Games as training worlds: guides for studios and labs
Practical answers on how licensing games to AI labs works, what it pays and how to do it safely.
Earn from AI labs
What labs pay for, which games fit, and how to license without risk to your IP or players.
- How game studios can earn revenue from AI labsPillar guide
- Do you hold the rights to license your own game?Guide
- Licensing a video game for AI training: the contract checklistChecklist
- Should a studio work with AI labs at all?The objections
- Which games fit, and which ones labs pass onLanding
- What it takes from your sideLanding
- A one-page brief to forward internallyBrief
- What AI labs buy from games: data, evaluation, environmentsThe category
- Gameplay data or a training environment?Comparison
- What a game training environment costsPricing
Game environments for agents
How a commercial game becomes a training environment, and what to require from a provider.
- Turning a commercial game into an AI training environmentPillar guide
- Distillation risk and how isolation removes itExplainer
- The EU AI Act and where your training data came fromRegulation
- Where your environment runs, and what that doesHosting
- Why Europe holds the deep simulatorsSupply
- benned Worlds: environments built for long-horizon planningProduct
What is happening, and what it means
The guides explain how this works and do not change much. The notes are shorter, dated, and about something that just happened or an opinion worth arguing with.
Read the notes- The twelve months in which this became a market
September 2025 to October 2026, in five dated steps: what was funded, what was bought, and why the money went somewhere other than where the argument pointed.
- Equity instead of a licence: reading the DeepMind–EVE deal
Google DeepMind took a stake in the studio behind EVE Online rather than licensing the game. What that structure signals, and why the licence is still the repeatable deal for everyone else.
- Reward hacking is a procurement question, not a research one
When a lab buys an environment, the scoring is most of what it is paying for. Four things to test before signing, and why a provider that calls scoring a configuration step has told you what it is selling.
- The second lab costs less than the first — and the studio keeps the difference
Most of the engineering that makes a game agent-ready is done once. What the second lab pays for is its own tasks and its own hardware, which is why one integration can become several revenue lines.
The dates, with their sources
A maintained list of what has publicly happened in training environments and game-world data — funding, launches, deals, estimates — each line with a link to where it was reported and a mark on how firm that reporting is.
Open the timeline- The category is estimated at about $1bn a year
An analyst piece carries a SemiAnalysis estimate putting combined annual revenue for RL environments at roughly $1 billion, about a year after the category existed at all. Same piece reports Prime Intellect at a $1bn valuation after a $130M round.
15 July 2026 · The category - Arkadium launches GameLab
Evaluation and training against Arkadium’s own catalogue of casual games, with public leaderboards comparing named models to a human baseline. It owns the games it uses, so it needed no licence to start.
16 June 2026 · Evaluation - HUD raises a $16M Series A
One of several rounds into environment vendors building against enterprise software rather than game worlds.
June 2026 · The category - Origin Lab raises $8M for game-world data
A seed round led by Lightspeed, to license recorded gameplay and assets from publishers and package them for world-model builders. The company reported partnerships with more than twenty publishers at announcement.
13 May 2026 · Data
The words this market uses
Observation space, black box, determinism, reward hacking, distillation risk, concurrent agent. Studios and labs often mean the same thing in different words, so every term is defined once and linked from everywhere else.
Open the glossary- Training environment
A world an AI agent can observe, act in, reset and be scored in, so it can learn by trial and error.
- Observation space
Everything the agent is allowed to see on each step: pixels, game state, or both.
- Action space
Everything the agent is allowed to do on each step.
- Black-box environment
The game’s normal build runs on a server and the agent drives it through pixels, keyboard and mouse, without engine access.
- Engine integration
A plugin inside the game project that exposes state, actions, resets and time control directly.
- Determinism and seeds
The ability to replay an episode exactly, by fixing the random seed and the starting state.