- A hosted environment is a running copy of your game that a lab’s agents play. The studio keeps the build; the lab gets access.
- The public price shape is an up-front engineering fee, a managed-service fee and a usage or concurrency fee. Captured-data deals use revenue shares instead.
- The terms that matter: no resale or sublicense, no training to recreate your game, isolation per lab, deletion, and what comes back to you from agent play.
- No buyer has published a price for a hosted game environment as of 19 November 2026.
What does it mean to license a game as an AI training environment?
It means letting an AI lab run software agents inside your game, usually on a private copy that you or a host operate. The lab trains or tests its agents there. You keep the build, the code and the assets. This is different from licensing recorded gameplay, where the lab gets files and never runs the game. Checked 19 November 2026.
Labs want these environments. TechCrunch reported in 2025 that "leading AI labs are now demanding more RL environments", and that leaders at Anthropic had discussed spending more than $1 billion on them over a year, citing The Information. A realistic game is a ready-made environment: physics, rules, goals and a world that reacts.
The table lists what companies and founders have said in public about price and terms.
| Company or source | Stated price or terms | Records or access asked for | Date checked |
|---|---|---|---|
| General Intuition CEO (LinkedIn post, 15 Oct 2026) | Up-front engineering fee, managed-service fee and a concurrent-user number; no figure | A privately hosted, managed version of a realistic game | 19 Nov 2026 |
| ZENOS (company FAQ and launch post) | Costs recovered, then net revenue split 50/50; non-exclusive; cloning barred | Licensed capture of released games; no source code or assets shipped | 19 Nov 2026 |
| ZENOS CEO (comment, 15 Oct 2026) | Agent environments priced on concurrency and usage, across 200+ titles planned | Input control, full game state and data outputs | 19 Nov 2026 |
| Origin Lab (FAQ; unrelated to Origin Data Partners) | Core data priced per player-hour; no number published | Gameplay capture with inputs and engine data | 19 Nov 2026 |
| benned (environment host, 23 Oct 2026) | What returns to the studio is a negotiated license term | Separate hardware per lab; host keeps none of the runs | 19 Nov 2026 |
| Buff (press, 14 Jul 2026) | $300,000 pilot with an unnamed AI company | Gamer behavior data, not an environment | 19 Nov 2026 |
How can an indie game studio license its released and unreleased games to AI companies?
There are three routes, and a studio can use more than one for the same game.
- A hosted environment. The lab’s agents play a running copy. This needs a build that runs reliably and an interface agents can use: a way to send inputs, read the game state and reset. General Intuition’s CEO said the more accessible the interface, the better labs keep coming back.
- Captured gameplay through a game data company. The company records play in your build and licenses the files. ZENOS says "Any build works: released, pre-release or an early dev build," and that it ships no source code or assets.
- Code, assets and project history. A lab or a data company licenses the repository and the engine project for training. The game studio data guide covers this route and its engine and store terms.
Unreleased games can work for the second and third routes. A hosted environment needs a build stable enough for thousands of agent sessions, which rules out most prototypes.
How much do AI companies pay a game studio to license its Unity or Unreal projects and assets as an environment?
No buyer has published a price for a hosted game environment. The CEO who described the market gave a price shape, not a number: "some up front fee for the engineering effort with a managed service fee and some concurrent user number tied to the services provided."
That shape tells a studio what to quote. The engineering fee pays for turning a game built for players into one agents can run: headless builds, an agent interface, resets and logging. The managed-service fee pays for keeping it running for one lab. The concurrency or usage fee scales with how many agents play at once.
Captured-data deals are priced differently. ZENOS says it recovers direct costs, then splits net revenue 50/50 with the studio (company terms). Origin Lab, which is unrelated to Origin Data Partners, prices core data per player-hour without publishing the rate. The one priced deal on record in games is Buff’s $300,000 pilot for gamer behavior data, which is not an environment.
What raises the price, by the CEO’s account: "The more realistic the game and the wider state/action space and customizable the more customers." A small arena game with few actions is worth less than an open world with physics and many ways to act. See what your game could be worth.
Some AI companies publish their terms: Patronus AI limits a sample to evaluation by two named reviewers and says it pays on acceptance after a full review, as our Patronus AI review sets out.
A lab's valuation is not a price for data: our General Intuition review explains why its headline figure comes from an investment round, not from a payment for gameplay.
Why would a stranger pay good money for a hosted game instead of code they could rip off for free?
Because a ripped copy is a legal risk and a poor training tool. Labs that train on games they do not license take on copyright exposure in every model they ship. A license with a clear scope, a named rights holder and an audit trail is what their lawyers ask for. ZENOS sells on exactly that point: every session it ships carries the license it was captured under (company claim).
A copy of the code is also not an environment. Someone has to build it, keep it running, connect agents to it, reset it thousands of times and log the state. That engineering is what the up-front fee pays for. A studio that already knows its engine does it faster than a lab that has never seen the code.
Labs also want the copy kept apart. benned, which hosts game environments, says frontier labs work with it because of isolation: "separate hardware per lab and the operator storing none of their runs." A lab will not run its agents on a shared server where another lab could learn from its runs.
What’s to stop a lab from repackaging the game and reselling it?
The contract and the delivery model. Put these in the license:
- No resale and no sublicense. The lab may use the environment for its own training and evaluation only.
- No training to recreate the title. ZENOS says "Labs are contractually barred from training models to recreate titles." Ask for the same clause.
- Host it; do not ship it. A hosted build never leaves your servers or your host’s. ZENOS says it never ships "source code, assets, or mods".
- One lab per copy. Isolated hardware per lab keeps one lab’s access from becoming another’s.
