How long should a data licence last? Terms, renewals and what perpetual costs

A data licence for AI training should usually run two to three years when the holder is a living company whose records keep growing, and be perpetual only when the records are sold once, as a closed company's are; a perpetual licence should then be priced as the sale it is. That is what the disclosed record shows. Reddit's January 2024 data licences ran terms ranging from two to three years, Meta's licence from News Corp runs at least three years and HarperCollins licensed books to Microsoft for three years. Buyers' own standard terms point the other way: micro1's master terms take a licence that is nonexclusive, irrevocable, perpetual, Kled's is perpetual and irrevocable and sublicensable, and every licence on Polyshares' track-record page is a Perpetual exclusive license to the operational record.

The price of "forever" is visible in the few catalogues that sell the same data both ways. Japan's national language institute licenses its Corpus of Spontaneous Japanese commercially for two years at 500,000 yen and with no end date at 5 million yen: ten two-year terms for the open-ended right. A sound-effects catalogue on a data exchange offers a yearly licence at €20,000 a year and a one-off purchase at €60,000, three years of the annual fee. A holder that grants a perpetual licence for the price of one term has given away every renewal it would otherwise have sold.

This guide is for the founder, CFO or counsel of a company (a holder) weighing the term a buyer has proposed. It explains what "term" covers in a data licence and what survives it, what the 111 licensing deals on our deal ledger and the 186 clauses we have read in buyers' published terms say about length, why a licence that is perpetual, irrevocable and sublicensable at once amounts to a sale, how a living company earns from renewals and refreshes, where short evaluation licences fit, and what to ask for. It draws on the 15 terms in our licence terms register and the 72 levers on our negotiation page, and cites only what those records show.

This guide is research from public sources, not legal advice.

Contents

1. What "term" means in a data licence
2. What disclosed deals show
3. What buyers' published terms ask for
4. Two kinds of holder: a living company and a closed one
5. Perpetual, irrevocable and sublicensable: a sale by another name
6. Renewals and refreshes: the recurring revenue of a living company
7. Evaluation-only and other short licences
8. Worked comparison: one term, renewals or forever
9. What to ask for
10. How the term is decided in a sale we run
11. Checklist before you agree a term
12. Questions people ask
13. What this means for you

1. What "term" means in a data licence

A licence is permission. The holder (the licensor) lets the buyer (the licensee) do specified things with specified data, and keeps everything it did not grant. The term is how long that permission lasts. It sounds like one number, but in a licence for AI training it is really three clocks that can run for different lengths, and most disputes about "how long" are disputes about which clock a draft meant.

The access clock

The first clock is how long the buyer may receive the data and train on it. For a living company that keeps producing records, this is the clock that matters most, because during it the holder may deliver new records (a refresh) and the buyer may train new models on them. Reddit's licence to Google, for example, gives real-time access to Reddit's data through Reddit's own interface, which only means something for as long as the access clock runs. When it stops, new deliveries stop, and the buyer has only what it already holds.

The copy clock

The second clock is how long the buyer may keep the copy it received. A licence can end the right to train while saying nothing about the files on the buyer's servers, which leaves the holder's records sitting in someone else's systems indefinitely. Our register's position is to require deletion of the dataset (not the models, unless agreed) within a fixed number of days, with a written certificate, and deletion on breach. NVIDIA's own data agreement for model training is a published example: when it ends, the company will stop using the Datasets and delete and destroy copies of Datasets within fourteen days. Some buyers tie their retention to the licence's own length: Ooak says Only the anonymized dataset is retained for the licence term. Grepped says it deletes the source data and every working copy within 30 days of final delivery, of your request, or of the agreement ending, whichever comes first, and confirms it in writing.

The model clock

The third clock is how long the buyer may keep and use models already trained on the data. This is the clock buyers care about most, because a trained model cannot easily forget what it learned, and retraining a large model to remove one holder's records is expensive. So buyers ask for models to survive the licence, usually forever: one template phrases it as a perpetual, worldwide license for training and use of models trained. micro1's terms state that deleting the data does not require Licensee to delete or retrain any model, weights, or derivative works.

