The worlds worth training in were mostly built in Europe.
Not a patriotic claim — a supply observation. The games that model work rather than spectacle, and that have been deepened for a decade rather than replaced every two years, cluster on this side of the Atlantic. If you are sourcing environments, that is where the catalogue is.
Three reasons that compound. There is a durable European audience for games about work — driving, farming, flying, running a railway or a city — where the appeal is the system behaving correctly rather than the spectacle. Those studios tend to maintain one title for a decade, adding regions, machinery and rules, instead of shipping a new game every two years. And public funding and tax arrangements in several European countries make a slow-compounding product viable where a purely hit-driven market would not. The result is depth of systems rather than scale of audience — which is exactly what an agent needs and what a recent hit often lacks.
What that catalogue actually contains.
Described by what they simulate rather than by studio, because the thing a lab is buying is the behaviour of the world and not the name on the box.
| Category | What makes it useful to train in |
|---|---|
| Trucking and logistics | Long routes, load physics, fatigue, scheduling and fuel — consequence that arrives hours after the decision. |
| Farming and land management | Seasons, weather, machinery, yield. A planning problem where the feedback loop is a year long. |
| Flight and air traffic | The oldest simulation lineage there is, and still the strictest about fidelity. |
| Trains and infrastructure | Signalling, timetables, braking distance. Narrow action space, unforgiving rules. |
| Military and defence training | Built to be realistic because the buyer required it, long before AI wanted it. |
| City building and public systems | Many interacting subsystems, no single win condition, failures that compound quietly. |
| Economic and grand strategy | Long horizons, adversarial agents, hidden information and markets that respond. |
Depth is a function of time, not budget.
A studio that has shipped the same simulator for ten years has spent that decade adding edge cases: the load that shifts, the weather that closes the route, the machine that breaks in a particular way. Nobody designed those for AI training. They are there because a demanding audience kept asking for them.
That is accumulated world-modelling that cannot be bought quickly. A lab commissioning an equivalent simulation from scratch is looking at years of work and a team that has never had to defend it to players who know the job.
Most of these catalogues have never been looked at this way.
The first public deals in this category are recent, and the studios holding the deepest back catalogues are generally not the ones with a business-development team watching frontier AI. The result is a supply of suitable worlds that nobody has inventoried.
It is also why there is no playbook, and why a studio's first question is usually whether its game qualifies at all. The three properties that decide it.
Same timezone, same law, same train.
An operator in the Netherlands sits inside this supply rather than across an ocean from it: the same working day, the same contract law, and a studio in Prague, Munich or Warsaw within a few hours. For a lab, that means the licensing conversation and the engineering happen in one place.
Tell us what your agents need to learn- Supply here, demand thereThe deep simulators are largely European; the labs buying worlds are largely not.
- One jurisdiction for the whole chainStudio, operator and hardware under European law. Where it runs.
- Documented provenanceA licensed European rights holder is a clean entry in a training-data summary.
About the supply
Why are so many simulation games European?
A long-running European market for games about work rather than spectacle, studios that maintain one deep title for a decade instead of shipping a new one every two years, and public funding and tax arrangements that make a slow-compounding product viable. The result is depth of systems rather than scale of audience.
Why does that matter to an AI lab?
Because depth of simulation is what decides whether an agent learns something that transfers, and depth is exactly what a decade of maintenance produces. The supply of worlds worth training in is concentrated where those studios are.
Does the studio have to be European to work with benned?
No. benned is based in the Netherlands and the environments run here, but a studio anywhere can license a game. The European concentration is an observation about where the deep simulators were made, not a requirement.
Are these studios already licensing to AI labs?
Mostly not. The category is young and the first public deals are recent, which is why most of these catalogues are unexamined from this angle. That is the opportunity and also the reason there is no playbook yet.