Licensing one of our games to an AI lab
A new revenue line for a finished game, with no public release, no change to the live product, and no engineering budget from our side. One page, written to be read cold.
What is being proposed
Licensing one finished game to one named AI lab as a private training environment — a copy of the game their agents practise in, running on dedicated hardware, operated by benned. Not a sale, not a data transfer, and not a public release. The studio keeps ownership and licenses one defined use.
Why a lab wants it
AI agents learn by acting in a world with consistent rules and real consequences. Building such a world takes years; a finished simulation game already is one. Google DeepMind took a stake in the studio behind EVE Online in 2026 to use it this way, and spending on training environments across the frontier labs runs into hundreds of millions.
What it pays
The lab pays three things and the studio takes an agreed share of each: a one-off integration fee for the engineering, a monthly managed-service fee, and a usage fee per concurrent agent that scales with how hard the lab trains. The third is the reason this is a recurring line rather than a one-off licence fee.
What it costs us
A licence, a build, and a few hours from one engineer who knows how that build is produced — front-loaded into the first week. No art, design, production or QA time. No engineering budget: the lab pays the integration fee, so the studio funds none of the work.
What it does not touch
The live game, the player base, the roadmap and the shipped build are all unaffected. Training copies run on separate hardware with no connection to live servers. No announcement is required. Source code is not needed for the fastest route.
Is this the AI thing people are angry about
No — it is the opposite transaction. The grievance is about work used for training without consent and without payment. Here the studio decides, per title, what a named lab may do; it is paid; the terms are written, auditable and terminable; and nothing generative enters the studio’s pipeline. No art is generated and no role changes.
What we would restrict in writing
One named use and one named lab. No redistribution or sublicensing. No generative reproduction of our assets or characters. Deletion of every copy at the end of the term. A right to audit where the build runs. If a lab will not accept that shape, that tells us something about the lab.
The risk we should weigh
An agent practising in our simulator learns the task, not how to build our game. The narrower risk — a model prompted toward something resembling our assets — is handled by the licence rather than by trust. The remaining question is one of position rather than terms: whether we want to supply frontier labs at all. That is a leadership call, and a legitimate no.
What is actually being asked now
A fit check. It is free, runs under NDA, costs a conversation and a build, and obliges nothing. It tells us whether any title in our catalogue qualifies, which technical route it would take, and roughly what the deal would look like. Most conversations can reasonably end there.