Perpetual or a 2-year license: which is better for your data?
For a company that is still trading, a 2 to 3 year non-exclusive license usually beats perpetual. Models trained during the term are hard to undo either way, so the term mainly decides whether the buyer pays again for renewal and fresh data. Perpetual or an outright sale suits a one-off archive, such as a company that is closing.
What each option gives the buyer
| Term license (e.g. 2 to 3 years) | Perpetual license | Outright sale | |
|---|---|---|---|
| Buyer can train new models | Until the term ends | Forever | Forever |
| Models trained during the deal | Usually kept, unless the contract says otherwise | Kept | Kept |
| You can license it to others | Yes, if non-exclusive | Yes, if non-exclusive | No: it's theirs |
| You get paid again | At renewal, and for fresh data | Only if the contract adds it | No |
| Fits | A company still creating data | A fixed archive, priced for forever | A company that is closing |
Fixed terms are the norm in the deals that have been made public. Reddit's filing for its stock market listing in February 2024 showed data licensing contracts worth $203 million in total, with "terms ranging from two to three years", TechCrunch reported. Reported deals between publishers and OpenAI ran from 2 years (AP) to 3 (Axios, Axel Springer), 5 (News Corp) and 6 (Shutterstock), according to a review of AI content licensing deals published in May 2025.
The catch: a term doesn't take your data back
When a term ends, the buyer stops getting your data. What it already built from your data is another matter. UK law firm Bristows notes that developers will want to "continue to use any model that has already been trained on content prior to termination". And removing what a model has learned isn't a reliable option: "current unlearning techniques do not reliably remove learned patterns," a March 2026 analysis in IAM found.
Foley & Lardner warned in 2025 that "too many AI contracts are silent on what happens to data, models, and outputs after a contract is terminated." So whatever you sign, spell it out. One practitioner's suggested structure makes the split explicit: a perpetual license for the models trained, and a finite license for access to the dataset itself.
Price a term license as if the models trained during it are permanent, because in practice they usually are. What the term buys you is control over everything after: new models, fresh data, renewal and other buyers.
Why a term usually suits a trading company
- You get paid again. When the term ends, the buyer renews or loses access to your new data. Reddit's deals with Google and others came up for renewal in 2026, with reports that it wanted usage-based fees next time.
- Your data keeps growing. A company still in business adds a year of tickets, threads and decisions every year. One-time deals give the buyer "no updates, no fresh content", in the words of the same May 2025 review, so fresh deliveries are something to charge for.
- Prices are moving. The market is young. Locking in a forever price now means you can't benefit if data becomes worth more.
- You can change your mind. If the deal causes problems with staff or customers, a term ends it, at least for anything new.
When perpetual or an outright sale makes sense
- You're closing the company. There's no new data coming, and no one left to manage renewals. Shut-down startups sell their archives outright; see how to sell company data to AI.
- The archive is fixed. Old records from a system you no longer use won't grow, so renewal has little to sell.
- The price reflects forever. Most of the money is paid up front, and it's well above what a term would bring.
Even then, keep it non-exclusive where you can, so a sale to one buyer doesn't stop you licensing a copy to another.
A worked example
Illustrative numbers only, to show the trade-off. A 60-person support company has eight years of tickets and chat. It gets two offers:
| Offer A: perpetual | Offer B: 2-year term | |
|---|---|---|
| Up front | $300,000 | $180,000 |
| Fresh data | Not included | Quarterly deliveries included |
| After 2 years | Nothing more | Renew at a new price, or license to someone else |
| Over 4 years, if renewed once at the same price | $300,000 | $360,000 |
Offer A wins only if the company expects to close, or if nobody would renew. Offer B carries the risk that the buyer doesn't renew, which is why its first payment should still be meaningful on its own. Before comparing any offers, work out what's left after broker, legal and preparation costs; see how much company data is worth.
Which suits you?
Where is your company?
Trading, and adding data every month
- Ask for a 2 to 3 year term, non-exclusive.
- Charge separately for fresh deliveries.
- Get the renewal terms, or how they'll be set, in the contract.
Old records that won't grow
- A perpetual license can make sense if the price is mostly up front.
- Keep it non-exclusive so you can license a copy elsewhere.
- Require deletion of the raw files once training is done.
Winding the company up
- An outright sale is usually simplest.
- Leave out customer and HR records, as Spirit Airlines' sale did.
- Check who gets the proceeds: creditors and investors usually come first.
Exclusivity offered for a higher price
- Keep the exclusive period short, such as a year.
- Make the premium big enough to cover what other buyers would have paid.
- Limit exclusivity to AI training, not every use.
Clauses to ask for, whichever you choose
- What happens to trained models at the end: are they allowed to stay, and can new models be trained afterwards?
- Deletion or return of the raw data at the end of the term.
- No resale or sub-licensing of your data to others.
- A price for renewal, or a clear way it will be set.
- Payment: a fixed amount up front, plus any fee for each new delivery.
The full list is in data licensing agreement terms. Not sure your data would attract an offer at all? Take the free data check.
Questions
Is a perpetual or a term license better for selling company data to AI?
For a company still trading, a 2 to 3 year non-exclusive term usually is: you can charge again at renewal and for fresh data. Perpetual suits a fixed archive or a company that is closing.
How long do AI data licensing deals usually last?
Public examples run 2 to 6 years. Reddit's filing showed terms of two to three years; reported publisher deals with OpenAI ranged from 2 to 6.
Does the AI company have to delete my data when a term license ends?
Only if the contract says so. Ask for deletion or return of the raw data at the end, and spell out what happens to models already trained.
Can an AI model unlearn my data after the license ends?
Not reliably. A March 2026 analysis found current unlearning techniques do not reliably remove learned patterns, so treat trained models as permanent when you set the price.
What is the difference between licensing and selling my data?
A license lets the buyer use your data on agreed terms while you keep it and can license it elsewhere. A sale transfers it, so you can't sell it again.
Sources
- Reddit says it's made $203M so far licensing its data (22 February 2024) TechCrunch
- The 7 deal points of AI content licensing agreements (9 May 2025) Rob Kelly, Media & the Machine
- Licensing content to train AI: essential points to consider (13 November 2024) Bristows LLP, law firm
- AI contracts in health care: avoiding the data dumpster fire (25 June 2025) Foley & Lardner LLP, law firm
- The machine unlearning problem: AI's hidden licensing and compliance challenges (25 March 2026) IAM
- AI training data licensing: a usable agreement Sergei Tokmakov, Terms.Law
- Reddit weighs cutting Google's AI access as its deal expires (22 July 2026) AI Weekly, reporting CNBC
- Google scoops up Spirit Airlines' data in bankruptcy sale to train AI (17 August 2026) Skift