Where AI training data comes from.

German legal units with 250+ (and 50-249) dependent employees

In reporting year 2024, Germany had 18,632 legal units with 250 or more dependent employees and 84,701 with 50 to under 250, out of 3,543,865 in total.

EU enterprises with 50-249 and 250+ persons employed

In 2023, the EU business economy had 53,000 large enterprises (more than 249 persons employed) and 246,000 medium-sized enterprises (50-249), per preliminary SBS data.

micro1 (buyer)

micro1 committed to spend $1B over the next 12 months to buy and license enterprise operational data through its Company Data Partnerships program, financed by Citi and Hercules Capital.

Startup workspaces evaluated for sale (Slack, Drive, GitHub)

Of the startup workspaces SimpleClosure evaluated for sale, 95% use Slack, 87% Google Drive and 82% GitHub.

UK medium (50-249) and large (250+) private businesses

At the start of 2025, the UK had 38,435 medium-sized private sector businesses (50 to 249 employees) and 8,335 large ones (250 or more).

US businesses filing for bankruptcy

From January 1 to September 30, 2026, there were 25,467 US commercial bankruptcy filings, up 7% from 23,789 in the same period of 2025 (September 2026 alone: 2,627).

US firms with 100-499 employees

In reference year 2022, the United States had 93,341 employer firms with 100 to 499 employees.

US firms with 20-99 employees

In reference year 2022, the United States had 561,160 employer firms with 20 to 99 employees.

US firms with 500+ employees

In reference year 2022, the United States had 21,041 employer firms with 500 or more employees (2,262 of them with 5,000+ in the detailed-size table).

Companies licensing operational data through Labelbox Prism

Labelbox Prism says it has distributed more than $100 million to over 1,000 companies licensing their data, with about 21 days to the first payment.

Companies licensing their data through Handshake AI

Handshake AI's company-data page says it has paid out $250 million to partners and has more than 1,000 partner companies; it quotes payouts of $100K to $4M a company. The page does not say whether the $250 million includes its expert workers.

Closing startups listing code and workspace data on SimpleClosure AssetHub

SimpleClosure expects to return more than $10 million to startups listing data on AssetHub by the end of 2026, after about $1 million across nearly 100 deals in its test year.

An unnamed wireless-power technology company, offered through Hilco Global

Hilco Global is selling, under NDA, a wireless-power company's engineering archive: about 1.25 million lines of company-authored code (6.0 million with vendor SDKs), 202 firmware projects and a 32.6 GB archive of 53,712 engineering emails from August 2017 to January 2024. No price, bid deadline or court case is stated; the listing does not mention AI buyers.

Spirit Airlines estate (In re Spirit Aviation Holdings, Bankr. S.D.N.Y.)

The Spirit enterprise data on sale to Google derives from about 190 million flights and touches the personal data of about 97 million passengers; after the 5 October 2026 narrowing the structured passenger systems are out and corporate email stays in, de-identified by Tonic.ai.

Spirit Airlines' bankruptcy estate: the Microsoft 365 records in the de-identified data sale

The asset schedule of the Spirit Airlines de-identified data sale marks as included about 100 million emails across 80,000 email accounts, 17,082,644 OneDrive items, 20,577,677 SharePoint items and 500,000,000 Teams items, as the Association of Flight Attendants states in its objection (Dkt. 1489).

Leading AI labs' spending on training data (a forecast)

Former OpenAI researcher Will DePue predicts that by the end of the decade leading AI labs could be spending more than $100 billion a year acquiring and creating the data needed to train future models. It is one researcher's forecast, not a budget.

Software customers' contracts with their vendors (Common Paper standard cloud service agreements)

Across 16,140 signed agreements on Common Paper's standard forms, language prohibiting AI training on customer data appeared in under 1% of cloud service agreements in 2024, 11% in 2025 and 14% in 2026.

