When and How AI Data Buyers Pay: Upfront, Milestones, Revenue Share and Clawbacks

AI data buyers pay when a contract event occurs, usually delivery of the data or the buyer's acceptance of it, and rarely at signing. Mercor states a wire on delivery, micro1 invoices only after acceptance, and Labelbox Prism states an upfront licensing fee at signing (all checked November 9, 2026). Every timing figure here is the buyer's or intermediary's own claim.

Payment terms are one part of AI training data licensing, and not selling remains a valid result.

DataPacts Editorial Team checked every figure here against its primary source on November 9, 2026; the page on how we source, date and check every figure sets out the rules.

What Triggers Payment in an AI Data Deal

Five events trigger payment in published AI data terms: signing, delivery, acceptance, the buyer's own receipt of money, and each later sale or use. The table compares each stated trigger, as of 2026-11-09, grouped by trigger and alphabetical within each group.

Court-supervised sales add a judge's order. The Spirit Airlines bill of sale (Bankr. S.D.N.Y. 25-11897 (SHL), Dkt. 1463, 2026-08-14) provides payment contemporaneous with delivery, conditional on a sale order not entered as of November 9, 2026 (hearing November 17, 2026, Dkt. 1779), so sales of data assets in insolvency pay only after that order.

Four Ways AI Data Buyers Structure the Payment

Payment for company data comes as a single fixed fee, as installments or milestone payments, as a revenue share, or as an upfront fee combined with a revenue share.

Upfront Fixed Fee Paid Once

A fixed fee is one agreed sum, and in published programs "upfront" usually means paid once on delivery or acceptance, not before the data moves. Endless, an aggregator and intermediary, states a fixed payment agreed before data leaves, and Collinear AI states payment according to the contract after delivery is confirmed (both checked November 9, 2026).

Installments and Milestone Payments Over a Term

Larger licenses split the price into an upfront payment and later installments or per-delivery payments, and public-company filings show the schedules. Tempus AI's 10-K for fiscal 2025 (filed 2026-02-24) records USD 200 million of data license fees over three years from Pathos, including USD 50 million paid upfront as of April 2025. Its third-quarter 2026 10-Q records USD 100 per delivered record under a 2026-10-22 order form, and USD 137.5 million of fixed payments to 2031-12-31 under a 2026-10-14 statement of work that BMS may end on 12 months' notice.

The filings are summarized under Tempus AI clinical data licensing deals. Data Licensing Advisors, an intermediary, states a success fee and payment on buyer milestones (checked November 9, 2026).

Revenue Share on Each Sale or Use

Under a revenue share the seller is paid a percentage each time the buyer or intermediary sells or uses the data, so income depends on later sales the seller does not control. Four programs state this model (checked November 9, 2026):

• Gradient Health: revenue share each time data is used; its own example is USD 3 million x 40 % contribution x a "typical" 20 % share = USD 240,000.
• Protege: payouts "with each use", percentage not public.
• SuperAnnotate: payment on delivery and on each relicense.
• Turing Project Lazarus: "Nothing upfront", paid every time the record sells, split not published.

Apart from Gradient Health's "typical" 20 % in its own example, none of these programs publishes a split. The choice between models belongs to AI data monetization models.

Upfront Fee Plus Revenue Share

Some buyers combine a fee at signing or delivery with a share of later income, which pays the seller early and keeps a stake in reuse. Labelbox Prism states "an upfront licensing fee at signing plus a revenue share on downstream use", and Replay states offers are "typically structured as an upfront cash payment plus a perpetual revenue share" (both checked November 9, 2026).

Payment Timing Claims: Days From Signing, Delivery or Acceptance

Published payment timing runs from under three weeks to about three months, and each figure is the buyer's or intermediary's own claim, checked November 9, 2026. The table gives each claim and the event it is counted from, as of 2026-11-09.

No seller statement or filing confirms when any program actually paid. Every program's AI data partnership payouts and program terms compared is dated on the terms tracker.

Where Buyers' Own Pages Disagree on When They Pay

Two buyers state two different payment points on the same page, and the difference decides whether the seller waits one month or two. micro1 says "We pay in full within 30 days of acceptance" and, in its FAQ, that it has up to 30 days to review and approve, with payment within 30 days of invoicing (checked November 9, 2026).

Mercor states payment "typically within 2-4 weeks of a signed agreement" and also "as soon as processing is complete" (checked November 9, 2026). The first counts from signing and the second from the end of processing. DataPacts shows both statements and does not choose between them.

Acceptance as the Gate Before Payment

Where payment follows acceptance, the buyer's review period decides when money is owed, and silence can count as acceptance. micro1's Data License Master Terms (updated 2026-11-08) set four mechanics:

• Rejection window: 30 business days to reject (s.1.3).
• Deemed acceptance: delivery not rejected in that window is deemed accepted (s.1.3).
• Invoice after acceptance: invoices follow acceptance, and the License Price in the Order Form is the sole consideration (s.3).
• License after payment: the grant attaches after acceptance and payment.

Patronus AI states up to 30 days of review (checked November 9, 2026). The acceptance and payment clauses of an AI data licensing agreement are explained on the contract page.

Clawbacks, Refunds and Holdbacks After a Buyer Pays

A clawback lets the buyer recover money already paid, and in published seller terms it is tied to the accuracy of what the seller described, not to how the buyer's model performs. The table compares the published terms, as of 2026-11-09.

Three patterns follow for sellers:

• Inaccurate manifest: micro1 states a refund if the Data Manifest is materially inaccurate, and the right survives acceptance.
• Latent defects: micro1 allows claims within 60 days after acceptance.
• Revocation without refund: Labelbox states deletion within 30 days under its kill clause, trained models "cannot be unlearned", and no refund term is stated.

