21 programmes and channels publish terms for data used in AI training. For company records, Replay, Handshake AI, micro1, Polyshares, Turing, Grepped, Mercor, Troveo, Pewter and SimpleClosure publish the most: ranges from $10K to $5M, one-off or revenue-share payment, and entry bars from none to 30 employees.
Page updated 19 Oct 2026. Figures checked 18 Oct 2026 at their sources.
Company-data programmes with thinner public detail
These take operational records too, but publish too little for a profile. Each card quotes what is on record and says where nothing is.
Ooak Data
Company data
What they buy
Company tool data (Gmail, Slack, Notion, Jira, Drive, SharePoint), anonymised into a "digital twin" and turned into RL environments.
Published price
Not published.
How they pay
Not published.
Referral fee
None published.
Notes
YC Summer 2026; founders Pierre-Louis Vouteau (CEO), Thomas Aubry (CTO), Grégoire. "working with multiple frontier labs"; "well above 7 figures in revenue"; data from 20 companies so far. Its own pages do not state that or how it pays data-providing companies (only a dataset-worth "simulator" is mentioned).
Data of startups being wound down (code, workspace data) resold to AI buyers.
Published price
Not published.
How they pay
Not published.
Referral fee
None published.
Notes
Startup wind-down firm; MIT Technology Review says its data resale business is "rapidly growing" and it values assets by headcount, years in operation, data volume and tool use. Own site (sunsethq.com) blocked our fetch, so fees and figures unconfirmed. MIT Technology Review describes a "rapidly growing data resale business"; CEO Brendan Mahony values assets by head count, years in operation, data volume and how the team used its tools. Own site (sunsethq.com) blocked fetch; no terms verified.
Agent interaction data from companies building AI agents: "traces, corrections, completed workflows". Intake asks where agent data lives (e.g. Langfuse, Postgres, S3) and monthly trace volume.
Published price
Not published.
How they pay
"A revenue share, paid as your data licenses." Share % not stated; "Terms are in the data agreement we sign during onboarding."
Referral fee
None published.
Notes
Backed by Y Combinator (badge on site). Buyers: "AI labs and researchers". Narrow niche (AI-agent companies only).
Internal company data, "starting with manufacturing and industrial companies"; codebases, project files and payments records incl. from bankruptcy courts; says it has transacted construction plans, autonomous-driving recordings and codebases. Targets legacy businesses, startups and companies in bankruptcy or pivots.
Published price
Not published.
How they pay
Not published.
Referral fee
None published.
Notes
YC company. Launch post: "sources suppliers, cleans and restructures the datasets, certifies legal clearance, and resells" to AI startups and labs. Own homepage (petrarch.co) now frames it as helping operators prepare records "cleared and anonymized on premise" and does not state it buys. Contact founders@petrarch.co.
Mostly media, books, creator content or dataset listings. They matter here as reference points for commission levels and licence structures.
Defined.ai
Marketplace / media
What they buy
Text, audio, image, video or multimodal datasets, incl. "hard-to-source datasets"; partners must hold "the right to commercialize the data for AI training". No fixed minimum size. High-demand: cinematic video, anime, coding, STEM.
Published price
Commercial terms "depend on the dataset type, usage, and customer needs" and are "always aligned with and approved by the partner". Press release (27 Jan 2026): "Several partners now generate more than $1 million annually".
How they pay
"Licensing fees from data marketplace sales". Split % not stated.
Referral fee
None published.
Notes
Marketplace model (partner lists datasets, Defined.ai sells to AI developers). 1,200% increase in third-party partner datasets in 2025 per the release. Partners can set usage limits.
Domains: Healthcare, Video, Audio & Speech, Spatial & Physical, Other Domains. Series A post (13 Aug 2025): "over 100 data partners across healthcare and media"; over 300,000 hours of video, over 500,000 hours of audio, billions of clinical notes, hundreds of millions of medical images.