Software customers' contracts with their vendors (Common Paper standard cloud service agreements)

Language mentioning AI appeared in 53% of the cloud service agreements signed on Common Paper's standard form in the year ending June 2026, up from 47% in 2025 and 38% in 2024.

Startups that shut down through SimpleClosure, January to June 2026

Non-AI companies that shut down through SimpleClosure in the first half of 2026 had a median of $13,000 in the bank when they stopped; AI companies had a median of $30,000.

App publishers on Google Play and the Apple App Store

As of 16 October 2026, there are 2,635,872 apps on Google Play and 2,718,308 on the Apple App Store, each backed by a private codebase.

Permissively licensed public code (The Stack v2, built from Software Heritage)

The Stack v2 contains over 3 billion files in 600+ languages, totalling 67.53TB uncompressed (about 900B training tokens per the card's table).

COBOL code in production at enterprises

A Micro Focus global survey estimated 775 to 850 billion lines of COBOL in daily use.

Private repositories on GitHub

In 2025, 81.5% of contributions on GitHub (4.97B) happened in private repositories, though 63% of all repositories were public.

Repositories hosted on GitHub

GitHub hosts 630 million repositories. 121 million were added in Octoverse year 2025 (Sep 2024 to Aug 2025), including 58 million new private repositories (up 33%).

Startups selling assets through SimpleClosure AssetHub

The companies that offered assets on AssetHub since its April launch represent nearly 500 million lines of code across 3,000 repositories, with 1.7 million commits and 200,000 pull requests.

Public source code archived by Software Heritage

The Software Heritage archive holds over 27 billion unique source files from 421 million projects.

Agencies and software companies licensing private code through Pangea.ai's partner

Pangea says $600,000 has been paid to code owners across its network for 380 codebases, about $1,579 a codebase.

Early-stage startups' private repositories

Scale AI's SWE-Bench Pro benchmark keeps a commercial set of 18 proprietary repositories from early-stage startups under formal partnership agreements; the paper says Scale purchased them.

University of Liverpool (lecture recordings)

One UK university alone has recorded and shared more than 85,000 hours of teaching through lecture capture since 2013.

Pond5 (stock footage marketplace, Shutterstock-owned)

Pond5 passed 30 million licensable royalty-free video clips, with 100,000 artists uploading 15,000 new videos a day.

Game studios publishing on Steam

SteamDB counted 20,008 games released on Steam in 2025. Other trackers counted 20,282 to 20,558.

Vimeo (business and creator video hosting)

Vimeo says an average of more than 350,000 videos are uploaded to it every day, by 1.7 million or more paid subscribers in 190+ countries.

YouTube (creators, platform)

YouTube says an average of more than 20 million videos are uploaded to it every day.

Video creators and media companies licensing through Troveo

Troveo says it has paid more than $50 million to creators, counts about 7,000 content owners and has licensed more than 7 million hours of video. Its business-data page gives $20 million or more paid to licensors on the same day.

Troveo (creator, studio and archive video licensed for AI training)

Troveo states it has paid more than $50 million to creators, has about 7,000 content owners, and has licensed more than 7 million hours of video; its homepage adds that it delivered more than 1 million hours of footage to one enterprise client, and its resource pages state 95 percent of licensors signed exclusively.

GoPro subscribers (AI Training programme, US cloud subscribers who opted in)

GoPro's AI Training programme, launched in Q3 2025, had received more than 500,000 hours of subscriber video by its Q4 2025 report; subscribers earn 50% of the licence revenue GoPro expects to generate on their behalf, revenue was to be recognised from Q1 2026 and first payouts made later in 2026. No dollar figure has been disclosed (the Q2 2026 10-Q names the programme only).

Veritone Data Refinery (rights holders' archive video and audio: sports federations, broadcasters)

Veritone cites qualified bookings and a near-term pipeline for its Data Refinery valued at $40M at the end of Q3 2025, up 400% over Q1 2025; in one instance it processed 500,000 hours of video to deliver 26,000 hours for model training.