What survives, and why silence matters

Every licence has a list of clauses that survive its end: confidentiality, payment already due, indemnities, and the deletion and certification duties themselves. Survival of models belongs on that list only if the licence puts it there. The register's position is short: Any survival must be written in; silence grants none. That cuts both ways. A holder who wants trained models to stop being used at the end of the term must say so; a buyer who wants them to survive must say so too, and should pay for it.

Derived data is the clock people forget. A buyer may make cleaned, labelled or synthetic versions of the records during the term; micro1 says certain datasets may undergo synthetic rewrites that preserve workflow structure. If the licence lets derived data survive while the original is deleted, the deletion clause protects much less than it appears to. A derived copy that could stand in for the original should follow the original's clock. Our guide to what a data licensing agreement contains covers models, derivatives and deletion clause by clause.

Model survival is the realistic trade

Model deletion is hard to win, and a holder rarely gains by spending its leverage on it. What a holder can do is limit the surviving right and price it. Licensing counsel quoted in our register recommends a perpetual right to models already trained plus a finite licence to the data itself, limited to a single model with the right to extend on additional fee. In health and personal-data contexts, the register's position is to limit post-term rights to de-identified data. The result is a licence whose access clock is short, whose copy clock ends in a certificate, and whose model clock is long but narrow: the model trained during the term lives on, and a new model needs a new licence.

2. What disclosed deals show

Disclosed AI data licences that state a length are for a fixed term. Of the 111 deals on our ledger, those that state one are set out below, with one licence from our price records (Kled's); every one of them is finite. They are mostly publishers and content platforms rather than operating companies licensing internal records, so they show what a buyer accepts from a holder with continuing data, not what a buyer pays for a one-off set of records.

Four things stand out.

Two to three years is the centre. Reddit's filing gives the clearest single statement, and our register adopts it as the default for a living company: a two to three year term with renewal and refresh rights keeps the holder's leverage. Longer terms (five and six years) appear where the licensor is a very large content business with a continuing flow of new material, and the six-year Shutterstock licence was itself a renewal.

Many deals say "multiyear" and nothing more. CNN's licence to Google is a multiyear agreement, Universal Music's with ElevenLabs a multi-year licensing agreement, AFP's with Mistral a multi-year agreement. The parties chose to disclose that the licence ends, not when. None of these press statements describes a licence without an end.

Money arrives by the year. Where a deal on the ledger discloses money, it is usually an annual figure: Meta pays News Corp up to $50 million a year, OpenAI pays Dotdash Meredith a minimum of around $16 million per year, and Reddit's licences to Google and OpenAI were reported at about $60 million per year each. An annual price is what a finite, renewable term produces: the buyer pays for each year of access, and the holder gets to negotiate again.

Renewals happen, and they reopen the price. Shutterstock's six-year licence was a renewal and expansion of an earlier one. Tempus AI extended its licence to Recursion by replacing discretionary fees that could have totalled $84 million over two years with committed payments over three. Reddit's licence to Google shows the leverage a term gives a holder: the ledger records a renewal under discussion in 2026. A holder that granted perpetual rights would have nothing to discuss.

3. What buyers' published terms ask for

The buyers that publish standard terms for company records are mostly firms that license data on to AI labs, and their terms tell a different story from the disclosed deals. Several of those that state a grant ask for a licence without an end.

Why firms that resell ask for forever

The pattern has a simple reason. A firm that buys records to package and license on to AI labs needs rights that outlast its own downstream sales. If its licence from the holder ends in two years, every licence it grants to a lab must end then too, which makes its product harder to sell. So it asks the holder for perpetual, sublicensable rights, and sells the labs something narrower. Defined.ai's own licence to its buyers is non-exclusive, non-sublicensable, and non-transferable: the firm in the middle keeps the broad rights and passes on limited ones.

Our negotiation page states the consequence as a lever: Read the grant: a perpetual, sublicensable licence is a resale right. micro1's terms take that right for the one fee the holder agreed, and do bar one use: micro1 may not resell the data as received from Licensor and as a standalone data product. That leaves a great deal it may do: build the records into its own products, license training rights on them through several tiers, and keep doing so after the holder has moved on.

