rationale:

The median estimate of 50.0% reflects a steady shift toward automated environment generation, while the central 50% interval of 32.0% to 70.0% and wide 80% confidence interval of 18.0% to 85.0% capture severe definitional ambiguities, the lack of external measurement, and a broadened upper tail from rapidly scaling procedural automation. Anthropic is currently the most likely candidate for the largest AI lab by revenue, reporting an annualized run rate above $65B compared to OpenAI's estimated $50B, though Anthropic includes cloud-partner sales while OpenAI does not https://techcrunch.com/2026/10/08/openais-revenue-is-reportedly-20-billion-less-than-previously-projected/ https://techcrunch.com/2026/10/08/openais-revenue-is-reportedly-20-billion-less-than-previously-projected/. Ultimately, the automation dynamics at both labs are similar: OpenAI runs a dedicated "Synthetic RL" team Hiring RL Environment Engineers: 2026 Labs Guide Hiring RL Environment Engineers: 2026 Labs Guide, meaning the underlying automation share remains comparable regardless of which lab takes the top revenue spot.

The Case for a High AI Share The resolution criteria specify that the share is measured "by count," which heavily favors AI systems, particularly because procedural automation can scale much faster than high-level experiment design. Automated pipelines can cheaply generate environments in the tens of thousands by reading code repositories, writing setup recipes, and injecting synthetic bugs RL Environments and RL for Science: Data Foundries and ... RL Environments and RL for Science: Data Foundries and .... Industry sources note that models now draft task variations, test cases, and reward functions, shifting humans to "editor and reviewer" roles Hiring RL Environment Engineers: 2026 Labs Guide Hiring RL Environment Engineers: 2026 Labs Guide. Contrasted with the much lower rates of high-level experimental design by AI, the sheer volume of automated task generation supports a median share of 50%. This aligns with Anthropic's reporting that Claude "leads" 26% of its AI R&D work and that over 80% of merged code is AI-authored Measurements for understanding the pace of AI .... Furthermore, testimony from Ryan Greenblatt in August 2026 highlights that labs are "using huge amounts of AI labor to build RL environments" Ryan Greenblatt – What happens once AI can automate AI ... Ryan Greenblatt – What happens once AI can automate AI ....

Ongoing Reliance on Human Vendors Despite automation, labs continue to spend massively on human-built environments, keeping the median from rising above 50%. Anthropic works with more than a dozen environment contractors RL Environments and RL for Science: Data Foundries and ... RL Environments and RL for Science: Data Foundries and ... and reportedly discussed spending roughly $1B on environments https://epoch.ai/gradient-updates/state-of-rl-envs. Mercor reached a $2B annualized run rate in mid-2026, largely driven by lab demand RL Environment Vendors: 2026 Directory & Rankings | RL List. Human-built tasks are expensive—often costing thousands to tens of thousands of dollars each—and remain critical for quality control RL Environments and RL for Science: Data Foundries and .... Dario Amodei recently attributed alignment incidents to broken vendor-built environments We Must Pace the Frontier - Dario Amodei, confirming that external human contractors still produce a vast proportion of the environments currently trusted in training.

Definitional Ambiguity The extreme width of the distribution is driven by unobservable internal definitions of both "environment" and "primary author." If labs count individual task instances—including cheap, AI-generated variants—the AI share could easily exceed 80%, because rapidly scaling procedural generation has a sheer volume advantage over high-level experiment design. Conversely, if a "training environment" is defined as an overarching suite or distinct design framework credited to the human who directed the work, the share could sit firmly in the 15-25% range. Without public measurement, a median near 50% balances the staggering and rapidly scaling volume of automated task generation against the undeniable persistence of billion-dollar human data pipelines.