What AI labs are buying
What AI labs are buying right now, and whether your company has it.
Buyers describe what they want as the systems a company already runs on: its chat, email, tickets, code, customer records and books. Here is what one lab buyer went for on 12 Oct 2026, what makes records worth more, what buyers turn down, and which of your records fit.
5 lots one lab buyer went for on 12 Oct 2026
“A lab buyer just bought 5 corpuses a few minutes ago and wants to close in under a week. Here's exactly what they went for:”
The message ended: “Labs want real company archives with years...” and was cut off there. No prices were given.
Source: A buyer-side referral network's announcement to its members, 12 October 2026 (reported to us).
VC-backed software startup
Company archive
PeopleAbout 250 at peakHistory3+ yearsSizeAbout 3.7 TBWhat was inside- Email, Google Drive and Slack, 3+ years
- 65 private repositories, 6.4M lines of code
- 13.7K tickets
- About 300M product and telemetry records
Creator marketplace
Company archive
People55History5 years (2021 to 2026)SizeAbout 2.1 TBWhat was inside- 96 repositories
- 313K CRM contacts
- About 83K sales-call recordings
- 57K support tickets
- About 200K Drive files
- 120K+ Slack messages
- 58M analytics events
Builder of engineering training environments
Training environment
PeopleNot statedHistoryNot statedSize12K+ containerised tasksWhat was inside- 3,400 CAD and PCB design tasks
- 9,400 chip-design tasks
- 1,500 Lean proof tasks
- All graded automatically
Industrial parts manufacturer
Company archive
People10History20+ years in businessSizeAbout 500 GBWhat was inside- 11 years of CRM, accounting, inventory and orders
- CAD drawings of client orders
- Contract email
Builder of a scientific-research environment
Training environment
PeopleNot statedHistoryNot statedSize400+ evaluation tasks across 7 tracksWhat was inside- Agents write hypotheses, experiment designs and NIH-style grant sections
- Agents respond to peer review
- Scored on rubrics
What we read from it
- Whole archives sell, not one system at a time: each company archive among the lots held several systems together, from code and tickets to CRM, accounting and CAD.
- Records few holders think of as data were inside: about 300M product and telemetry records, and 58M analytics events, about 83K sales-call recordings and 313K CRM contacts.
- A small staff did not rule a company out: a 10-person manufacturer with 20+ years in business was one of them.
- Builders sell ready-made environments too: 12K+ automatically graded engineering tasks, and a science environment scored on rubrics.
- The buyer wanted to close in under a week. A holder whose exports, rights file and de-identification plan are ready can meet that; one starting from nothing cannot.
What makes records worth more
Six factors recur when buyers say what they pay for. The share is how many of the 156 kinds of data in our catalogue each one moves as a main factor: our reading of what buyers state, not a measured weight.
- Outcomes attachedWhat it meansThe record shows how it ended: resolved, paid, won, failed, settled.Share of 156 kinds
47%
74 of 156
What a buyer said“Salesforce or HubSpot CRM, service histories, support tickets, and the record of how requests and issues got resolved.”
NyneIts own page, 12 Oct 2026 - LinkageWhat it meansRecords connect across systems, so a buyer can follow one piece of work from request to result.Share of 156 kinds
46%
71 of 156
What a buyer said“Connected data is worth more than big data.”
Ooak DataIts own page, 12 Oct 2026 - SpecialisationWhat it meansWork in a field that is hard to replicate with hired contractors.Share of 156 kinds
39%
61 of 156
What a buyer said“Specialized expertise that makes your data hard to replicate.”
Scale AIIts own page, 12 Oct 2026 - HistoryWhat it meansYears of continuous records in the same systems.Share of 156 kinds
33%
52 of 156
What a buyer said“Value scales with your systems, team size, and years of history.”
TroveoIts own page, 12 Oct 2026 - ScarcityWhat it meansData that exists nowhere on the open web or in any catalogue.Share of 156 kinds
32%
50 of 156
What a buyer said“What you receive scales with how complete, rare, and well-preserved the record is.”
