AIBRIEF
OpenAI says a swarm of 10,000 agents solved a long standing math problem. Mathematicians pushed back, and the compute behind the run is the new unit of work driving the data center buildout.
AI assistants used to answer a question and stop. The new model is closer to a worker that gives itself hundreds of follow-up prompts to finish a job: building a multi-page website, drafting a research memo, or attempting a math proof. That shift in behavior, from one-shot Q&A to hours-long self-prompting runs, is the load-bearing fact under the current data-center and power-plant buildout (Wired: AI Agents Are Thirsty for Power).
OpenAI published the clearest case study. The company said a swarm of more than 10,000 agents exchanging 2.7 million messages had produced new work on a long-standing math problem. Mathematicians pushed back on the claim, and the public has no way to audit the run: private AI labs publish selectively on environmental metrics, and the company has not disclosed how much compute, or energy, the exercise consumed (Wired).
Wired's Maxwell Zeff put the order of magnitude in dollar terms: that one experiment likely cost "tens of millions" in processing power, "though how much exactly is tough to say." That approximation, not a measured figure, is the public's working number.
The wider trend is not contested. The IEA tracks AI-specific electricity demand at the global level (IEA: Energy Demand from AI), and Pew Research and Our World in Data place US data-center use inside the AI boom (Pew, Oct 2025; Our World in Data). The new uncertainty is who pays for the agent runs, and who measures them.