Human Motion Data Recorder app contributors (household-chore egocentric video, mostly the Philippines)

Launched in April 2026 to fill one robotics company's need for 50,000 hours of egocentric video, the app had collected 38,000 hours from 3,900 contributors by mid-summer 2026, paying US$3 per approved hour.

US contact centres (enterprises and BPOs)

The US had 2,666,000 customer service representative jobs in 2025, and BLS projects a 5% decline for 2025 to 2035.

Mozilla Common Voice (volunteer speech)

Common Voice Scripted Speech v27.0 (data cut-off 2026-09) has 42,593 total hours, 29,295 validated hours and 295 languages.

Contact centres on Five9 (4,000+ customers)

Five9, one contact-centre platform, says 14+ billion call minutes are recorded on it each year for its 4,000+ customers.

Podcasters worldwide

Listen Notes counts at least 3,824,033 podcasts and 195,710,248 episodes.

Shaip (speech data vendor)

Shaip's speech catalogue lists 217,795 hours of English speech across 61 datasets, and 55 call-centre datasets.

Spotify (podcast and audiobook catalogue)

Spotify hosts 7 million podcast titles and 700,000 audiobooks.

Off-the-shelf speech datasets licensed by their producer (Haitian Ruisheng, China)

In 2020 Haitian Ruisheng licensed 199,980 hours of its own off-the-shelf speech datasets (198,951 hours of speech recognition data, 1,029 hours of speech synthesis data), up from 136,685 hours in 2019 and 115,732 in 2018; its catalogue held 762 datasets at the end of 2020.

US hospitals (imaging and clinical record holders)

The American Hospital Association counts 6,100 hospitals in the United States, each a holder of imaging archives and clinical records.

Getty Images (Getty, iStock, Unsplash)

Getty Images has more than 662 million visual assets on its sites and distributes the content of more than 600,000 contributors and more than 360 premium partners.

Getty Images archive (historical photography)

Getty Images says it maintains one of the largest privately owned photographic archives in the world, with over 150 million images, separate from its creative and editorial stock.

LAION (web-crawled open dataset)

LAION-5B contains 5.85 billion CLIP-filtered image-text pairs: 2.3 billion in English and 2.2 billion in more than 100 other languages.

NCI Imaging Data Commons (public research archive)

NCI's Imaging Data Commons release V23 holds 95.33 TB across 160,199 imaging studies from 79,889 cases.

Segmed (aggregator of hospital and clinic imaging)

Segmed's imaging platform holds approximately 150 million de-identified imaging exams across CT, MRI, X-ray, ultrasound and mammography, sourced from healthcare sites and licensed to AI developers.

Shutterstock (stock images, footage, music, 3D)

Shutterstock's Data, Distribution, and Services revenue (licensing its image, footage and music library and metadata, including to AI developers) was $56.1 million in Q2 2026 against $67.2 million in Q2 2025, and $77.2 million in the first half of 2026 against $106.9 million.

US hospitals and imaging centres

About 93 million CT examinations were performed on 61.5 million patients in the United States in 2023.

Photographers and creators supplying Wirestock

Wirestock, which supplies multimodal data to six of the largest foundation model makers, has paid $15 million to its contributors on an annual run-rate revenue of $40 million.

Shutterstock's image, video and music library and its contributors

Shutterstock earned $104 million of AI licensing revenue in 2023 across all its AI customers.

Shutterstock (stock images, footage, music, 3D)

Shutterstock's Data, Distribution, and Services revenue (which includes licences to the metadata of its images, footage, music and 3D models for machine learning and generative AI training) was $203.3 million in 2025, up 16% from $175.3 million in 2024 and $137.3 million in 2023.