Turing (Project Lazarus)Its own page, 12 Oct 2026 - Reasoning visibleWhat it meansThe record shows why, not only what: notes, reviews, trade-offs, corrections.Share of 156 kinds
24%
37 of 156
What a buyer said“The reasoning matters more than the finished output.”
SuperAnnotateIts own page, 12 Oct 2026
Size alone is the factor buyers warn against
“More records do not automatically mean a higher offer.”
What buyers say they do not want
In the buyers' own words, or in reporting on what they bought and left out. Most of it is not worthless: it is cheap, already everywhere, or blocked by someone else's rights.
- Commodity web text and anything already scrapedWhere it standsCommoditisedWhat was said
“commodity web text is cheap”
Troveo (licensing guide)Its own account, 23 Jul 2026 - Code as a snapshot with no history, and ordinary web appsWhere it standsCommoditisedWhat was said
“A snapshot of source code with no history is worth little to a lab, because there's already a lot of code in the world.”
Troveo (code guide)Its own account, 3 Sep 2026 - A plain mailbox archive with nothing linking itWhere it standsWorth less than holders expectWhat was said
“'ten years of emails' ... 'less than you'd like'”
Grepped (its founder's article)Its own account, 9 Oct 2026 - Volume for its own sake, and duplicate media priced by the terabyteWhere it standsNot wantedWhat was said
“roughly four thousand copies of Frozen 2”
Grepped (its founder's article)Its own account, 9 Oct 2026 - Demos, staged work and synthetic data offered as company recordsWhere it standsNot wantedWhat was said
“not demos or synthetic data”
Scale AIIts own page, 12 Oct 2026 - Short fragments with no thread or intent behind themWhere it standsNot wantedWhat was said
“data that expresses human intention (e.g., long-form writing or conversations rather than disconnected snippets), across any language, topic, and format”
OpenAI (Data Partnerships call)As reported, 9 Nov 2023 - Maths problems and static benchmark-style tasksWhere it standsCoolingWhat was said
“math tasks are easy to create but don't transfer as well to other capabilities.”
Epoch AI interviewsAs reported, 12 Jan 2026 - Quickly built replica websites sold as environmentsWhere it standsNot wantedWhat was said
“what everyone does is vibe code a buggy website which isn't useful.”
Epoch AI interviewsAs reported, 12 Jan 2026 - Consumers' personal data inside a company archiveWhere it standsLeft out for rights, not for lack of interestWhat was said
“does not contemplate the transfer of the personal consumer data or e-mails of Spirit consumers”
Spirit Airlines estate sale (docket summary)As reported, 8 Sep 2026 - Media for copying a person's name, image, likeness or voiceWhere it standsExcluded by the buyer's termsWhat was said
“We do not permit Name-Image-Likeness (NIL) duplication / copying / use from media companies, including voice.”
ProtegeIts own page, 12 Oct 2026 - Raw system access, credentials and identitiesWhere it standsNot wantedWhat was said
“never the raw systems themselves, only structured, de-identified records.”
Scale AIIts own page, 12 Oct 2026
Which of these is your company?
What buyers want, what raises and lowers the value, and the rights to settle first differ by the kind of company. Each page below is written for one kind.
Software companies and startups
SaaS companies, startups and agencies that have shipped software for a few years: the founders, engineering leads and finance leads who decide what the company can license.
Typically holds: code with its full history, tickets and the work behind them, team chat, email
Marketplaces and consumer apps
Two-sided marketplaces, creator platforms, e-commerce brands and consumer apps with a few years of operating history: founders, heads of data and the finance lead who signs.
Typically holds: product analytics and events, orders, listings and payouts, customer and partner records, sales and success calls
Industrial, manufacturing and engineering firms
Manufacturers, machine shops, machine builders, engineering bureaus and distributors, often small and long-established: owners, managing directors and the people who run the office and the drawing vault.
Typically holds: orders, quotes and inventory, accounting, customer records, CAD models and drawings
Professional services firms
Law firms, accounting and tax practices, consultancies and insurance brokers: partners, managing directors and general counsel deciding whether years of work can be licensed.
