AI data partnership programs pay companies in four main ways, as of October 2026: an upfront payment plus a revenue share (Replay, Labelbox Prism), a fixed sum agreed before data moves (Endless, Turing, Mode), payment on each sale (Troveo, SuperAnnotate, pre.dev), or a price set company by company after a sample (Anthroly). Their published terms also differ on headcount floors, exclusivity and who removes personal data. Below, those terms sit side by side, taken only from each program’s own pages. For a range on your own records first, get a confidential estimate.
Owners usually say they want to “sell” their data. Almost every program licenses a copy: you keep the records and the systems, and the buyer gets defined rights to a de-identified copy. We use “license” below for that reason.
What is an AI data partnership program?
It is a buyer’s published offer to license a copy of a company’s operating records for AI training and evaluation. Mode describes its company data partnership as buying “an agreed copy of company-owned operating records” while the company keeps its originals (Mode, observed Oct 19, 2026). Some programs are AI labs or data suppliers that buy for their own account. Others sit in front of them. Endless, for example, says the data it licenses goes on to AI labs “such as micro1, Mercor and Handshake” (Endless, observed Oct 19, 2026). That matters when you compare offers: two programs may end up selling to the same lab.
Our buyer directory lists every program we track with ten terms each. The market index holds the published price bands.
How do the payment models differ?
As of October 2026, the published programs fall into four groups. Terms are from each program’s own page, observed Oct 19, 2026.
| Model | Program | What it publishes |
|---|---|---|
| Upfront plus revenue share | Replay | Upfront cash payment plus a perpetual revenue share; percentage not published (Replay) |
| Upfront plus revenue share | Labelbox Prism | Upfront fee at signing plus revenue share on later use; about $50K to $1M+ (Prism) |
| Fixed sum | Endless | A fixed sum agreed before anything is shared, paid after transfer (Endless) |
| Fixed sum | Turing | “One payment,” up to $1M, with no recurring commitment (Turing) |
| Fixed sum | Mode | Indicative $100K to $5M; a fixed offer after review of the records (Mode) |
| Paid per sale | Troveo | A share “on every sale, not just the first”; percentage not published (Troveo) |
| Paid per sale | SuperAnnotate | Payment on delivery and on each relicense; indicative $100K to $1M+ (SuperAnnotate) |
| Paid per sale | pre.dev | Code only: $100 per million source tokens and $40 per million git-history tokens per non-exclusive license (Codebase Agreement) |
| Priced per company | Anthroly | No band; “every record is priced individually” after a de-identified sample (Anthroly) |
Upfront plus revenue share
You get cash at signing and a share of later licenses. Replay adds that because a company keeps producing data, “the revenue can recur” (Replay, observed Oct 19, 2026). The share is worth only what you can see and audit. Neither Replay nor Prism publishes the percentage, so ask for it in writing, with statements per license and a right to check them.
Fixed sum
You know the amount before any data moves. Endless says the payment is “fixed before anything leaves your company” and nothing is deducted later (Endless, observed Oct 19, 2026). Turing pays a single payment confirmed after sampling (Turing, observed Oct 19, 2026). You get certainty, and no share if the buyer licenses your records many times.
Paid per sale
You are paid each time the program licenses your data to a buyer. Troveo reports an average of four sales per dataset, a figure from its own page that it does not define further (Troveo, observed Oct 19, 2026). pre.dev’s agreement says each further non-exclusive license is paid separately, that the rates are “not a fixed price, a minimum, or a guarantee,” and that pre.dev can change them on 30 days’ notice (sections 13.1 and 13.2, observed Oct 19, 2026).
Priced per company
Anthroly publishes no band. It says it never prices a record it has not seen, and that depth of history, visible reasoning and scarcity of the field move the number more than volume (Anthroly, observed Oct 19, 2026). Mercor also publishes no figure (Mercor, observed Oct 19, 2026). Our guide to what company data is worth sets out the bands other buyers do publish.
Which programs accept small companies?