Wirestock (photos, video, design and 3D assets made to order for AI labs)

Wirestock raised a $23 million Series A in May 2026 after pivoting from stock distribution to AI training data; it sources images, videos, design assets and 3D content from more than 700,000 artists and designers, supplies six of the largest foundation model developers, runs at $40 million in annual revenue and has paid $15 million to contributors since 2023.

Kled (consumer app paying people for their existing photos and files)

Kled's operator, Nitrility Inc., claimed in early 2026 a $100 million valuation and 3 million files uploaded daily by consumers selling their camera rolls for AI training.

arXiv (Cornell, research preprints)

arXiv has received 3,215,245 submissions since August 1991.

Bartz v. Anthropic (N.D. Cal. 3:24-cv-05417)

The Bartz v. Anthropic settlement covers 482,460 works on its Works List. By April 16, 2026, claims had been filed for 440,490 of them, a 91.3% claims rate.

Bowker (via Publishers Weekly)

More than four million books with ISBNs were published in the US in 2025. More than 3.5 million were self-published, and 642,242 came from traditional publishers.

Cloudflare Radar

In 2025, Anthropic's crawlers fetched up to 500,000 pages for every visitor they referred back, settling at about 25,000 to 100,000 per referral after May. OpenAI's ratio peaked at 3,700:1 in March.

Common Crawl Foundation

The September 2026 Common Crawl archive (CC-MAIN-2026-39) holds 2.17 billion web pages (361.4 TiB uncompressed) from 40.4 million hosts. Only 587.2 million of its URLs had never been crawled before.

Data Provenance Initiative

Between 2023 and 2024, robots.txt restrictions cut off about 5% or more of all tokens in C4 and 28% or more of its most actively maintained sources, and 45% of C4 was restricted by terms of service.

Crossref

Crossref holds 189,130,320 registered metadata records, including 126,615,864 journal DOIs.

Epoch AI

Epoch AI estimates the effective stock of public human-generated text at around 300 trillion tokens, and expects models to fully use it between 2026 and 2032 if current trends continue. This is still the June 2024 estimate: Epoch's page shows no 2025 or 2026 update.

Harvard Law School Library / Institutional Data Initiative

Institutional Books 1.0 releases 983,004 public-domain volumes (242 billion tokens). They come from 1,075,899 Harvard Library volumes in over 250 languages that Google Books scanned.

Internet Archive

The Wayback Machine passed one trillion preserved web pages in October 2025.

Newspapers.com (Ancestry)

Newspapers.com, owned by Ancestry, holds 1,264,636,501 digitised newspaper pages from more than 31,500 newspapers dating back to 1690.

PubMed Central (US National Library of Medicine)

PubMed Central had 11,327,616 full-text biomedical articles available in fiscal year 2025, up from 10,307,502 in FY24.

ProRata.ai

ProRata says more than 1,000 publications and content creators have partnered with it to license content to AI answers on a revenue-share basis.

Alibaba Qwen (buyer-side consumption)

Alibaba's Qwen3 models were pretrained on approximately 36 trillion tokens covering 119 languages and dialects, nearly twice the amount used for Qwen2.5.

Reddit, Inc.

Reddit's other revenue, which is mostly content licensing, was $140.0 million in 2025, up from $114.7 million in 2024 and $15.2 million in 2023. In Q2 2026 it was $43.3 million.

RSL Collective / RSL Technical Steering Committee

More than 1,500 media organizations, brands and technology companies endorsed the RSL machine-readable AI licensing standard when version 1.0 launched.

Stack Overflow (new public Q&A)

Stack Overflow's own API counts 986 questions asked in September 2026 and still on the site, against 24,732 for September 2024 and 119,730 for September 2021 on the same query.

Stack Overflow (Prosus)

Stack Overflow's public API reports 24,134,193 questions on the site in total.

TAUS Data Association (pooled translation memories)

TAUS put its pooled collection of about 7.4 billion words of non-public human translation data in 483 language pairs on sale for LLM training at discounts of more than 95% in March and April 2024.

