What a game training environment costs, and who pays which part.
No two games take the same work, so there is no price list. The structure, though, is the same every time, and knowing it is enough to judge an offer or plan a first deal.
benned prices every engagement in three parts: a one-off integration fee for making the game agent-ready, a monthly managed-service fee for the dedicated hardware and its uptime, and a usage fee per concurrent agent that scales with how many instances the lab runs at once. The AI lab pays all three; the game studio earns an agreed share of each, and funds none of the upfront engineering.
Integration
Making the game run headless, exposing an observation and action space, wiring up resets and seeds, and building the scored tasks the lab asked for.
Managed service
Dedicated hardware for that lab alone, in a Tier III facility in the Netherlands, with uptime, updates and support.
Usage
Metered on instance hours and concurrency — the only thing benned measures about a lab's runs.
Cost follows the build, not the box art.
A ten-year-old simulator with engine access can be cheaper to integrate and worth more to a lab than a recent hit that only runs as a black box.
- The technical route. Black box needs a licence and a build; engine integration needs project access and more engineering, and gives faster, exactly repeatable runs.
- How many tasks are scored. Each scored scenario is designed, tuned and checked against reward hacking. This is usually the largest line.
- How variable the world must be. Varying maps, rules, weather and starting states per seed is what stops agents memorising, and it is work.
- How much hardware runs against it. Dedicated machines per lab are not shared, so the service fee tracks what that lab actually reserves.
Why the lab pays the integration fee.
A studio entering this market should not be funding build risk in it. The integration fee covers the engineering, so the studio's exposure is the licence it grants and nothing else — and a lab unwilling to fund the integration is telling you it is not ready for a pilot.
Request a fit check- Share of all three partsAgreed per game, in writing, before anything is built.
- Recurring, not a one-offUsage pricing means the line grows with how hard the lab trains, long after the integration is finished.
- Repeatable across labsThe same environment can be licensed to several labs, each on its own isolated copy.
Before you ask for a quote
How much does an AI training environment cost?
There is no price list, because the engineering differs per game. Every engagement has the same three parts: a one-off integration fee, a monthly managed-service fee and a usage fee per concurrent agent. The integration fee follows the technical route and the number of scored tasks; the service fee follows the dedicated hardware a lab needs; usage follows how many instances run at once.
What does a game studio earn from it?
A share of all three parts, agreed per game before anything is built. Because the lab pays the integration fee, the studio funds none of the upfront engineering.
Who pays for the engineering work upfront?
The lab, through the integration fee. That is deliberate: a studio should not carry build risk for a market it is entering, and a lab that is not willing to fund the integration is not ready for a pilot.
Is there a minimum commitment?
A pilot is scoped as a fixed engagement with an end date, not an open-ended contract. Minimums, if any, belong to the renewal rather than the first term.
Does a bigger game cost more?
Not by size. Cost follows how much has to be built: whether the route is black box or engine integration, how many tasks need scoring, how far the world must be made variable, and how much dedicated hardware the lab wants to run against.