Early this month, there was a lot of hype around the claim that OpenAI had solved the Navier–Stokes Millennium Prize Problem. If you only read the headlines, it was easy to come away thinking that AI had finally figured out how water flows down a river, air moves around an airplane wing, or turbulence works.
That is not what happened.
The result is primarily a mathematical breakthrough: an AI system constructed a counterexample showing that, under the conditions allowed by the Clay Mathematics Institute’s official problem, a smooth three-dimensional Navier–Stokes flow is not guaranteed to remain smooth forever.
There is also a controversy about whether the route OpenAI took was inspired by unpublished work from two mathematicians working on a related forced Euler problem. I will come back to that. But first, what exactly was the problem?
From Euler to Navier–Stokes: a quick tour of equations
A fluid can be water in a glass, air moving around a car, smoke rising from a chimney, or oil moving through a pipe.
Instead of tracking every molecule, physicists describe a fluid as continuous fields of velocity and pressure. The equations governing those fields come in two flavors.
Euler equations (1757). Leonhard Euler wrote down equations for an ideal fluid with no internal friction. They capture inertia beautifully, but ideal fluids don’t exist in nature.
Navier–Stokes equations (~1840s). Claude-Louis Navier and George Stokes added a viscosity term — the internal friction that makes honey pour slowly and water splash freely. This is the version that actually models real fluids.
Here’s a key detail that matters later: the viscosity term is linear and mathematically well-behaved. It actually helps smooth things out. The hard part — the term shared by both Euler and Navier–Stokes — is the nonlinear convection term, where the fluid transports its own velocity. The fluid is moving while simultaneously changing the flow through which it moves. Its motion affects the surrounding motion, which then affects how it moves next. So there is a kind of feedback inside the equation. This nonlinearity is the source of turbulence, chaos, and the headaches that have kept mathematicians busy for nearly two centuries.
A useful simplification is:
Navier–Stokes is roughly Euler plus viscosity.
The nonlinear dynamics can create smaller and more intense structures in the flow, while viscosity works in the opposite direction, smoothing them out.
That competition turns out to be central to the story.
But don’t we already solve Navier–Stokes?
Yes, all the time.
Computational fluid dynamics numerically approximates Navier–Stokes to design aircraft, study airflow around cars, model pipes, weather and many other systems.
The Millennium Prize Problem was asking something very different.
Suppose a three-dimensional incompressible viscous fluid starts in a perfectly smooth state. Its initial velocity is well behaved and its total energy is finite.
Will the equations guarantee that the solution remains smooth forever?
Or can it break down after some finite time?
A simple picture is:
t = 0 ------------------------- T ------------------------- ∞
smooth solution ?
Can we always continue the smooth solution past every possible time (T), all the way toward infinity?
Or is there some allowed starting condition and evolution for which a singularity appears at a finite (T)?
This is a question about the mathematics of the equations themselves, not about whether engineers can numerically simulate useful fluid flows.
What did Clay actually ask?
The official Clay problem gives four possible routes to a solution.
| Route | Space | What would settle it? |
|---|---|---|
| A | Infinite space | Prove that unforced smooth solutions remain smooth forever |
| B | Periodic space | Prove the same for periodic flow |
| C | Infinite space | Construct allowed smooth data and smooth forcing for which a global smooth solution fails |
| D | Periodic space | Construct such a breakdown in the periodic setting |
The periodic version can be thought of as putting the fluid in a box whose opposite sides connect. Mathematically this is a three-dimensional torus. It is especially convenient because periodic functions can be studied using Fourier analysis: a complicated flow can be decomposed into spatial frequencies, much like an audio signal can be decomposed into frequencies.
The OpenAI result takes the counterexample route: it establishes C and D.
Finding one counterexample is logically easier than proving that every possible allowed flow remains smooth forever. But C and D were explicitly part of the Millennium Prize Problem. A valid counterexample satisfying those conditions settles the problem in the breakdown direction.
The interesting part of the counterexample
Simply making something blow up is not enough. The construction has to obey several restrictions at the same time.