Wiley's research and learning content

Wiley's AI licensing revenue was $14 million in the quarter to 31 July 2026, against $29 million in the same quarter a year earlier.

Taylor & Francis (Informa) archive content licensed to AI companies

Informa reported that its AI partnerships generated more than $75 million of non-recurring data access revenue for Taylor & Francis in 2024; its 2025 report says this was not expected to repeat to the same degree in 2025.

Wiley research and learning content licensed for AI (including content Wiley licenses from other publishers)

Wiley's AI license revenue was $13.7 million in the three months ended 31 July 2026, against $28.9 million a year earlier; AI license revenue in its Learning (books) segment was zero, against $13.1 million.

Character Technologies (Character.AI) estate, Chapter 11, D. Del. Case 26-11603

Character.AI filed for Chapter 11 on 12 October 2026 to sell all or substantially all assets; Disney's $20M stalking-horse bid is a non-exclusive licence to certain IP and claims and, per the reporting, excludes the app, brand, user data and conversation logs, which stay in the estate. A Kentucky complaint attached to a 40-state AG filing (Dkt. 33) counts 21.6 million US accounts by March 2025, 2.1 million declaring an age under 18. Bid deadline 20 November, auction 23 November, sale hearing 24 November 2026. The AGs ask that no buyer may train AI on chat content beyond users' consent, so the logs are unlikely to be saleable as training data.

Bloomsbury Publishing plc

Bloomsbury says the court listed 14,087 Bloomsbury titles in the Bartz v. Anthropic settlement, at about $3,000 a work before fees, divided equally between author and publisher, paid over several instalments from the second half of its financial year.

Cambridge University Press & Assessment

Cambridge University Press has about 45,000 books available in print and digitally, the pool it is asking authors to opt in to AI licensing; most agreements run a 2 to 5 year training term.

Arcadia Publishing (via Slate)

Arcadia Publishing, the US local-history publisher (History Press, Belt, Pelican imprints), has a catalogue of nearly 17,000 titles and sublicensed some of them to an unnamed technology company for AI training in two rounds in 2025.

AWS Data Exchange

AWS Data Exchange lists more than 3,500 data products from over 300 providers, and over 1,000 of them are free.

Publications Office of the EU (data.europa.eu)

The European data portal lists 1,751,247 datasets from 212 catalogues in 36 countries.

US General Services Administration (Data.gov)

Data.gov lists 586,902 datasets.

Datarade

Datarade lists data products from more than 2,700 data providers across more than 560 data categories.

Defined.ai

Defined.ai's dataset marketplace lists 819 off-the-shelf datasets. Its homepage cites a contributor crowd of 1.6M+ across 150+ countries.

Hugging Face

The Hugging Face Hub lists 1,092,755 datasets, almost all free to download.

Hugging Face Hub (datasets tagged rl-environment)

Only 148 of the Hub's roughly 1.09 million datasets carry the rl-environment tag, so finished training environments are almost absent from the free supply.

Snowflake Marketplace

Snowflake Marketplace lists more than 3,400 live data, agent and SaaS offerings from over 820 providers.

Zenodo (CERN open research repository, university researchers)

Zenodo's public API reports 7,458,033 records, of which 725,627 are typed as datasets and 306,429 as software.

Data owners selling datasets through Defined.ai

Defined.ai says its data partners earn more than $600,000 on average, more than five have earned over $1 million, and its dataset sales grew 319% in a year.

Data holders brokered by Protege (healthcare, media and enterprise records)

Protege's CEO says its total transaction volume rose from $30 million last year to at least $100 million this year.

Owners of large specialised content collections (code repositories and others)

Google says it has tested its Content Offer Pilot for over a year with thousands of partners worldwide, setting each fee in a one-to-one offer process; the deal is non-exclusive and the partner keeps its IP.

