How game studios can earn revenue from AI labs
AI labs pay for private, managed copies of realistic games so their agents can practise in them. A studio can license a finished game for this without new players and, with the right partner, without new engineering work. This guide explains what labs buy, how pricing works and how to start.
Yes, by licensing a private, managed copy of the game as a training environment rather than selling the game itself. The studio keeps ownership and grants a named lab a defined use; benned makes the game agent-ready and operates that lab's copy. The lab pays an integration fee, a monthly service fee and a fee per concurrent agent, and the studio earns a share of all three. No source code is required to start, no public release happens, and the live game is untouched.
Why do AI labs want video games?
The newest AI systems are agents: models that take actions to reach a goal. Agents improve by practising, which means acting in a world, seeing the result and trying again millions of times. The real world is too slow, expensive and risky for that. A good game is a ready-made world with consistent rules, many possible situations and a clear sense of success and failure.
Building such a world from scratch takes a studio years. That is why labs have started to buy access to existing ones. In 2026 Google DeepMind took a minority stake in Fenris Creations, the studio behind EVE Online, and runs an offline version of the game on local servers to research long-horizon planning and memory (InvestGame: Fenris Creations and Google DeepMind). Spending on training environments overall is large and growing: TechCrunch reported that Anthropic leaders discussed spending more than $1 billion on them in a single year (TechCrunch: Silicon Valley bets big on environments to train AI agents).
What exactly does a lab pay for?
Not a download and not a dataset. Labs pay for a private, managed environment: a copy of your game that their agents can play at scale, that nobody else can see, and that someone keeps running. In practice that includes:
- A version of the game that runs without a screen on servers, often faster than real time.
- An interface for agents: what the agent sees (pixels or game state) and how it acts (keyboard and mouse, or structured actions).
- Resets, fixed random seeds and snapshots, so experiments can be repeated.
- Tasks with automatic scoring, so the lab can measure whether its agent improves.
- Hosting that serves only that lab, with uptime and support.
Gameplay recordings are a separate market. Companies such as Origin Lab license recorded gameplay from publishers as training data; it reported exclusive partnerships with more than 20 publishers (Pulse 2.0: Origin Lab raises $8M seed). A playable environment is a different product: the lab's agent plays itself instead of learning from recordings.
How is it priced?
A common structure has three parts:
- An upfront integration fee for the engineering needed to make the game agent-ready for that lab.
- A monthly managed-service fee for hosting, uptime and support.
- A usage fee per concurrent agent, which grows with how many copies the lab runs at the same time.
The studio earns a share of each part. Because the lab pays the integration fee, the studio does not have to fund the upfront work. Contract values depend on the game and the lab; the larger labs have signed multiple seven-figure environment and data contracts (Wing VC: RL environments for agentic AI).
Which games are worth the most?
Three qualities decide demand:
- Realism. The closer the world behaves to reality, the more useful what an agent learns in it. Simulators of driving, flight, construction, farming and logistics score well.
- Breadth of situations and actions. Many ways to reach a goal and many things that can go wrong make a world harder to exhaust.
- Customisability. Labs want to vary maps, rules, weather and starting conditions, so agents do not just memorise one world.
How easily agents can connect also matters. If the environment works out of the box, labs stay longer. Release date matters much less: a deep simulator from ten years ago can be more valuable than a recent casual hit.
Do I need to share my source code?
No. There are two ways to make a game agent-ready:
- Engine integration. With access to the project, a small plugin exposes game state, actions, resets and time control. This gives the richest and fastest environment.
- Black box. The normal build runs on a server; the agent sees the screen and sends keyboard and mouse input, like a player. Resets use save games or machine snapshots. This needs only a licence and a build. It fits labs whose agents learn from pixels, which is a large part of the market.
What are the risks, and how do I limit them?
The main concerns are leaks of your build, use beyond what you agreed, and effects on your players. Limit them by keeping the build on controlled hardware, licensing each copy to one lab for a defined use, and keeping training copies completely separate from your live servers. Put scope, term and audit rights in the contract; our contract checklist covers the clauses to include.
Labs have a concern of their own: that an intermediary serving several labs sees how each lab's agents behave. That is why labs prefer to contract with the studio directly, with an operator in the background that keeps every lab physically separate. Read more about how isolation works.
How does a studio get started?
- Pick one or two titles with deep simulation and check whether you hold the rights to license them for this use.
- Have a fit check done under NDA: realism, breadth, customisability and the technical route (engine or black box).
- Agree a pilot with one lab, with the integration fee covering the engineering.
- Once the pilot runs, offer the same environment to other labs, each on its own isolated copy.
Want to know if your game fits? benned runs a free fit check under NDA.
Request a fit checkFrequently asked questions
Can a game studio sell its game to AI companies?
Yes. Studios can license a private, managed copy of a game to AI labs, which use it as a training environment for agents. The studio keeps ownership and licenses a defined use.
How much do AI labs pay for a game environment?
There is no public price list. Deals usually combine an integration fee, a monthly service fee and a fee per concurrent agent. Large labs have signed multiple seven-figure environment and data contracts.
Is licensing gameplay data the same as licensing an environment?
No. Gameplay data is recorded play that models learn from. An environment is a playable copy where the lab's own agent acts and learns by trial and error.
Which types of games are most valuable for AI training?
Realistic simulations with many possible situations and actions, such as driving, flight, construction, farming, logistics and management games, especially when maps and rules can be varied.