Three products, one conversation.
"AI companies want to pay for games" covers three different transactions with different buyers, different economics and different contracts. Knowing which one is on the table decides what you should charge and what you should refuse.
Three distinct things. Data: recorded gameplay and assets, which a model learns from by reading. Evaluation: running an existing model against the game to measure how capable it is. And a training environment: a running copy the lab's own agent acts inside, failing and retrying, which changes the model rather than measuring it. They are bought by different teams, priced differently and contracted differently — and a licence written for one can quietly cover another.
| Data | Evaluation | Environment | |
|---|---|---|---|
| What it is | Recorded play and assets | Running a model against the game | A playable copy the agent acts in |
| What it does | Teaches from what humans did | Measures a model that exists | Changes the model by practice |
| Who buys it inside a lab | World-model and perception teams | Eval and safety teams | Agent and RL teams |
| Access it needs | An export | A build and a harness | A build, hosting and operation |
| Shape of the money | One-off asset sale | Project or subscription | Integration, service, per agent |
| Studio effort | Export and rights clearance | Low | A licence and a few engineer hours |
| Easiest to start | Medium | Yes | No |
| Worth the most | No | Rarely | Usually |
The companies, and which column they are in.
Described from what each says publicly about itself. We are in the third column, so read the last row knowing that.
- Origin Lab
Data. Licenses recorded gameplay and game-world assets from publishers and packages them for world-model builders. Raised an $8M seed led by Lightspeed in May 2026 and reported partnerships with more than 20 publishers at announcement.
Data - Arkadium GameLab
Evaluation, and training on its own titles. Launched June 2026 on Arkadium's own casual games, with public leaderboards comparing named models against a human baseline. It owns the games it uses, which is why it needed no licensing to start.
Evaluation - benned
Environments. Licensed commercial simulators from third-party studios, made agent-ready and operated as one sealed copy per lab. The studio keeps the customer relationship and earns on every agent.
Environment
A fair reading of the table above: the three are complements rather than substitutes. A lab doing serious agent work tends to want all three, from whoever does each one properly. When each of these arrived, and what it was funded with, is on the timeline.
Games are a slice of this market, not the market.
Most of the money in training environments has gone to vendors simulating enterprise software— CRM systems, spreadsheets, internal tools. Those are where an agent's commercial value is most obvious today, and that is where the funding went.
The case for games is not volume. It is that a decade-old simulator models physics, consequence and time in ways an office application never will, and that what an agent learns there is meant to transfer out of it. That is a bet on depth, and it is worth stating as a bet rather than as a certainty.
Which games carry that depth- Name the product in the contractData, evaluation and training are separate grants. One clause can swallow the other two.
- Evaluation is the cheap first stepLess access, less commitment, and a reasonable way to work with a lab you do not know yet.
- Only one of the three recursUsage pricing on an environment keeps paying while the lab trains. The other two mostly do not.
About the three
What do AI labs actually buy from game companies?
Three distinct things. Recorded gameplay and assets, used as training data. Evaluation: running models against a game to measure how good they are. And training environments: a playable copy the lab’s own agent acts inside. They are bought by different teams, priced differently and contracted differently.
Is evaluation the same as training?
No, and it is the distinction studios miss most often. Evaluation measures a model that already exists; training changes it. Evaluation needs far less compute and far less access, so it is easier to agree to and worth less. It is also a reasonable first step with a lab you do not know yet.
Which one is worth the most to a studio?
An environment, because it is the only one with recurring revenue attached — usage grows with how hard the lab trains. Data is a one-off asset sale. Evaluation sits in between and is usually the smallest, though it is the easiest to start.
Can one game be used for all three?
Yes, and that is exactly why the contract has to name which one is being granted. A broadly worded data licence can be read as covering environment rights you would want to sell separately later, to a different lab.
Is most of this market about games?
No, and it would be misleading to suggest otherwise. Most funding in training environments has gone to vendors simulating enterprise software — CRM systems, spreadsheets, internal tools. Games are a distinct slice, argued for on realism and transfer rather than on volume.