Typically holds: contracts with their negotiation history, matter and engagement files, workpapers, underwriting and claims files
Support-heavy businesses and contact centres
Companies with a large help desk, outsourced contact centres, and the operations leaders who run support in software, telecoms, banking, insurance, utilities and travel.
Typically holds: tickets with thread and resolution, live chat and messaging, call recordings, quality scorecards
Contractors, builders and trades
General contractors, home builders, roofers, HVAC, plumbing, electrical and solar companies, and other field-service businesses: owners, operations managers and the people who run the office and the job board.
Typically holds: jobs, dispatch and service history, estimates and bids, won and lost, project records of commercial work, residential project records
Closing startups and estates
Founders winding a company down or selling it, acquirers holding records they will not use, and the assignees, trustees, liquidators and administrators who control an estate.
Typically holds: team chat, email, tickets and projects, code with its history
Builders of task sets and environments
Teams and experts who build training environments, graded task sets and private evaluations: in software, CAD and PCB design, chip design, formal proofs, science and professional work.
Typically holds: graded task sets, engineering design tasks, formal proofs, research and writing tasks scored on rubrics
Measure what you hold, whatever kind of company you are
Questions holders ask first
- Do AI labs buy company data directly?
- Mostly not. None of the AI labs we checked runs a referral or broker programme, and only Google runs a self-serve route for outside sellers: its Content Offer Pilot, which is non-exclusive. Lab money reaches holders through vendors: NVIDIA's own summary of its training content names datasets bought from Mercor, Turing and Scale. The lots on this page were announced by a buyer-side referral network.
- Is a small company's data worth anything?
- Often. Of the 24 buyers on our board that take company records, 5 publish a minimum head count, from 20 to 40 people; the rest publish none. One of the lots on this page came from an industrial parts manufacturer with 10 staff and 20+ years in business: what it had was 11 years of CRM, accounting, inventory and orders.
- Do I have to send my data to get a valuation?
- No. The valuation reads your exports on your own computer and sends only counts: how many records, over which months, how many people, how many link to another system, how many identifiers it found. No names, subjects or message text leave it. A de-identified sample, if you add one, narrows the range.
- Can you value a codebase or a task set on its own?
- Yes. The valuation prices a codebase the way buyers price code: per repository, per reviewed change, per task and as a company's whole code base, from the code prices on record, each times how many the code holds, leaving out any buyer whose own published rules would not take it. Where the middle lands follows what code buyers ask for in their own words: tests in CI, reviewed changes tied to their tickets, code written by people, never public. The valuation prices a task set from what the market's records say one task fetches in its field, times how many tasks it holds; where the middle lands follows what task buyers ask for: graded automatically, running as an environment, never published.
- Who pays you, and when?
- Only the holder pays us, a fee fixed in the mandate, and only if a sale closes. In a sale we run no buyer pays us, so our fee depends on what you are paid, not on which buyer pays it. A buyer pays us only if you choose to apply to it through its own referral programme, and you see that amount before you choose. Several buyers pay whoever introduces a holder. In a sale we run we decline that money, tell you beside each offer what that buyer pays introducers where it publishes it or tells us, and ask it to put that amount into your price instead. If you choose instead to apply to one buyer through its own referral programme, that buyer pays us its published payout as well as our fee, and you see the amount in writing before you choose. We take on sales we expect to reach $26,000 or more; below that we name marketplaces whose published rules may fit.
In this topic
What sells
- Start hereWhich data do AI labs want, and who holds it?Each kind of data the market prices, the businesses that hold it, what labs are asking for now and what they turn away.
- GuideWhat Are RL Environments? Who Builds, Buys and Prices ThemAn RL environment is a sandbox where an AI agent attempts tasks with tools and a verifier scores each try. Who builds them, who buys them and what they cost.
- GuideHow to Sell Videos to AI Companies: Prices, Buyers and RightsAI companies license video by the hour or minute. What buyers pay, which footage sells, and the rights you need in every frame before you sell a library.
- PageWhere data comes fromWho holds each kind of data, how much exists and how it reaches buyers.
- RegisterCompanies closingCompanies closing, sold or gone dark, and the records they leave.