Website replicas (UI gyms) bought from environment vendors

SemiAnalysis reports that UI gyms (website replicas for agent training) often cost about $20,000 per website and that OpenAI has bought hundreds of them for ChatGPT Agent.

Open-source RL environments built through Prime Intellect bounties

Prime Intellect crowdsourced more than 400 RL environments and evals through bounties in two months, with 80+ reviewed implementations.

RL environments sold to AI labs

An RL environment is often a one-time delivery with a half-life of about six months, and the same customer rarely buys the same environment twice, according to an investor's analysis of the AI data category.

AgiBot / OpenDriveLab

AgiBot World Beta has 1,003,672 robot trajectories (about 43.8T of data) under a non-commercial CC BY-NC-SA 4.0 license.

DROID consortium (13 institutions)

DROID holds 76,000 teleoperated demonstration trajectories, or 350 hours of robot interaction data.

Ego4D consortium (Meta and universities)

Ego4D holds 3,670 hours of first-person daily-life video from 923 participants in 74 locations across 9 countries.

Factory owners worldwide (counted by IFR)

Factories worldwide installed more than 600,000 new industrial robots in 2025.

Factory owners worldwide (counted by IFR)

About 5 million industrial robots were in operation worldwide at the end of 2025, up 9% on the year.

Meta (buyer)

Meta pretrained V-JEPA 2 on over 1 million hours of internet video, then post-trained its robot variant on under 62 hours of DROID robot video.

NVIDIA (buyer)

NVIDIA pretrained its Cosmos world models on 9,000 trillion tokens, including 20 million hours of driving, robotics, synthetic and related data.

Open X-Embodiment collaboration (Google DeepMind and research labs)

Open X-Embodiment pools more than 1 million real robot trajectories from 22 robot embodiments, collected by 21 institutions.

Tesla (customer fleet)

Tesla said owners had driven over 10 billion miles on FSD (Supervised) as of early May 2026.

Waymo

Waymo had driven 271.3 million rider-only (no human driver) miles through June 2026.

People filming everyday tasks through Figure AI's Index app

Figure AI's Index app shows $35 million paid to contributors, 28.5 million video uploads, 1 million app downloads and contributors in 111 countries. Figure said $15 million on 25 August 2026.

Embodiment-free human manipulation data (UMI, egocentric) produced in China

Chinese embodied-data service company Mifeng Technology (觅蜂科技) says it has produced over 1 million hours of embodiment-free data; Noematrix (穹彻智能) collected over 100,000 hours of UMI data and about 10,000 hours of egocentric data in about six months.

Real-robot teleoperation collection in China

A year before September 2026, eight hours of robot work typically yielded one hour of valid data, per industry bodies quoted by Xinhua Finance.

Studio-grade human motion capture (Uthana)

Uthana licenses 250+ hours of human motion captured with a 120-camera marker-based system in a AAA studio, for commercial model training under agreement.

High-precision motion capture for humanoids (Noitom HiPHI)

Noitom released HiPHI, 617.5 hours of motion capture from 132 performers, free for non-commercial research with enterprise access through ModalityNet.

US Bureau of Labor Statistics

There were 862,800 physician and surgeon jobs in the US in 2025, a credentialed pool expert-data platforms recruit from.

Handshake AI

Handshake says its AI business reached about a $1B run rate and pays about $60M a month to over 30,000 people.

Mercor

Mercor says it has nearly five million experts on its platform.

Outlier (Scale AI)

Outlier says it has 900K+ AI trainers (master's holders, PhDs and graduates) and has paid experts more than $500M.

Upwork

In Q2 2026, Upwork's gross services volume from AI-related work grew more than 22% year over year, while total gross services volume fell 4% to $966.4 million.

The six largest Chinese AI labs, buying from American data-labelling companies

The six largest Chinese AI labs spend about $500 million a year with American data-labelling companies, according to two founders in the sector who received market-size estimates from Tencent and ByteDance executives (Forbes).