Published price
Not published.
How they pay
Not published.
Referral fee
None published.
Notes
Site pitches "Revenue for Data Holders" / "Monetize your existing data assets". Says it has generated "tens of millions in revenue for its data partners". Payment structure not published (revenue-share claims seen only in secondary commentary, dropped).
Books from authors, agents and publishers: "training, reference, and transformative rights", titles "fully rights cleared".
Published price
Not published.
How they pay
Negotiates agreements for rights-holder approval; payments tracked in a dashboard. TechCrunch: Created by Humans "would take a cut of the deal" (no % disclosed).
Referral fee
None published.
Notes
Rights-holder (author) licensing, not company operational data. Buyers: AI developers ("One agreement, one invoice").
Audiovisual content (movies, TV series, news); catalog of more than 35,000 hours of TV and film; "License to Scrape" collective licence for creators' YouTube and other platform videos.
Published price
Not published.
How they pay
Creators under the collective licence "compensated according to a formula based on volume and subscriber count". No split % published.
Exclusivity
"AI model training licenses are non-exclusive and do not include exhibition rights."
Referral fee
None published.
Notes
Rights holders are film/TV owners and creators, not company operational data.
Marketplace listing for data providers; FAQ lists "AI training datasets" among publishable data types. Does not buy data itself: "Your data stays with you."
Headline ranges overlap so much that they rarely separate one programme from another. The structure of the deal does. Four points show the real differences:
Counterparty. Polyshares holds the licence itself; Grepped lets labs buy listings; Pewter places records with a licensing firm it does not name. The party that ends up with the copy is not always the party you talk to.
Payment shape. A single payment (Turing), an upfront sum plus a perpetual share (Replay), or a payment per resale (Troveo). Shares depend entirely on how often the data licenses again.
Cleaning and review. Mercor covers anonymisation itself, Polyshares uses a PII partner and micro1 lets companies review samples first. Whether a seller sees the cleaned output before transfer varies.
Who pays the middle. Pewter is paid by the licensing firm, SimpleClosure takes a success percentage, Datarade charges providers 15% to 30%. Polyshares and Troveo say they deduct nothing.
01Which programmes pay companies for data used in AI training?+
Programmes with public terms include Handshake AI, micro1, Replay, Polyshares, Turing, Grepped, Mercor, Troveo, Pewter and SimpleClosure. Most do not name the labs they supply. The directory above lists every one we track, with sources.
02Do frontier labs ever buy directly?+
Rarely in public. The best-documented case is Google paying $10 million for Spirit Airlines' data in a bankruptcy sale. Most company records reach labs through a programme that collects, cleans and licenses them.
03What separates a buyer, a marketplace and a broker?+
Polyshares and Turing buy or license for themselves. Grepped runs a marketplace where labs buy listings. Pewter brokers to a licensing firm that pays its fee. Replay says it is neither a marketplace nor a broker.
04Which programmes publish referral fees?+
Grepped (10% of the deal, up to $100K), micro1 ($50K, or 7.5% above a $700K deal) and Turing (up to $10k). Mercor pays a bonus once the referred company goes live but gives no amount.
05Is there a minimum company size?+
Several programmes start at 20 employees (Handshake AI, Replay's bands, Pewter, Troveo's estimator) and micro1 at 30. Polyshares, Grepped and Turing publish no minimum.
06Why do some programmes have a profile and others not?+
A profile needs at least 4 sourced facts from the programme's own pages or reliable reporting. Programmes with less stay in the directory so the site has no thin pages.
07Do programmes pay once or over time?+
Both models exist. Turing pays once; Replay pays upfront plus a perpetual revenue share; Troveo pays on every resale; Magma pays a revenue share as data licenses; Polyshares pays Net 30 to 60 after approval.
Stay current
Get the monthly price update
Once a month: deals that disclosed a value, programmes that changed their published figures, and corrections to this site. Free, and one click to leave.
Ooak Data