For three years the story of what makes a frontier model was a story about text: scrape more of it, filter it better, tokenize it, repeat. That resource is now widely described inside the industry as exhausted, not because the text ran out but because the marginal token stopped teaching the model anything new. What is being funded in its place, at scale and in public, is a different kind of input entirely. Not data you collect, but a world you build: an environment in which an agent acts, fails, and gets a reward.
The last week of June 2026 put a price on that shift. The clearest single datapoint is General Intuition, a New York lab that raised $320 million at a $2.3 billion valuation to train models on gameplay. But it is not alone, and it is not the beginning. It is the moment an idea that had been circulating for a year acquired a market-clearing number.
The idea has a paper trail
This did not come from nowhere. The framing was named almost a year earlier. In September 2025, TechCrunch reported that “Silicon Valley bets big on ‘environments’ to train AI agents,” describing a rush of startups building interactive simulations because static benchmarks and consumer agents kept failing at real multi-step work. Around the same time, Prime Intellect launched its Environments Hub, calling interactive training environments “the playgrounds where agents learn” and naming the core problem directly.
RL environments are the playgrounds where agents learn. Until now, they’ve been fragmented, closed, and hard to share.
The thesis underneath all of it is about post-training. Applied Compute, founded by veterans of OpenAI’s Codex and o1 work, built its company on a one-line bet that Modal’s writeup captured cleanly: as frontier models commoditize, the competitive layer moves to post-training, and the firms that win are the ones that own their reward functions and their evaluation environments. Once you believe that, the environment stops being test infrastructure and becomes the actual product. The money followed the belief.
The map
Here is the sector as the June 2026 rounds define it, sorted by the layer each company occupies. The pattern is that no two are doing the same thing, which is what a real category looks like before consolidation.
| Company | Raised | What it builds | The bet |
|---|---|---|---|
| General Intuition | $320M Series A, $2.3B valuation | Foundation models trained on billions of gameplay clips (“acting in space and time”) | Action data transfers from games to robots and the physical world |
| Patronus AI | $50M Series B ($70M total) | Digital World Models: simulated environments to train and test agents | Static evaluation is dead; simulate the job instead |
| Applied Compute | $80M | Custom post-training for enterprises (DoorDash, Cognition, Mercor) | Owning the reward function is the durable moat |
| Prime Intellect | Backed by Founders Fund, Karpathy, Tri Dao | Open Environments Hub + full-stack RL post-training | The environment layer should be open, not lab-locked |
| Mechanize | Early | RL environments that simulate entire economic systems | The end state is full economic autonomy, not task automation |
Read the “what it builds” column and the layers separate: General Intuition makes the model from environment data, Patronus and Mechanize make the environments, Prime Intellect makes the marketplace and training stack, Applied Compute makes the reward functions for specific enterprises. That is a supply chain forming in real time, and every tier just took on capital.
General Intuition is the sharpest version of the bet
Of the group, General Intuition makes the cleanest claim, so it is worth being precise about what it actually is. The company spun out of Medal, a gaming-clip platform, giving it a proprietary archive of billions of clips of humans acting inside virtual worlds — action-labeled data that written text does not contain. On its own site the pitch is unambiguous.
The most powerful foundation models are trained on written words. But human intelligence far exceeds language.
The company frames itself as “the frontier lab for acting in space and time,” and its stated destination is not better game AI but training environments for robots and physical tasks. Backers reportedly include Jeff Bezos and Eric Schmidt through Khosla and General Catalyst. The detail that tells you how the market values this input: the company reportedly rejected a $500 million acquisition offer from OpenAI before raising, and per Axios is already working on a Series B. Proprietary action data is being priced like the scraped text corpora of 2022, except this time only a handful of companies hold it.
The catch nobody prices at the round
Every environment is authored. That is the fact the valuations do not directly confront. A simulated world is written by people who decide what the agent can see, what counts as success, and which failures matter. That authorship is a feature when it lets you generate hard, fresh, ungameable tasks. It is a liability when the world you built is not the world the agent will actually operate in.
This is the sim-to-real gap, and it is the same structural risk that runs through the eval-to-simulation pivot: you have not removed the trust problem, you have moved it from “do I believe the score” to “do I believe the world is representative.” A game replica that transfers to a warehouse robot is worth $2.3 billion. A game replica that only transfers to more games is worth a games company. The entire sector is a bet that the transfer rate is high, and the transfer rate is precisely the number that no round announcement reports, because nobody yet knows it.
Why it matters
The static-text era had a legible economics: the corpus was a commons, the scrape was cheap, and the moat was compute and talent. The environment era inverts that. The corpus is now something you must manufacture, the inputs are proprietary and expensive, and the moat is whoever owns the world and its reward function. That is why the capital concentrated so fast in the last week of June, and why a gameplay archive can be worth more than most model labs.
The question that decides the whole category is not whether environments are a better training substrate than text. On long-horizon agentic tasks, they clearly are. The question is transfer: how much of what an agent learns in an authored world survives contact with the real one. Until someone publishes that number honestly, every valuation in this map is a wager on an unmeasured constant. The environments are real, the funding is real, and the thing they are all quietly betting on is a transfer rate none of them has yet proven.