For the last three years, the dominant assumption in AI has been simple: more data equals better models. Companies hoarded datasets. Data labeling became a billion-dollar industry. “Data moats” became boardroom vocabulary.
That assumption is now breaking.
The frontier labs (Anthropic, OpenAI, Google DeepMind) have quietly shifted how they train their most capable models. They are not racing to collect more text. They are building environments: simulated worlds where AI learns by doing, failing, and doing again. The shift is so significant that Scale AI recently reported that nearly half of all their new model training projects now involve reinforcement learning (RL) environments, up from a fraction just eighteen months ago.
This is not an incremental upgrade. It is a rethinking of what AI learning actually means.
The Data Wall Was Real
To understand why environments matter, you have to accept something uncomfortable: the internet is nearly exhausted as a training resource.
The models behind ChatGPT, Claude, and Gemini were trained on essentially everything humanity has written and published: books, code, scientific papers, web pages, forum threads. That corpus is finite. Researchers now estimate that high-quality human-generated text will be effectively saturated as a training source within this decade. You cannot solve that problem by scraping harder.
The response from AI labs was predictable: synthetic data. Generate more text using AI itself. But synthetic data has a ceiling too. Models trained primarily on AI-generated content tend to amplify existing biases and gradually degrade in quality, a phenomenon researchers have started calling “model collapse.” You cannot improve a model by feeding it its own outputs indefinitely.
Something different was needed. And it turns out the answer was hiding in a fifty-year-old branch of computer science.
What an RL Environment Actually Is
Reinforcement learning (RL) is not new. It’s the same family of techniques that taught AlphaGo to beat the world champion at Go. The idea is simple: an agent takes actions in an environment, receives a feedback signal, reward or penalty, and adjusts behavior accordingly. Over millions of iterations, the agent learns what works.
What’s new is applying this at scale to language models, and doing it in environments that reflect the real world rather than board games.
An RL environment, in this context, is a simulated system: a sandboxed browser, a virtual operating system, a mock database, a coded replica of a business workflow. The AI agent is dropped into this environment with a goal. It tries things. It either succeeds or fails. It learns from both. Critically, the environment can be reset and run again, thousands of times, across countless variations of the same scenario.
As Andrej Karpathy, one of the founders of modern deep learning, observed in his year-in-review: by training language models across a wide range of verifiable tasks in different environments, the models begin to spontaneously develop strategies that look, to human observers, like genuine reasoning.
That word, verifiable, is crucial. In a well-designed RL environment, you know definitively whether the agent succeeded. It either completed the task, or it didn’t. That binary signal is far cleaner than human feedback or labeled datasets, where quality is inconsistent and subjectivity creeps in.
Why This Matters More Than Better Data
Consider what data-based training actually teaches a model. It learns patterns. It learns to predict the next word given prior context. Given enough high-quality text, it becomes extraordinarily good at that. But prediction is not the same as competence.
A model trained on surgical textbooks can discuss anatomy fluently. That is not the same as being trained to navigate real uncertainty, adapt when things go wrong, and complete a multi-step procedure under conditions no textbook described.
RL environments bridge that gap. Rather than learning from descriptions of actions, the model learns from the consequences of its own actions.
The practical difference shows up in what today’s most capable AI can actually do. The reasoning models (OpenAI’s o-series, Anthropic’s Claude with extended thinking) that have surprised researchers with their problem-solving ability are, not coincidentally, the ones most aggressively trained through reinforcement learning in simulated environments. The environments taught them to plan, backtrack, try alternative approaches, and verify their own outputs.
This is not something you can teach through text alone, no matter how much of it you have.
The Investment Signal Is Clear
If you track where the money goes in AI, the pattern is unmistakable.
In September 2025, The Information reported that Anthropic had discussed spending over one billion dollars on RL environments over the following year. OpenAI’s total data and environment investment for 2026 is projected to be substantially higher. Scale AI, the largest third-party data provider in the industry, restructured a significant part of its business around building and operating RL environments for frontier labs.
A new category of startup has emerged entirely around this space. Wing Venture Capital, which has tracked the market closely, expects the field to consolidate from roughly twenty early-stage companies today down to three to five dominant platforms by 2030, with the winners being those that earn deep trust with frontier labs early and build the research infrastructure to scale.
This is not speculative. The budget allocations at the world’s most sophisticated AI organizations have already moved.
What Changes for Enterprises
If you run a company deploying AI (and in 2026, most significant companies do), this shift has direct implications.
First, “data strategy” is no longer enough. The question is not just what data you have, but what environments you can create or access to train AI on your specific workflows. A company with rich proprietary data still has an advantage, but only if that data feeds into systems where the AI can practice on realistic simulations of real tasks, not just learn patterns from historical records.
Second, the quality bar for AI deployment is rising. RL-trained models go through structured simulation before they ever touch a live system. Frontier research teams test agents across environments that replicate the exact UI logic, error conditions, and database states of production systems before deployment. If your organization is still deploying models that were only validated on benchmark tests or static evaluations, you are behind the methodology that the best labs now consider standard.
Third, this creates a new competitive layer. The companies that will extract the most from AI over the next five years will not be those with the most data. It will be those that build the richest environments for AI to learn their specific business context. That means simulating your customer service flows, your supply chain decisions, your underwriting processes, and your code review cycles. Not just collecting logs of them.
The analogy is imperfect but useful: data is the textbook; environments are the flight simulator. You would not certify a pilot based on how many aviation textbooks they had read.
The Deeper Shift
There is something philosophically significant happening here that goes beyond training methodology.
For years, the dominant model of AI progress was passive accumulation: gather more human knowledge, compress it, make it retrievable. The intelligence was borrowed from the corpus. Environments represent a different theory: that genuine capability comes from experience, from the feedback loop of action and consequence, from learning what the world actually pushes back with.
That is, roughly, how humans develop expertise. Not by reading everything ever written about surgery, negotiation, or engineering. It comes from doing those things under conditions where failure has real meaning.
The labs are not trying to build better search engines. They are trying to build systems that can act, not just respond.
That distinction, acting versus responding, is what the next phase of AI is actually about. And the environment is where the lesson gets learned.
If you find this useful, forward it to a colleague making AI investment decisions this year. The strategy layer is where most organizations are still behind.
Click here to read this article on Dave’s Demystify Data and AI LinkedIn newsletter.