Holdbacks appear on the introducer side, and FileYield states that parties handle payment directly.

How the Money Reaches the Seller: Wires, Processors and Deductions

Company data payments arrive by bank wire or through a payment processor, and the price can shrink by an intermediary's fee or by tax withheld at source. Mercor states a wire with no setup fees or deductions, and Centific states no deductions from payment (checked November 9, 2026). SimpleClosure and Data Licensing Advisors, both intermediaries, state a percentage or success fee, and DataVendor keeps 20 % of vendor earnings when withdrawn.

Marketplaces add a processor. Snowflake states that Stripe pays providers after receiving payment from consumers, and Pangea.ai pays within 15 days of the buyer's payment, so the seller waits for the buyer's customer. Stripe Connect cross-border payouts are limited to the US, UK, EEA, Canada and Switzerland. Fees and payout rules of AI training data marketplaces are compared separately.

Benefits, Risks, Downsides and Alternatives of Each Payment Structure

No payment structure is safe by default: a fixed fee caps the upside, a revenue share defers it, and both depend on terms signed before any money moves.

Benefits

• A fixed fee gives a known sum, as Endless describes.
• An upfront fee plus revenue share pays early and keeps a stake.
• Filed schedules show fixed payments over years, such as the BMS statement of work.

Risks

• A revenue share can pay nothing: Turing states "Nothing upfront" and Pangea.ai states "licensing not guaranteed".
• Fixed payments can end early: BMS may terminate on 12 months' notice.
• Pass-through timing depends on the buyer's customer paying first.

Downsides

• The grant is broad: micro1's terms are irrevocable and perpetual.
• Trained models cannot be unlearned, Labelbox states, so deletion does not undo training.
• micro1 states no duty to delete or retrain Licensee Materials (s.4.4).

Alternatives

• A fixed fee paid on delivery with a short acceptance window.
• A revenue share only.
• A time-limited license, as SuperAnnotate describes.
• Not selling or licensing at all.

The prior decision is whether to sell at all, which the page asking should you sell or license your data to AI companies treats first.

Payment Questions to Settle Before Signing

Six payment questions decide what a seller actually receives and when, and each should have a written answer in the order form.

1. Which event triggers payment?
2. How long is the acceptance window, and does silence count as acceptance?
3. How many days pass between the trigger and the money?
4. Is the price fixed, milestone-based or a revenue share, and on what base?
5. On what grounds, and for how long, can the buyer claw money back?
6. Who bears costs and fees, including tax withheld at source?

The answers feed into evaluating an offer for your data against other offers.

Where Payment Sits in the Process to Sell or License Company Data

Payment comes after the contract is signed and the data is delivered and accepted, and before tax treatment and the obligations that continue after the sale. The full sequence to sell or license company data to AI companies runs as follows:

1. Selling direct or through a broker
2. Negotiating an AI data licensing deal
3. Term sheets and order forms
4. Delivering licensed data to an AI buyer
5. Payment (this page)
6. Tax on data licensing income
7. Accounting for the license revenue
8. After the sale

The elapsed time is covered in how long an AI data licensing deal takes, and alternatives to selling data to AI labs covers other outcomes.

Payment Terms in Each Buyer Program

The payment terms above come from these programs' own pages, each checked November 9, 2026. The overview of company data partnership programs and the profiles of Mercor and micro1 hold more.

• Buyer programs: Centific, Collinear AI, Gradient Health, Labelbox Prism, Mercor, micro1, Mode, Patronus AI, Polyshares, Replay, SuperAnnotate, Turing Project Lazarus
• Marketplaces: DataVendor, Pangea.ai, Protege, Snowflake Marketplace
• Intermediaries (they sell data for companies and do not buy it): Data Licensing Advisors, Endless, Marin Data, SimpleClosure

How Introducers and Referrers Get Paid

Introducers are paid under their own referral terms, usually only after the referred company's deal closes or the buyer receives its money, and those fees can be clawed back. Handshake AI states its bonus is paid when the deal closes and Replay pays the day it pays the upfront fee (checked November 9, 2026); the referral programs of AI data buyers set their own triggers.

Tax and Accounting Treatment of Data License Payments

A US buyer withholds 30 % of license payments to a foreign corporate seller under 26 U.S.C. 1442(a) unless a treaty reduces the rate. The rules are set out in withholding tax on data license payments. Reddit's FY2025 10-K recognizes license revenue ratably, whereas Getty Images' FY2025 10-K, discussing 2024, says data licensing typically results in a greater portion recognized up front. The receipts are covered under accounting for data licensing revenue.

Frequently Asked Questions

When do AI data buyers pay?

Most AI data buyers pay after delivery or acceptance of the data, typically within weeks to about three months by their own statements. Mercor states 2 to 4 weeks, micro1 30 days and Mode about three months (checked November 9, 2026). Mercor's own feed shows two paid-to-companies totals, USD 17,000,000 and USD 50,000,000 (checked November 9, 2026); see reported payouts to data sellers.

Do AI data buyers pay upfront?

Few do: among published terms only Labelbox states a fee at signing, and most fixed fees are paid on delivery or acceptance. Tempus AI's filed deal with Pathos includes USD 50 million paid upfront as of April 2025. Mercor and Centific state no setup fees, so a request for money from the seller deserves a check on whether a data licensing email is legitimate.

What is a revenue share in AI data licensing?

A revenue share pays the data holder a percentage of what the buyer or intermediary earns each time the data is sold or used. Turing and Protege state this model without publishing the percentage, and Gradient Health's own example uses a "typical" 20 % share (checked November 9, 2026).