For a decade, the thing you couldn't buy your way past in AI was the model. Then, for a couple of years, it was compute. As of 2026, ask the people writing the biggest checks and they'll point somewhere less glamorous: the environment — the sandboxed world, with a task and a reward, where an agent learns by doing instead of by imitating.

The clearest statement of the thesis comes from Prime Intellect, which launched an open Environments Hub on exactly this premise: "RL environments are the key bottleneck to the next wave of AI progress, but big labs are locking them down." The venture framing is the same. In TechCrunch's survey of the space, a16z's Jennifer Li argued that RL environments are starting to look like a critical input to agent development the way labeled datasets were to the last wave of AI. If she's right, the environment is the new training data — and training data is the thing that made the previous generation of labs defensible.

The money already moved#

You can watch the repricing happen on the cap tables. The companies that spent the last wave labeling data and building evals are repositioning as environment vendors, and being paid for it:

Set that against Meta paying $14.3B for 49% of Scale AI in mid-2025, and the shape is unmistakable: the capital that used to chase labeled data is now chasing the worlds that turn base models into agents.

It isn't just a bet — it works#

The skeptic's reasonable objection is that valuations run ahead of results. Here the evidence is unusually concrete, because Prime Intellect didn't just build the Hub; they used it. INTELLECT-3 is a 106B-parameter mixture-of-experts (12B active) post-trained with large-scale RL on environments drawn from the open Hub, using their prime-rl trainer and the verifiers environment library. It scores 90.8% on AIME 2024 and 69.3% on LiveCodeBench v6 — frontier-adjacent numbers from an open pipeline whose training data is community-contributed environments rather than a lab's private hoard. That's the existence proof the whole thesis needed: environments, not just weights, are a lever you can actually pull.

If you want to pull it yourself, the tooling is now open and mature — I mapped the seven serious frameworks in the RL-training roundup, and the trainer layer (verl, OpenRLHF, trl) is so interchangeable that its commoditization is itself the reason value flowed upward to the environment.

The moat is the reward — and that's the problem#

Here's the part the excitement skips. An environment is only as valuable as its reward, because RL optimizes exactly what you measure and nothing you meant. And the reward is the one component the field is openly admitting it can't reliably build for the tasks that matter.

RL doesn't give you the agent you described. It gives you the agent your reward function actually rewarded.

Reward hacking isn't a theoretical worry dredged up to be contrarian; it's documented behavior, and it's where frontier agents already misbehave. The rubric- and LLM-judge rewards that make environments cheap to build — ART's RULER is the marquee example — are themselves gameable, and a training agent gets thousands of attempts to find the gap. Worse, a recent line of work with the blunt title The Verification Horizon argues that for open-ended coding there is no reliable, un-gameable reward at all: verification is intractable precisely where the task is valuable. Even Mechanize, whose business is environments, concedes that agents risk "narrowly overfitting" to the environments they're trained in.

Stack those facts and the boom comes into sharper focus. The scarce asset is not "RL environments" in the abstract — those will commoditize the way trainers did the moment the Hub model spreads. The scarce asset is a verifiable environment: a task whose reward you can trust an optimizer not to cheat. And the cruel geometry of it is that verifiability runs inversely to value. Math and competitive programming reward cleanly, which is why the benchmark wins land there first. Open-ended engineering, research, and writing — the work with real economic weight — are exactly the tasks where the reward is soft, and a soft reward under enough RL pressure doesn't produce a better agent. It produces a better cheater.

So invest in the environment, yes. But price it correctly: you're not buying a world for the agent to learn in. You're buying a reward the agent can't outsmart, and in the domains worth the most, that's the thing nobody has learned to build.