The fluid starts smoothly. It remains incompressible. In the whole-space construction, its total kinetic energy stays bounded. The viscosity is positive. And the external force used to drive the fluid must itself remain smooth.
Yet the velocity becomes unbounded in finite time. That means finite total energy does not prevent an increasingly large velocity from concentrating inside an increasingly tiny region.
You can imagine a spike becoming taller while also becoming narrower:
|
| /\
| / \
| /\ / \
| / \ / \
--------------------+------------------------
region shrinks
peak grows
The peak can grow without bound while the region containing it becomes small enough that the total energy remains finite.
There is another subtle part. If we first invented a velocity field that blows up and then simply declared that whatever is left over in the Navier–Stokes equation is the external force, that force could itself become singular. That would not solve the Clay problem.
Instead, the construction arranges large contributions from the dynamics, pressure and viscosity so that they cancel in just the right way, leaving an allowed smooth forcing term even while the velocity becomes singular.
To me, this balancing act is the beautiful part of the result.
So did we discover something new about real water?
Not in the way the headlines might suggest. The result is mainly about what the mathematical model permits.
It does not mean that a glass of water can suddenly go crazy and reach infinite velocity. Navier–Stokes treats a fluid as a continuous material. Real fluids are made of molecules, and at sufficiently small scales that continuum description eventually stops being the right physical picture.
So there are really two different questions:
Mathematics: If we trust the Navier–Stokes equations exactly, at every scale, can a smooth solution break down?
Physics: How far can we trust this continuum model as a description of real matter?
The counterexample answers the first question under the conditions allowed by Clay. It does not suddenly change how we model a river tomorrow.
But we should also not downplay what was done. Some critics have waved this away by saying “we already knew turbulence could become wild.” That misses the mathematical point. Physical intuition and turbulence can suggest that complicated behavior is possible. They are not a proof that a perfectly smooth solution satisfying the precise mathematical assumptions of the Clay problem must, or even can, develop a finite-time singularity.
Why Euler matters to the controversy
Euler comes back into the story because mathematicians Tristan Buckmaster and Levent Alpöge were working on a related problem involving the forced Euler equations.
Their result and OpenAI’s Navier–Stokes result are not the same theorem. Euler has no viscosity; Navier–Stokes does. Buckmaster and Alpöge’s Euler result used external forcing, while OpenAI says its AI system also found an unforced Euler blow-up result before attacking forced Navier–Stokes.
But the problems are related, and some of the broad ideas overlap. This is where the controversy begins.
AI did not just produce a proof — it formalized it
There is another part of the result that I think deserves attention.
While the proof was produced by an internal OpenAI model, GPT-6 Astra formalized and verified it in Lean. The formalized proof is published in Lean 4, a proof assistant that mechanically checks the logical steps.
The speed at which AI is becoming capable of this is remarkable. Mathematician Kevin Buzzard received £1 million in funding for a five-year project that includes formalizing Fermat’s Last Theorem in Lean. Again early this month an Anthropic system produced a complete Lean formalization through a different, older route in just 11 days. Buzzard downloaded it, compiled it and confirmed that it checked. His own project is pursuing a more modern proof and has broader goals, including building reusable mathematical infrastructure and tools for human understanding.
Conclusion
What the AI system did was construct a counterexample satisfying the breakdown alternatives in the official Clay formulation: under allowed smooth external forcing, a smooth three-dimensional Navier–Stokes flow can fail to remain globally smooth.
The harder question — whether blowup can happen without any external forcing at all — remains open. The Millennium Problem, as most working mathematicians think of it, is not yet settled.
But strip away the controversy and the hype, and something remarkable did happen. We have used Navier–Stokes successfully for generations. The unresolved question was never whether the equations were useful. It was whether, if we trust the mathematics completely, the equations guarantee that a perfectly smooth solution stays smooth forever.
Apparently, they do not.
And it was not a human mathematician who found the counterexample. It was a swarm of 10,000 AI agents, running for 88 hours, that produced a 166-page proof and then had another AI system verify every line of it.