A promise about the past, too

micro1's terms also ask the holder to warrant that it has not previously entered into any transaction granting a third party rights to license, access, use or otherwise exploit the same Data on the same or substantially similar terms. Combined with a perpetual grant, that matters for term planning. A holder that licenses its records for two years to one buyer and then wants to license them to another must check whether the second buyer's terms treat the first licence as a disqualifying earlier grant. Our guide to exclusive and non-exclusive licences discusses how to narrow that warranty.

What the terms leave out

Several buyers publish nothing about length at all. Mercor's company-data page, for example, says We license it, package it, and pay you for it and publishes no licence term. Where the published terms are silent, the length is whatever the licence the holder signs says, and that document is where the holder's attention belongs. The buyer programs compared guide sets the published terms side by side.

4. Two kinds of holder: a living company and a closed one

The right term depends less on the buyer than on what kind of holder is licensing. Two cases cover most of the market.

A living company

An operating company keeps producing records every month: new tickets, new messages, new transactions, new code. It also keeps using its records in its own business. For that holder, a licence is a recurring product, not a one-time event, and a short term is how it captures the recurrence. SimpleClosure, which handles the data of companies that are winding down, describes the split plainly: for a company still operating, licensing can create a new revenue stream while allowing you to continue using the underlying assets. The same lever records that one buyer's estimator values the same company's records at $51k to $71k while operating and $33k to $47k once closed, about 1.5 times more for the operating company.

For a living company, a perpetual grant gives away three things at once: the renewals of the same records, the refreshes of new records (unless the licence carefully excludes them), and the leverage that comes from a buyer needing to come back. That is why the register's default for this holder is the two to three year term with renewal and refresh rights.

A closed company

A company that has shut down is in a different position. It produces no new records, it will not be around to negotiate a renewal, and it usually needs to turn its assets into cash within a fixed time. Under Delaware law, a dissolved corporation continues for the term of 3 years from such expiration or dissolution (or longer if the Court of Chancery directs) for the purpose of winding up, and must adopt a plan of distribution for its remaining assets before that period ends. A licence that needs the licensor to exist in year four to renew it, audit it or enforce deletion is a poor fit for a licensor that will not exist.

For that holder, a one-off sale with perpetual rights is the natural shape, and the register says so: price the sale as perpetual. SimpleClosure's guidance agrees that a company winding down can capture value through an outright sale. Depending on the asset and the buyer, its deals can be licensing the data while you retain the underlying assets or selling assets outright. And Polyshares' disclosed licences of single companies' operational records (a roofing company's at $450,000, a general contractor's at $300,000, a solar company's at $275,000 and a marketing agency's at $110,000) are all perpetual and exclusive: the shape of a sale, whatever the state of the company that granted them.

Even a closed company should know what "perpetual" hands over. SimpleClosure's own terms say that if it or an affiliate acquires the assets, it will thereafter have the right to use, reproduce, modify, distribute, license, sell, or otherwise exploit the Digital Assets for any lawful purpose, without further payment. That is the honest description of any sale, and the price should reflect it. Our guide to selling customer data in bankruptcy covers the court-supervised version of the same sale.

Between the two sits a company that is selling one fixed set of records (say, a product line it closed) while continuing to operate. For that holder the best structure is often a perpetual licence limited to the defined, closed set of records, priced as a sale of that set, with every later record outside the grant. What matters is that the grant names its records, so "perpetual" cannot reach data the company has not yet created.

5. Perpetual, irrevocable and sublicensable: a sale by another name

Each word in a grant like micro1's ( a worldwide, royalty-free, fully paid-up, nonexclusive, irrevocable, perpetual license, with the right to sublicense through multiple tiers) removes one thing a holder would otherwise keep.

