OpenAI Solved the $1 Million Navier-Stokes Problem in 88 Hours as the Trillion-Dollar AI Race Heats Up

AI has just cracked one of mathematics’ hardest open questions, and the moment reads as both triumph and warning. The WSJ short traces a year of accelerating AI progress in math that outpaced expectations; in 2000 a Fields medalist called the seven Millennium problems far beyond computers, and a year ago forecasters put the chance of solving any by 2030 near 20%. Now OpenAI says it has published a solution. A Fields medalist interviewed in the video is highly optimistic, while twelve other top laureates worry that corporate obsession with solving problems could erode what mathematics is for, and the same news cycle revived fears about control, with talk of slowing development and questions about whether AI could threaten humanity.

A Century-Old Equation: Why Navier-Stokes Matters

The Navier-Stokes equations, rooted in the 19th-century work of Claude-Louis Navier and George Gabriel Stokes, describe fluid motion from air over a wing to ocean currents. In 1934 Jean Leray proved existence in a generalized sense, but whether solutions stay smooth remained unresolved for about ninety years. In 2000 the Clay Mathematics Institute made smoothness and existence a Millennium Prize Problem with a $1 million bounty; only one of the seven had ever been solved. OpenAI’s claim is that a finite-time singularity can form: a flow starting smooth and at rest can, under a smooth finite-energy force, spin into a vortex that shrinks and speeds up like stretching spaghetti, driving velocity unbounded at a point while the external force stays smooth and energy stays finite, settling statements C and D in the official formulation.

The company describes an unusual production line behind the result. After hearing a rumor on September 1 that two Millennium problems had been resolved, it shifted agents to Navier-Stokes, first running about 100 agents for roughly 50 hours on the related Euler problem, then scaling to about 10,000 autonomous agents on full Navier-Stokes. The agents reached a proof on Saturday, September 5, about 88 hours after launch, with Lean formalization taking another 17 hours via GPT-6 Astra. Nearly five million messages were exchanged, and the computation was said to be powered by an internal model significantly more capable than GPT-6 Astra. Quanta cites a cost of several million dollars, while New Scientist, citing the press conference, puts a customer-equivalent price near $15 million; the BBC corroborates the 10,000-agent, 88-hour scale.

The Economics: $1 Million Prize Versus $15 Million in Compute

Through an economy lens the arithmetic is striking. The Clay prize is $1 million, the reported spend is up to fifteen times that, and OpenAI says it does not intend to claim the prize. The payoff is not the purse but the trillion-dollar race for AI infrastructure and model leadership. With interest rates still high, capital is expensive, yet leading labs are committing trillions in implied valuation to data centers, chips, and talent, and a flagship math proof is priced as a down payment on perceived supremacy. WSJ frames the spend in the millions, and the fact that rival announcements landed within hours normalizes the burn rate; the symbolic prize is dwarfed by the market signal that the next frontier model is within reach.

Priority and competition are entangled with that economy. Quanta and New Scientist note that hours before OpenAI’s announcement, Tristan Buckmaster of NYU and Levent Alpoge of Anthropic announced progress on forced Euler, described as stepping stones to Navier-Stokes, and that OpenAI acknowledges the rumor accelerated its effort. OpenAI cedes priority on forced Euler while asserting its Navier-Stokes proof is distinct and was not built on their prompts or text. Buckmaster’s statement leaves open whether agents may have indirectly benefited from work done with OpenAI models. When being first is priced into trillion-dollar valuations, scientific priority and financial first-mover advantage become hard to separate.

Summit, Fear, and Calls to Slow Down

The video widens from proof to culture and safety. The same day the solution appeared, an Anthropic researcher posted on X that he puts the chance of human extinction in the next decade above 10%, and another Anthropic researcher resigned over safety, saying builders privately believe the technology could wipe everyone out. By the weekend, leaders of four of the biggest AI companies agreed development should slow. OpenAI, meanwhile, hosted a summit for mathematicians at its San Francisco offices starting from a shared premise: AI will soon surpass humans in mathematics. Attendees debated which problems to prioritize, what to teach, and what the shift means for the next generation of mathematicians, framing the breakthrough as an institutional acknowledgement that how knowledge is produced is changing.

Verification will be deliberately slow. Clay president Martin Bridson said the evaluation would be unhurried and absolutely rigorous, even though the Lean formalization gives early confidence. The one previous Millennium solution took years of refereeing before acceptance. While mathematicians line-check the argument, markets will keep pricing mathematical supremacy into trillion-dollar valuations, universities will revisit curricula, and large investments funded in a high-rate environment will be justified with milestones like this one. OpenAI has already signaled it is close on a second Millennium problem, suggesting the pipeline that spent $15 million in 88 hours is being reloaded rather than paused.