- Deletion and audit. Logs, recordings and any copies are deleted at the end, with written confirmation and a right to audit.
Labs worry about the reverse risk too. General Intuition’s CEO wrote that "labs will be reluctant to work with rollups because of fear of distillation": a host serving many labs might learn from one lab’s agents and help a rival. One host replied that its environments run on the lab’s own machines, offline, so no agent traces come back. Ask any host which model it uses.
What should a game studio check before licensing its game and source code for AI training?
Ownership first, then terms. This page is general information, not legal advice.
- Who owns the game: the studio, a publisher, or both. A publisher that owns the IP must sign.
- Engine and store terms: Unity’s terms and the asset stores restrict AI training, as the game studio data guide explains. Remove store-bought assets and anything tagged NoAI.
- Licensed music, voice and motion capture: check each contract for AI use.
- Player data: leave out accounts, chat and recordings of players who did not agree. If the game is directed to children, the COPPA Rule applies; it defines a child as "an individual under the age of 13" (16 CFR Part 312).
- Artists and staff: tell the team. The partnerships lead of Origin Lab, a game data company unrelated to Origin Data Partners, described a studio whose artists made clear they did not want their assets used to train AI (company claim).
- What comes back: benned argues the exploit findings from agent play are a valuable dataset and that getting them is a license term best agreed before the deal, not after.
- Money terms: fee schedule, revenue share base (gross or net), statements, audit rights, and the right to opt out.
Is licensing a game environment a one-off payment or recurring revenue?
It depends on the model. Under the post that started this market, a commenter asked the right question: "Do you see this as a long term revenue driver or more one-off?" Once a lab has trained on a game, it may move to the next one.
A hosted environment can recur, because the managed-service and usage fees run for as long as the lab keeps agents in the game. A revenue share on captured data recurs for as long as the data keeps selling. ZENOS argues that a one-off buyout caps the studio’s upside, and says "Labs buy libraries, not single titles", so a niche game earns inside a larger catalog. A flat fee for the archive is paid once.
Non-exclusive terms keep more doors open. An exclusive license should cost much more, as the guide to exclusive and non-exclusive licenses explains.
Which companies help game studios license their games as AI environments and gameplay data to AI labs?
In public: game data companies that capture and license games for a share of revenue (ZENOS; Origin Lab, which is unrelated to Origin Data Partners), environment hosts that run isolated copies for labs (benned), and labs that may deal with studios directly (General Intuition’s CEO said his company would be a customer of a service that rolls games up). Each sets its own terms. Ask every one of them the questions above, in writing.
Origin is an independent data partner. We help you license your data. For a game studio that means listing what the studio owns, valuing the build, code, assets and recordings, and working out terms for a hosted copy or a data license. If an offer has already arrived, run it through the offer check or see what your game is worth.
Origin Data Partners works with companies that own data and helps get it bought. We are independent of the services on this page, and a buyer may pay us when a deal closes (how Origin is paid). See what your data is worth.
Common questions
An AI lab offered to license our studio’s game as a hosted environment for its world model. How do we know if the price is fair?
Check the shape of the offer. A fair hosted deal usually pays for the engineering work up front, then a monthly service fee and a usage or concurrency fee. Compare it with what a revenue share on captured data would earn. Check exclusivity, term, resale and cloning bans, and isolation per lab. No buyer has published a price for a hosted game environment, so the number is negotiated.
I am the lawyer for an indie studio. What license terms stop a lab from cloning the game it trains on?
Ask for a use limited to the lab’s own training and evaluation, no resale or sublicense, and an express ban on training models to recreate the title, which one game data company already offers. Deliver as a hosted build rather than code. Add deletion with written confirmation, audit rights and a governing-law clause the studio can enforce. This is general information; draft against the studio’s own facts.
Does a game have to be released to work as an AI training environment?
No, but it has to run. A hosted environment needs a build stable enough for thousands of agent sessions, with a way to send inputs, read the game state and reset. Unreleased games that meet that bar can work. Prototypes usually suit a license of captured play, code or assets better.
Should a studio give an AI lab its source code so the lab can build its own environment?
Usually not as a first step. A hosted copy keeps the code and assets on the studio’s side and lets the studio charge for running it. If the lab needs code, license it separately with a narrow scope, no resale, no training to recreate the game, and deletion at the end.
What can a studio ask to get back from AI labs that run agents in its game?
Ask for the exploits and bugs agents find, aggregated findings on how agents play, or raw agent runs for anti-cheat work. Labs may refuse raw runs because they want their work kept private, so agree it in the license before signing. One environment host says this is easier to add before a deal than afterwards.
Can a small studio license its game as part of a library of games instead of alone?
Yes. Game data companies license many titles together, and one says labs buy libraries, not single titles. Read who keeps what: some recover their costs first and then split net revenue, and some ask for exclusivity. General Intuition’s CEO warned that labs are wary of roll-ups that serve many labs, so ask how each lab’s work is kept separate.
Sources and further reading
- Pim de Witte, CEO of General Intuition, on LinkedIn (15 October 2026), with replies
- TechCrunch: Silicon Valley bets big on environments to train AI agents (21 September 2025)
- ZENOS Data: FAQ
- ZENOS CEO on LinkedIn: ZENOS Data launch (25 August 2026)
- benned on LinkedIn: what studios should ask back from AI labs (23 October 2026)
- Origin Lab: FAQ
- Origin Lab partnerships lead on LinkedIn: a studio whose artists said no (1 October 2026)
- Ynetnews: Buff signs a $300,000 AI data pilot (14 July 2026)
- eCFR: 16 CFR Part 312, Children’s Online Privacy Protection Rule