• Perpetual removes the end date. There is no renewal, so no second payment for the same records.
• Irrevocable removes the holder's right to end the licence, even if the buyer breaks its terms. The holder's remedy for a breach is then usually a claim for money, not the return of its records.
• Sublicensable through multiple tiers lets the buyer grant the rights on, and lets those it grants them to grant them on again. The holder no longer knows who uses its records.
• Transferable (in Kled's and Handshake's grants) lets the buyer sell the licence itself, for example as part of selling its own company.
• Fully paid-up means the one fee agreed is the whole price, for every use, forever.

Put together, those words leave the holder with a bare title: it still owns the records, but it has given away the economically valuable rights to them, permanently, to a party that may pass them on. That is a sale in substance, even though the document says "licence" and the holder keeps the original files. A grant that is also exclusive, like Kled's or Polyshares', goes further, because the holder can then never license those records to anyone else.

What the law says about labels

The law in several places looks through the label to what is actually granted. Under US copyright law, a "transfer of copyright ownership" includes an exclusive licence whether or not it is limited in time or place of effect, but not including a nonexclusive license. An exclusive grant of copyrighted records (documents, code, written messages) therefore transfers ownership of the rights it covers, and is not valid unless it is in writing and signed.

US tax regulations draw a similar line for digital content. A transfer of a copyright right is a licence, not a sale, when not all substantial rights have been transferred. A copy the buyer must destroy after a period is treated like one it must return: a transaction in which a person acquires a copy of digital content on a disk subject to a requirement that the disk be destroyed after a specified period is generally the equivalent of a return, which the regulation classes with leases rather than sales. And the regulation states that Neither the form adopted by the parties to a transaction, nor the classification of the transaction under copyright law, shall be determinative. A term-limited licence with a deletion duty, in other words, reads as a licence or a lease; a perpetual grant of every substantial right can read as a sale. Whether a payment is a royalty or sale proceeds changes where it is sourced and how it is taxed, especially for a holder outside the US; our guide to US withholding tax on data licence payments covers that, and a holder should take its own tax advice.

Why the price should be a sale price

If the substance is a sale, the price should be measured against what the holder gives up for good, not against one year of access. Three records give a sense of the gap.

Catalogues price forever as many terms. The national language institute's open-ended commercial licence costs 5 million yen against 500,000 yen for two years, the same corpus for the same number of users. The sound-effects catalogue's one-off purchase at €60,000 is three years of its €20,000 a year licence, and that catalogue's own monthly rate, €3,000 / month, shows the same logic at the short end: the longer the commitment, the lower the price per month, and the open-ended right costs most of all.

One-off deals do not repeat. Getty Images guided its 2026 revenue to $948M to $988M, and attributed the decline exclusively to one-time revenue recognition from 2025 licensing deals. A perpetual grant is the extreme case of a one-time deal: the revenue is booked once, and the next year starts from zero.

Exclusivity multiplies again. When the perpetual grant is also exclusive, the holder gives up every other buyer as well as every renewal. Our valuation shows an exclusive licence at 2.5 to 5 times the non-exclusive figure, and the exclusivity guide explains how that range depends on scope.

The grant follows the company

A perpetual licence also outlives the holder's own plans. When the holder's company is itself sold, the buyer's lawyers will ask what rights to its data it has granted. In Ooma's purchase of FluentStream, at a $45,000,000 enterprise value, the target represented that, except as scheduled, it had not let its data (including deidentified data) to be used as Training Data by any other Person. A two-year licence that has ended is a short line in a disclosure schedule. A perpetual, sublicensable licence is a permanent encumbrance on one of the company's assets, and an acquirer will price it.

6. Renewals and refreshes: the recurring revenue of a living company

For a living company, the term is not only a limit on the buyer. It is the mechanism by which the holder gets paid more than once.

What recurring access earns

Our negotiation page records the lever as Sell access to refreshing data, not a one-off file. Reddit's January 2024 licences carried $203.0M of aggregate contract value over two to three years, with a minimum of $66.4 million expected in the first year. Microsoft paid Informa an initial data access fee and then recurring payments over the following three years. The first payment buys the existing records; the recurring payments buy the continuing flow.

