Notes

The twelve months in which this became a market

Daniël Paaij · 23 October 2026 · 5 min read

In short · When did training environments become a market?

Inside about twelve months. In September 2025 TechCrunch reported that Anthropic leaders had discussed spending more than $1 billion on training environments in a single year; by mid-2026 analyst estimates put the whole category at roughly $1 billion in annual revenue. In games specifically, Origin Lab raised an $8M seed in May 2026 to license game-world content as data, Arkadium launched GameLab in June 2026 to evaluate models against its own titles, and Google DeepMind took a stake in the studio behind EVE Online to use the game for research. What is striking is where the money did not go.

Categories rarely announce themselves. This one did, roughly, between the autumn of 2025 and now, and the dates are worth laying next to each other because the pattern they make is not the one the argument predicted.

We keep the dates themselves on a maintained timeline, with a source on every line and a mark on how firm each one is. The short version: environments turned from research plumbing into a budget line in September 2025, the first game-side companies launched in the spring and summer of 2026, and by mid-2026 analysts were sizing the whole category at around $1 billion a year. This note is about what that sequence shows.

Where the money actually went

Not to game worlds. The funded environment companies are overwhelmingly simulating enterprise software — CRM systems, spreadsheets, internal tools. Fleet AI builds gyms that replicate Salesforce and Excel. Mercor bought two companies in that space inside six months. Prime Intellect, HUD and others raised against the same thesis.

That is worth sitting with rather than explaining away. The intuitive case for games is strong: a decade-old simulator models physics, consequence and time in ways a spreadsheet never will. The money went elsewhere anyway, because an agent that can operate a CRM has an obvious buyer tomorrow and an agent that can drive a simulated truck has one later.

One entry on that list is a different shape from the rest: Google DeepMind took equity in a studio rather than a licence, which says something on its own about how labs currently think about access to a game world.

And in games, the easy two went first

Look at what the first movers chose. Origin Lab took data: export, clearance, sale. Arkadium took evaluation, on games it already owns, which is why it needed no licence from anyone to launch. Both are real products and both are faster to stand up than the third one.

The third — a licensed commercial simulator, made agent-ready, operated as a sealed copy per lab — is still largely unoccupied. Not because nobody thought of it, but because it needs two slow things at once: a licence from a studio that has never sold one of these, and an architecture a frontier lab will accept, which rules out the aggregation that makes the business easy to scale.

We are in that third column, so read this knowing it. The honest summary is that the fast parts of this market were taken first, exactly as you would expect, and the slow part is slow for reasons that have not changed. The three products, side by side.

What would change the picture

One signed deal between a named studio and a named lab, with terms anyone can read. Right now every figure above comes from a company describing itself or an analyst estimating from the outside. That is enough to see a shape and not enough to price anything, which is roughly where everyone in this market currently stands.

Sources