Navier–Stokes, Explained Simply: The $1 Million Water Problem AI Just Shook

Pour a little milk into a cup of tea and watch for a few seconds. The white curls in, folds over itself, makes tiny whirlpools, and then fades into one even colour. You have seen it a thousand times.

Physicists can write down the rule behind every one of those swirls. It fits on a single line. And yet, for more than ninety years, nobody could prove that this rule always makes sense.

That gap, between an equation engineers use every day and an equation mathematicians still don't fully understand, is the Navier–Stokes problem. It is one of the seven Millennium Prize Problems, each worth US$1 million from the Clay Mathematics Institute. In September 2026 it was suddenly all over the news, because OpenAI said an AI system had cracked it.

So what is the actual question? Did AI really solve it? And why should a school student care? We'll go slowly. School maths is all you need to follow along.

First, what is a "fluid" to a mathematician?

Water, air, honey, the blood in your veins: all fluids. A glass of water holds an unthinkable number of molecules, and nobody can track them one by one. So mathematicians cheat, in a useful way. They pretend the fluid is smooth, continuous stuff, and at every point they attach an arrow showing which way the fluid is moving there and how fast. Put all those arrows together and you get the velocity field, written $\mathbf{u}$.

The Navier–Stokes equations tell you how that field of arrows changes from one moment to the next. Underneath, they are just Newton's second law, $F = ma$, applied to every tiny blob of fluid at once. Mass times acceleration on one side. On the other side, every force acting on the blob: pressure from its neighbours, friction inside the fluid, and any outside push such as gravity, a fan, or a spoon.

The equation, one piece at a time

Here is the version the Clay Institute uses for the prize, with the density set to 1 to keep it tidy:

It looks scary. It isn't, once you read it like a sentence. Each piece has a job:

Why one little term causes all the trouble

Look again at $(\mathbf{u}\cdot\nabla)\,\mathbf{u}$. The velocity appears twice, multiplied by itself. That makes the equation nonlinear.

In school you mostly meet linear rules: double the cause, double the effect. Nonlinear rules don't play fair like that. Fast-moving fluid carries itself along faster, so fronts can pile up and get steeper, a bit like waves stacking up just before they break on a beach. Meanwhile viscosity pulls the other way and smooths sharp edges out. Every flow is a tug of war between the two.

In two dimensions (think of a very thin film of soap) mathematicians proved long ago that viscosity always wins: a smooth flow stays smooth forever. In three dimensions, the world we actually live in, nobody could show which side wins every time.

The 3D twist: stretching

Picture a figure skater spinning with their arms out. When they pull their arms in, they spin faster. Fluid can do something similar. In three dimensions, a spinning tube of fluid can be stretched longer and thinner, and as it gets thinner it spins faster, which can stretch it even more. Mathematicians call this vortex stretching.

A flat, two-dimensional swirl has no extra direction to be stretched into, which is a big part of why the 2D case was settled and the 3D case wasn't.

So what exactly is the million-dollar question?

Start with a perfectly well-behaved flow: no sudden jumps, no infinite speeds, a finite amount of energy. Now let the equation run forward in time. Does the flow stay smooth forever? Or can it, at some definite moment, blow up, with the speed somewhere inside the fluid shooting up without any limit?

A blow-up wouldn't mean real water suddenly moves infinitely fast. A real fluid can't do that. It would mean the equation has stopped describing the fluid well at that point, and you would need to go back to thinking about individual molecules. Jim Denier of the University of Adelaide makes a lovely point about this in his explainer for the International Maths Challenge: in simpler fluid models, singular answers have often been a hint of new physics that the model was missing.

Wait, doesn't energy keep things under control?

This is a fair question, and the answer shows why the problem is so slippery.

Sounds safe, right? The catch is the word total. Energy is added up over the whole fluid. A tiny region can have an enormous speed while holding very little energy, the way a single spark is extremely hot but carries almost no heat. A blow-up hides in exactly that loophole: the speed goes to infinity in a shrinking spot while the total energy stays finite.

Four ways to win the prize

The official problem description, written by the mathematician Charles Fefferman for the Clay Institute, splits the question into four statements. Prove any one of them and you have solved the problem.

1. A: With no outside force, in all of space, every smooth starting flow stays smooth forever.

2. B: The same as A, but inside a "wrap-around" box.

3. C: With a smooth outside force, in all of space, some smooth starting flow breaks down.

4. D: The same as C, but inside the wrap-around box.

The wrap-around box (mathematicians call it a torus) works like an old video game screen: walk off the right edge and you come back in on the left. It removes any worries about what happens far away at infinity.