A common structure combines a floor with a variable part. OpenAI pays Dotdash Meredith a minimum of about $16M a year, the fixed component only, with a variable component on top. The register's position on price structure follows: Anchor on a fixed minimum plus refresh or usage fees. In a term licence, the fixed minimum pays for access during each year and the refresh fees pay for new records as they are delivered.

What a refresh is

A refresh is a delivery of new records during the term. For a company whose records keep growing, refreshes can matter as much as the original delivery, because current records show buyers how the work is done now, with today's tools and today's ways of working. Centific, which sources data for AI labs, says most of its supply relationships then move to a quarterly refresh after the first accepted contribution.

Refreshes need terms of their own. The register's position is to Specify format, schedule and acceptance; avoid service levels the holder cannot meet; tie refresh fees to new data. The warning about service levels comes from Reddit's own filing, which lists among its risks that Our data license agreements are subject to terms and conditions including API performance requirements, that we may be unable to meet. A holder that promises a delivery schedule turns its licence into a service, and a missed delivery into a breach. A holder that offers refreshes as an option, priced per delivery, keeps the upside without the obligation.

Renewal: selling the same records again

A renewal is a new term for records already licensed. The clearest disclosed sequence is 23andMe's with GSK. GSK chose to extend its exclusive period over 23andMe's database for a fifth year, at a cost of a one-time payment of $50 million. When the exclusive period ended, GSK took a $20 million upfront payment for a one year, non-exclusive data license. The extra year was sold separately, and so was the year after it, first exclusively and then non-exclusively, because the holder had kept the right to sell each one. A perpetual grant at the start would have ended that sequence before it began.

Renewals also reopen the terms in the buyer's favour sometimes. Tempus AI's extension with Recursion replaced fees that could have totalled $84 million over two years with committed payments of $42 million over the next three years. The holder accepted less upside for more certainty and a longer commitment. That is a legitimate trade, and it is only available to a holder whose licence had an end to negotiate around.

Renewal is not guaranteed

The value of a renewal right depends on the buyer wanting to renew. Getty's guidance shows the downside: revenue from 2025 licences was one-time, and declines attributed exclusively to one-time revenue recognition from licensing deals in 2025 followed. Our negotiation page draws the practical conclusion: price refreshes, not repeats. A holder should not price a term licence on the assumption that every renewal will happen. It should price the first term to stand on its own, and treat renewals and refreshes as upside that the term structure keeps available.

A multi-year term does not mean waiting for the money

A longer term can still pay early. Getty recognised about $40M of two multiyear deals worth $65M in the quarter they were signed, and Wiley realised $17M of a $21M content-rights project in its first quarter. A holder can ask for most of a multi-year licence fee at signature and still keep the term finite. The two questions (how long the licence lasts, and when the money arrives) are separate, and our guide to when AI data buyers actually pay covers the second.

What a renewal clause should say

• Who may renew. Renewal by mutual agreement keeps the holder free; an option for the buyer alone should come with a price fixed in advance or a formula.
• At what price. Either a stated renewal fee, or a reopening of price with the records' current size and the market as the basis.
• Notice. A window before the term ends in which either side says whether it will renew.
• What happens without renewal. The copy is deleted and certified; models trained during the term survive only as the licence says.
• Refreshes. Whether the buyer may take new records during the term, at what price per delivery, and whether new records delivered late in the term carry their own minimum period.

7. Evaluation-only and other short licences

Not every licence is meant to be the main event. Buyers often want to test records before committing, and short, narrow licences exist for that purpose.

Evaluation-only licences

An evaluation licence lets a buyer measure whether records improve its models, without the right to ship a model trained on them. It is short (weeks or months), narrow (testing only), and ends with deletion. Its price is a fraction of a full commercial licence: our valuation shows an evaluation-only licence at 0.2 to 0.5 times a full commercial one. That figure is our reading of two data catalogues that list narrower use beside full commercial use for the same data, and those catalogue rows are not yet records on this site, so treat it as an indication rather than a measured market rate.

