Anthropic published a piece this week from its institute on recursive self-improvement. The essay includes a set of internal numbers worth pulling out. As of May 2026, more than 80 percent of Anthropic's production code was authored by Claude. Engineers merged eight times more code per day in the second quarter of 2026 than they did in 2024. In an April research project, roughly $18,000 of compute over 800 cumulative hours recovered 97 percent of a performance gap that two human researchers had addressed at 23 percent in a week.
The Cooling Report is not an AI safety publication. The reason the Anthropic numbers belong here is that they describe a shifting elasticity in compute demand. For most of the last three years, forecasts of accelerator and cooling demand have been gated by how fast human research teams can use the available compute productively. If model development itself becomes substantially automated, the rate-limiting step moves from human researcher hours to available accelerator hours, and the demand curve for compute becomes much less elastic with respect to headcount.
For the cooling industry, that is a forecast input. The grey-space inversion at the data center and the Nvidia wattage roadmap both assumed continued human-paced research throughput. The Anthropic disclosure is the strongest public signal yet that the assumption is loosening at one of the major labs.
Treat the absolute figure as a research result, not a unit economics quote. The interesting structural fact is that the same task ran on automated compute and on human research time, and the automated path closed a gap the humans closed only partially in the same calendar window. If even a fraction of the internal research at frontier labs follows this curve, the aggregate accelerator demand at training time grows on a path that current vendor forecasts do not capture, and the cooling load grows with it.
The number to track is what fraction of new model development cycles at the major labs runs as automated optimization versus human-directed experimentation. The number to discount is the timeline speculation. Anthropic is careful in the essay to say full recursive self-improvement is not here. The cooling read does not require it to be. A partial multiplier on research productivity at a few major labs is sufficient to bend the demand curve for the liquid cooling supply chain past what the procurement teams modeled a year ago. The vendors with available capacity in 2027 and 2028 are the ones who took this signal seriously this year.