Can Water Break Its Own Laws? The Navier–Stokes Problem Boiled Down

Stir a cup of coffee and let go. The swirl spins, wobbles, spreads out, and within a few seconds the surface is still again. You have watched that happen thousands of times. Mathematicians have never been able to prove that it always has to.

The rules that govern that swirl were written down in the 1820s by the French engineer Claude-Louis Navier, and put on firmer ground two decades later by the Irish physicist George Gabriel Stokes. The Navier–Stokes equations are now everywhere. They shape aircraft wings and Formula 1 cars, they run inside every weather forecast, and they model blood moving through arteries. They work astonishingly well.

Nobody knows whether they always work. Hidden in the equations is the possibility that a perfectly ordinary flow, given time, could wind itself so tight that the mathematics simply breaks, at a point where the fluid would have to spin infinitely fast. In 2000 the Clay Mathematics Institute put a $1 million prize on settling the question, one way or the other.

On 8 September 2026, OpenAI said an AI system had done it.

When I told my partner, she asked the obvious question. Wasn’t this the same technology that, not long ago, couldn’t count the R’s in “strawberry”? How do you get from that to one of the hardest unsolved problems in maths?

It’s a fair question, and it’s the one this piece is really trying to answer. To get there you first need to see what it would look like for the equations to fail.

A million dollars, four ways to win

The Clay prize is one of seven Millennium Problems. Only one has been solved so far: the Poincaré conjecture, by Grigori Perelman, who turned the money down. The Navier–Stokes problem is unusual because its official statement spells out both ways to win.

The official problem statement, written by the Princeton mathematician Charles Fefferman, lists four statements, and proving any one of them wins. They come in two pairs. One pair says the fluid never blows up: whatever you start with, the solution stays smooth forever, which mathematicians call global regularity. The other says it can. Fefferman’s statement calls the two sides “existence and smoothness” and “breakdown”.

Each pair also comes in two settings. One is an endless ocean. The other is a “wrap-around” box, where anything that leaves one side reappears on the opposite one, like Pac-Man. The box is a mathematician’s convenience: no walls, no infinity, the same physics.

The crucial difference is the hand. In A and B you stir the fluid once and let go; nothing pushes it ever again. In C and D you may keep stirring for as long as you like, as long as the stirring is smooth (no infinitely sharp spoons, no sudden jerks) and fades away with distance and over time.

That makes the two pairs less like opposites than they first appear. “Left alone, water never blows up” and “with the right hand, water can be made to blow up” can both be true at once. Proving C does not settle A.

What OpenAI says it did

OpenAI’s claim is on the C and D side. According to the company, an internal system of roughly 10,000 AI agents worked for about 88 hours and produced a proof that water starting completely at rest can be driven to blow up by a smooth, carefully designed force. The force touches only one patch of water, acts only for a limited time, and never pushes infinitely hard. A further 17 hours or so went into translating the argument into Lean, a programming language in which a computer checks every logical step. OpenAI has said it will not claim the prize money.

As the paper describes it, the heart of the argument is a self-similar collapse. The hand sets up a slender vortex. Fluid spirals in towards its axis and streams out along it, carrying spin inward: the figure skater again. The core keeps the same shape while its ruler shrinks. Measure it in ever-smaller units and ever-faster clocks and it looks identical, right up to a definite moment that mathematicians call time T. That is the singularity.

Why not just cheat?

If you are allowed any force you like, there seems to be an obvious shortcut: draw whatever exploding flow you want, then work backwards to the force that would produce it. The rules close that door. A flow that blows up would, in general, need a force that blows up too, and the force has to stay smooth right through the moment of blow-up.

So the hand can only nudge, and the water has to do the blowing up itself. OpenAI’s paper says that its shrinking vortex, left to itself, would demand exactly the kind of exploding force the rules forbid. The proof’s trick is to add small, oscillating ripples of flow whose combined push cancels the part of the force that would blow up. That is a genuine piece of mathematics, and it is why researchers take the result seriously even while they argue about what it means.

Why the asterisk

Read literally, the Clay problem allows statement C, and OpenAI’s result targets it. Scientific American asked whether OpenAI had solved the wrong Navier–Stokes problem. Luis Silvestre of the University of Chicago told the magazine that the Clay problem may be settled, but “the most important problem is unsolved.” Others, Diego Córdoba among them, point out that every real fluid is pushed by some outside force, and that it is hard to call something a loophole when the official rules spell it out.

Several other questions are still open.

• Is it right? The Lean check is strong evidence that the formal argument holds. But on 6 October three mathematicians posted a preprint arguing that the Lean version does not match the written proof, so a passing check does not vouch for the paper humans are reading. They do not claim the result is wrong. The written proof runs to 166 pages, and more than a month on Terence Tao said that “no expert I’ve talked to has actually understood the whole argument.” It has not been peer-reviewed. The Clay Institute said on 11 September that the problem “has apparently been settled” but that its evaluation is “deliberately unhurried”; its rules need a peer-reviewed publication and two years of acceptance before a prize committee even sits.
• Could the hand be removed? Not along this road, it seems. On 17 September Peter Constantin, Mihaela Ignatova and Vlad Vicol posted a preprint showing that in a construction of OpenAI’s type, the force cannot be switched off near the point where the fluid blows up. The hand has to be touching the exact spot at the exact moment. Whatever eventually settles A and B will need a different idea.
• Whose idea was it? The mathematician Tristan Buckmaster says he and Levent Alpöge, who now works at Anthropic, had been developing a similar approach, partly using OpenAI’s own coding tools. OpenAI denies that their work fed into its result, while conceding it “cannot rule out” that de-identified usage data helped improve its models. OpenAI cites earlier forced blow-up work by Diego Córdoba and Luis Martínez-Zoroa, though mathematicians who have read the proof say the resemblance is superficial.
• What does it teach us? Terence Tao, one of the world’s leading mathematicians, has warned that a landmark problem solved by a system nobody can question can stall a field rather than spark it. Once a solution exists, nobody gets to explore the problem the slow way again, and if nobody understands why a proof works, there is little to build on.

What hasn’t changed is the question this started with. Stir your coffee, let go, and watch. Every experiment ever run says the swirl will settle. Whether the equations agree, with nobody’s hand in the cup, is still anybody’s guess.

So how does a model go from fumbling the R’s in “strawberry” to this? By being handed a problem where being wrong is cheap and being right is checkable. That’s the whole trick, and it’s a good one. It’s also why Tao’s warning matters more than the headline. A proof you can verify but not understand is an answer without an explanation, and the Navier–Stokes question that keeps physicists up at night, the one with no hand in the cup, still needs the explanation. The strawberry got sorted. The coffee is still spinning.

Watch: Terence Tao on what this means

Numberphile also covered the announcement itself in Navier–Stokes is cracked by AI, with Tony Padilla, and heard Tristan Buckmaster’s side of the credit dispute in The Buckmaster Interview.

Further reading

A note on the graphics: the animations are built to show the ideas, not to simulate the Navier–Stokes equations or OpenAI’s construction. The story is still moving, and this piece reflects what had been reported by 17 October 2026.