The risk in an evaluation licence is that it becomes a free training licence. A buyer that evaluates by fine-tuning a model on the records has trained on them. So the evaluation licence should say what evaluation means (measuring, not deploying), require deletion of the copy and of any model trained during the evaluation, and give the holder the right to a certificate. NVIDIA's data agreement, with its duty to delete and destroy copies of Datasets within fourteen days at the end, is the kind of clause that fits. The register's position on licensed purpose is to Name the permitted uses (pretraining, fine-tuning, evaluation, retrieval) and exclude what is not paid for; an evaluation licence names one of those and excludes the rest.

Time-limited exclusivity

Exclusivity is often the part of a licence that should run shortest. A buyer that wants sole use is usually buying a head start, and a head start is worth most at the beginning. Disney's three-year licence to OpenAI carried about one year of exclusivity, and Troveo, which licenses video on holders' behalf, says its exclusivity is time-limited and it has to be earned. A holder can grant a non-exclusive licence of two or three years with an exclusive window inside it, priced separately, and recover the right to license others when the window closes.

De-identification has its own expiry

For records that contain personal data, the term interacts with de-identification. US health regulators note that the Privacy Rule does not explicitly require that an expiration date be attached to an expert's de-identification determination, but that experts recognise changes in technology and available information over time, so determinations often carry one. Our own de-identification standard gives each release an evidence file and a certificate that expires after a year, or sooner when the data, the recipient, the purpose or the known attacks change. A licence that runs longer than the de-identification behind it should say what happens at the certificate's expiry: re-certification before each refresh, and a duty on the buyer not to combine the records with new data that would undo it. A perpetual licence of personal-data records is the hardest case, because the buyer's copy outlives every certificate. The HIPAA de-identification guide covers the health side.

8. Worked comparison: one term, renewals or forever

The point of this comparison is the ratio between structures, not a valuation of any particular company. It uses only ratios taken from price records and the exclusivity multiple our valuation publishes.

The setting

A living company has records that a buyer wants for AI training. The buyer offers three structures for the same records. Call the fee for a two-year non-exclusive licence F.

• A. A two-year term, renewable. The holder receives F for the first term, and F again (or a renegotiated amount) at each renewal. New records can be delivered as priced refreshes.
• B. A perpetual, non-exclusive, one-time licence. The holder receives one payment and may still license the same records elsewhere, but never renews with this buyer.
• C. A perpetual, exclusive, sublicensable licence. The holder receives one payment and gives up every other buyer and every renewal.

What the records say each should cost

Structure B against A. The two catalogues that price the same data both ways give a range. The language institute's open-ended licence costs 5 million yen, which is ten times its two-year licence: ten F. The sound-effects catalogue's one-off purchase at €60,000 equals three years of its yearly licence, which on a two-year basis is one and a half F. So the catalogues that offer both price the open-ended right at somewhere between one and a half and ten two-year fees. Both are small catalogues, not company records, so the range is a guide to the shape of the trade, not a rate.

Structure C against B. Exclusivity multiplies the price again. Our valuation's multiple for an exclusive licence is 2.5 to 5 times a non-exclusive one; the low end rests on the one disclosed same-buyer pair, 23andMe and GSK, where an exclusive year cost $50 million and a later non-exclusive year $20 million. On those ratios, structure C should cost a multiple of B, which is itself a multiple of F.

Reading the comparison

Three conclusions follow.

First, the break-even is a count of renewals. If the buyer offers B at three F, the holder is better off with A only if it expects more than three terms (six years) with this buyer, or if refreshes add enough to make up the difference sooner. If the buyer offers B at one F (the price of a single term for perpetual rights), the holder is giving away every renewal for nothing, and should not accept. The 23andMe sequence shows that renewals can be large: an extra exclusive year at $50 million, and then a non-exclusive year at $20 million.

Second, A carries risk that B does not. Renewals may not happen; Getty's experience is a reminder that one-time revenue does not recur by itself. A holder that values certainty, or that may not be operating in three years, has good reason to prefer a larger one-time payment. That is exactly the closed-company case in section 4.

