchasing requirement.
Dozens of AI training data copyright cases remain active in United States courts, with our AI training data lawsuits guide tracking the ones that matter for buyers.
Vendor economics
Meta paid 14.3 billion dollars for a 49 percent stake in Scale AI in June 2025, the deal that pushed several frontier labs to diversify their data vendors.
Mercor raised at a 10 billion dollar valuation in October 2025 and was reported in mid-2026 to be in talks at around 20 billion, with reported annualized revenue above 2 billion dollars (TechCrunch).
Handshake AI reached a reported 1 billion dollars in gross annualized revenue within roughly 18 months of launching its expert data business (Sacra).
Surge AI bootstrapped to reported revenues above 1 billion dollars before taking outside capital, unusual in a market where its peers raised heavily.
Specialist experts producing training data command 100 to 125 dollars per hour at the university-pipeline networks (Sacra).
Prime Intellect raised 180 million dollars and hosts more than 2,500 community-built RL environments, the open-source end of the agent training market.
Deeptune raised a 43 million dollar Series A before Mercor acquired it in July 2026, a marker of how quickly RL environment startups became acquisition targets.
The licensed data marketplace
Troveo, the licensed data marketplace publishing this page, contributes its own figures to the record: more than 8 million hours of licensed video and 4 million hours of audio cleared for AI training, sourced from more than 7,000 rights holders globally, around 95 percent of them signed exclusively, with more than 20 million dollars paid out to content owners to date. The catalog spans video, audio, text, gaming, robotics, and enterprise workflow data, browsable in Lens.
For how this market works end to end, see our guides to AI training data marketplaces and where AI labs source training data, or talk to us about licensing.