Headcount floors range from none to 40 employees, as of October 2026. Several programs set lower floors for professional firms.
| Program | Published floor | Source, observed Oct 19, 2026 |
|---|---|---|
| Mode | 6+ employees for law firms; 10+ for accounting, finance and medical; 30+ for others | Mode |
| SuperAnnotate | None: “any size, any industry, anywhere” | SuperAnnotate |
| Replay | Lowest band is 20 to 50 employees, at $10K to $100K | Replay |
| Handshake AI | 20+ full-time employees and 3+ years in operation | Handshake |
| Nyne AI | 20+ full-time US W-2 employees and 3+ years of work | Nyne |
| Endless | Roughly 30+ employees and several years of history | Endless |
| micro1 | 30+ employees | micro1 |
| Scale AI | 40+ employees preferred, assessed case by case | Scale |
| Anthroly | None published; US-headquartered companies only | Anthroly |
Mode’s lower floors for law and accounting firms come with harder rights questions. Client files may not be yours to license. Our pages for law firms and accounting firms cover what applies.
Which programs are exclusive?
Most programs do not publish an exclusivity position, as of October 2026. Of those that do, they split both ways. Labelbox Prism says its licenses are “exclusive by default” (Prism, observed Oct 19, 2026). Endless says its license is non-exclusive and “you can license the same data to anyone else” (Endless, observed Oct 19, 2026). pre.dev’s agreement is non-exclusive and needs a separate signed confirmation for an exclusive license (section 13.4). SuperAnnotate prices exclusivity into its valuation and lets you approve each buyer and block competitors (SuperAnnotate). Mode says the agreement defines “any exclusivity” (Mode). Replay, Turing, micro1, Handshake AI and Nyne AI said nothing on it on the pages we read.
A revenue share pays more when the data sells to many buyers, and an exclusive license stops that. Our guide to exclusive and non-exclusive licenses goes through the clauses.
Who removes personal data, and where?
Every program here says identifiers are removed before records reach a lab. They differ on who does it and where, as of October 2026 (pages observed Oct 19, 2026):
- Inside your systems, before anything leaves: Scale AI de-identifies in the company’s environment before extraction, and Endless strips identifiers “before anything leaves your company.”
- The program, after export: Replay during export, using a model it says it trained for Slack threads, tickets and commits; Mode in a secure workspace; Labelbox Prism as part of ingestion; Nyne AI on its side after transfer; Mercor through an automated pipeline its engineers do not review by hand.
- A third party: Polyshares uses an unnamed specialist partner before anything moves (Polyshares).
Where the work happens decides who holds your raw data, even briefly. Ask to see a de-identified sample before you sign.
When does the money arrive?
Some programs publish timing, as of October 2026. Polyshares pays Net 30 to 60 once the data is approved, de-identified and delivered (Polyshares). Mercor pays by wire on delivery, typically 2 to 4 weeks after signing (Mercor). Labelbox Prism cites under 21 days from first call to first wire (Prism). pre.dev pays within 15 days of the buyer’s payment clearing (Codebase Agreement, section 13.6). Endless pays after transfer and Anthroly after delivery and approval. All observed Oct 19, 2026.
Questions to ask any program
- Is the payment fixed, a share, or both? If a share, what percentage, and how will we see each license?
- Who is the end buyer, and will you license our records on to others?
- Is the license exclusive, for what use and for how long?
- Who de-identifies, where, and can we review a sample first?
- When are we paid, and what happens if the buyer rejects the data?
Each program’s own answers are only one offer. Comparing several on the same scope is what a seller-side advisor does; our guide on how seller-side advisors work explains who pays whom. Marin Data costs you nothing up front, and nothing at all if no deal closes; Marin Data is paid by the AI lab.
Get a confidential estimate. It takes two minutes and asks for no documents.
Sources
- Replay, Endless, Anthroly, Mode, Labelbox Prism, Turing Project Lazarus, Troveo, SuperAnnotate, all observed Oct 19, 2026
- Handshake AI, Nyne AI, micro1, Scale AI, Polyshares, Mercor, all observed Oct 19, 2026
- pre.dev Codebase Agreement (PDF), sections 13.1 to 13.6, read Oct 19, 2026