Third, C should be priced as a sale, against all the buyers it excludes. A perpetual exclusive grant forecloses not only this buyer's renewals but every other buyer's licences, for ever. The price has to be compared with what several non-exclusive licences of the same records could earn; Troveo, for example, advertises an average of 4 sales per dataset, labelled illustrative. Polyshares' perpetual exclusive licences of single companies' records, at $450,000 for a roofing company and $110,000 for a marketing agency, are the prices at which some holders have made that trade; whether it was right for them depends on how many other buyers the records could have reached.

The same test on one holder's records

A holder can run the comparison on its own offer in three steps. Get the two-year non-exclusive price first (that is F). Ask for the perpetual price as a separate number and divide it by F: that is how many terms the buyer is paying for. Then ask how many terms the holder realistically expects to renew, and how many other buyers could license the same records over that time. If the perpetual price does not cover both, the term licence is worth more. The offer check places a written offer's amount against the market and lists the buyer's published clauses, which makes the first two steps quicker.

9. What to ask for

A buyer's first draft will usually ask for the longest rights it thinks the holder will accept. These are the terms a holder can reasonably ask for in return, each with the record it rests on.

On audit, specifically

A term licence is only as good as the holder's ability to know whether the buyer kept to it. The register's position is to ask for records sufficient to show compliance with the licensed uses, an annual compliance certificate, and audit only on reasonable suspicion of breach. The compliance certificate matters most at the end of the term, when the buyer should certify that it deleted the copy, that only models trained during the term remain in use, and that no derived data outside the licence survives. The register also notes the mirror request: buyers want certification of de-identification from the holder before each transfer, so each side certifies the part it controls.

On the buyer's own sale

A licence will usually let the buyer assign it to a successor. For a short term that matters little. For a perpetual, sublicensable licence it means the holder's records can pass to whoever buys the buyer, with all the rights the grant contains. A holder granting a long licence can ask that assignment to a competitor of the holder, or to a buyer in a country our screening excludes, needs its consent. The same logic runs the other way: the holder's own acquirer will want to see every licence the holder has granted, as the FluentStream disclosure shows.

10. How the term is decided in a sale we run

AIData.Broker acts as the holder's agent: the holder itself licenses or sells its data to the buyer, and we do not take ownership of it. Under our mandate the holder alone decides whether to sell or license at all, to which buyer and on what terms, and the mandate names term, field of use and exclusivity among those terms. The holder may reject any offer, including the highest, without giving reasons. When the holder asks, we advise on the terms and help negotiate them, including the term and what survives it; the licence remains the holder's agreement, and the holder takes its own legal advice on it. We run sales only for holders based in the United States, the United Kingdom and Canada.

Two parts of the process bear directly on the term. Every buyer that fits receives the same datasheet with one deadline for offers, and we pass every offer to the holder with its full terms as the buyer stated them. So a perpetual offer from one buyer arrives beside term offers from others, for the same records, by the same date, and the holder can run the comparison in section 8 on real numbers rather than on one buyer's draft. And no buyer gets a first look or a last look; if a buyer's standard terms require that the data has not been offered or licensed to anyone else (micro1's prior-licence warranty is a promise of that kind), we tell the holder in writing what granting it would cost, and the holder decides.

Our fee is a share of what the holder is actually paid, band by band: 15% of the first $250,000, 10% from $250,000 to $1 million and 6% above $1 million. What the holder is paid includes payments for renewals and extensions of a licence made in a sale we run, so we have no reason to prefer a large one-time payment over a term licence that renews: our interest is the holder's, the highest total the records can earn over time. No buyer pays us in a sale we run. The full terms are on the pricing page and in how it works.

Our mandate has a length of its own, and it is a different thing from the licence's term. For 2 months we are the holder's only agent for the data, after which the mandate continues without that exclusivity. How long a buyer may use the data is decided only in the licence, by the holder.

11. Checklist before you agree a term

• Name the three clocks. How long the buyer may receive and train on the data, how long it may keep the copy, and how long models trained on it may be used. Write each one down.
• Decide which holder you are. A living company with continuing records, or a closed set of records sold once. Choose the term accordingly.
• Define the records. Records to a date and named systems; state whether later records are inside the grant, and if so only as priced refreshes.
• Limit model survival. Models trained during the term survive; new models need a new licence; extensions cost extra.
• Put derived data on the same clock as the original.
• Require deletion and a certificate within a fixed number of days of the end, and on breach.
• Write the renewal mechanism. Who may renew, at what price, with what notice.
• Price refreshes per delivery, and avoid service levels you cannot meet.
• If perpetual, price it as a sale. Ask for the perpetual price as a separate number and divide it by the two-year price.
• Strike "irrevocable" or narrow it, so a serious breach can still end the licence.
• Limit sublicensing and transfer, and bind any recipient to the same terms.
• Keep exclusivity inside a window, priced separately.
• Ask for audit and an annual certificate, and a final one at the end of the term.
• Check de-identification expiry against the term for any records about people.
• Get competing offers before accepting a long term from one buyer.

Questions people ask

How long should a data license last? For a company that keeps producing records, usually two to three years with renewal and refresh rights, which is what disclosed deals show: Reddit's licences ran terms ranging from two to three years. For a closed company's records sold once, a perpetual licence is normal, but it should be priced as a sale.

What is a typical term for an AI training data licence? Two to three years is the centre of the disclosed range: Reach licensed to Google for two years, HarperCollins to Microsoft for three years, and News Corp to Meta for at least three years. Some large content licences run longer, such as News Corp's with OpenAI at five years.

Should I accept a perpetual data licence? Only if the price reflects what you give up for good. A perpetual, irrevocable and sublicensable licence is a sale in substance: catalogues that price both ways charge from three years of fees ( one-off purchase) to ten two-year terms ( open-ended licence) for the perpetual right. For a living company, a term licence with renewals is usually worth more.

What happens to AI models when a data licence ends? Only what the licence says. Buyers ask for models to survive, and micro1's terms state that deletion does not require Licensee to delete or retrain any model, weights, or derivative works. Counsel's middle way is a surviving right for models already trained, limited to a single model with the right to extend on additional fee.

Does the buyer have to delete my data when the licence ends? Only if the licence requires it. Ask for deletion of every copy within a fixed number of days, with a certificate; NVIDIA's data agreement requires deletion within fourteen (14) days, and Grepped deletes within 30 days of the agreement ending.

Can I renew a data licence at a higher price? Yes, if the licence leaves the price open at renewal. 23andMe sold GSK a fifth exclusive year for $50 million, and then a non-exclusive year for $20 million. Write who may renew, at what price, and with what notice.

What is an evaluation-only data licence? A short licence that lets a buyer test whether records improve its models, without the right to deploy a model trained on them. Our valuation shows it at 0.2 to 0.5 times a full commercial licence. It should end with deletion of the copy and of any model trained during the evaluation.

Is a multi-year licence paid all at once? It can be. Getty recognised about $40M of two multiyear deals worth $65M in the quarter they were signed. Length and payment timing are separate terms.

What this means for you

Treat the term as three clocks and a price, not one number. For a living company, a licence of two to three years with priced refreshes and a renewal clause is what the disclosed deals show and what keeps the company's leverage; for a closed set of records sold once, perpetual rights are normal, but the price should be a sale price. Whatever the length, write down what survives (models trained during the term, narrowly), what does not (the copy, deleted and certified), and how renewal works. When a buyer's draft says "perpetual, irrevocable and sublicensable", read it as an offer to buy the records, and compare its price with what renewals and other buyers could pay.

If you hold an offer, run it through the offer check, which places its amount against the market and lists that buyer's published clauses, and read each clause against the licence terms register. The negotiation page carries every lever cited here with its sources, and the valuation shows exclusive and evaluation-only figures beside the main one, binds no one and needs no upload. The rest of the licence is in what a data licensing agreement contains, exclusivity in exclusive or non-exclusive, payment timing in when AI data buyers actually pay, the Reddit contracts in Reddit's AI data licensing deals, and the whole process in how to sell company data to AI labs. The rights hub holds the legal background, and how it works shows how a sale we run puts term offers and perpetual offers side by side before the holder decides.

This guide is research from public sources, not legal advice.