Notice something sneaky. A is not simply the opposite of C, because C is allowed an outside push and A isn't. So proving C would settle the prize question, but it would not tell us whether water with no outside push always stays smooth. That question would still be open.

Two hundred years on one line

• 1822 and 1845: Claude-Louis Navier and then George Gabriel Stokes write down the equations.

• 1934: Jean Leray proves that "weak" solutions always exist. They satisfy the equation in an averaged sense but could, in principle, hide rough spots. He also shows smooth solutions last forever if the starting flow is small enough.

• 1969: The two-dimensional case is settled, with smooth solutions for all time (the standard reference is Olga Ladyzhenskaya's work).

• 2000: The Clay Mathematics Institute announces the seven Millennium Prize Problems in Paris, Navier–Stokes among them.

• 2016: Terence Tao proves that an "averaged" toy version of the equation can blow up. His message: any proof that the real equation stays smooth must use some finer structure that the toy version doesn't have.

• 2023: Diego Córdoba and Luis Martínez-Zoroa develop a way to build blow-ups in related fluid equations. The 2026 construction builds on it.

• September 2026: OpenAI announces its claimed solution.

Only one Millennium Problem has been solved so far: the Poincaré conjecture, proved by Grigori Perelman, who turned down the prize money.

What OpenAI says it did

Here is the claim, in OpenAI's own framing from its 8 September 2026 write-up. A fluid starts completely at rest. A smooth force is applied to it. The energy stays finite the whole way through. And yet, after a finite time, the fluid develops a singularity. That would establish statement C, and OpenAI says D as well.

What does the blow-up look like? A vortex, a spinning swirl, that spirals inward and gets stretched out, which OpenAI describes as becoming "increasingly elongated, like spaghetti". The central region shrinks while it speeds up. The hard part is that this has to happen through the fluid's own motion, not because someone sneaks in an infinite force. All the terms in the equation have to grow huge and still cancel each other out precisely, so that the force left over stays perfectly smooth.

How the proof was found

OpenAI says it used an internal model that it describes as more capable than its GPT-6 Astra model, running as large teams of cooperating AI "agents" that could run code and read a cached copy of the internet. A few details from its account:

"This is not a culmination, but rather a snapshot in time" of AI progress, OpenAI wrote about the result.

What it doesn't settle (yet)

This is the part most headlines skipped. The honest status today is not solved but claimed, and under review.

What is Lean, anyway?

Lean is a proof assistant: a programming language where a proof only "compiles" if every single step follows from the last. Here is about the smallest Lean proof there is:

theorem two_add_two : 2 + 2 = 4 := rfl

The computer checks it and accepts it. For a proof hundreds of pages long, that kind of checking is enormously valuable, because humans get tired and miss things.

Why this matters, even if you never study fluids

• It runs the world quietly. These equations are used in aircraft design, weather forecasting and the study of blood flow. Computers solve them every day, but only approximately, on a grid of points. A simulation can't prove that nothing goes infinite between the grid points. That is why a proof was needed at all.

• It's the same habit you use in class. When you check whether $x = -2$ is allowed in $\sqrt{x}$, or whether you're secretly dividing by zero, you are asking "does my answer exist, and does it behave?" The Navier–Stokes problem asks that same question about every possible flow of water in the universe.

• It shows where maths research may be heading. People pose the question, AI systems search through huge numbers of ideas, proof checkers verify the steps, and people still have to check that the right question was answered.

A tiny glossary

• Velocity field: an arrow at every point showing the fluid's direction and speed there.

• Viscosity: a fluid's internal stickiness. Honey has a lot, water a little.

• Incompressible: it can't be squeezed into a smaller volume.

• Smooth: no jumps or sharp corners anywhere, however closely you zoom in.

• Blow-up (singularity): some quantity, here the speed, becomes infinite after a finite time.

• Euler equations: the Navier–Stokes equations with viscosity removed.

Frequently asked questions

Is the Navier–Stokes problem solved?

Not officially. OpenAI claims a proof of statement C (and D), which would be enough for the prize, but independent experts are still checking it and the Clay Institute lists the problem as active.

Who gets the $1 million?

Nobody yet. Clay waits at least two years after a solution is published before considering it, and OpenAI has said it won't claim the prize.

Did AI really solve it on its own?

According to OpenAI, researchers chose the problems, wrote the prompts, steered groups of agents and pooled their best ideas, while the agents produced the proof itself. The work also built on a 2023 method by two human mathematicians.

Why can't a supercomputer just settle it?

A simulation works with a finite grid and finite precision. Blow-up happens in a region that shrinks towards a single point, smaller than any grid can see. Only a proof can cover every point and every moment.

What's the difference between the Euler and Navier–Stokes equations?

Euler's equations ignore viscosity, so there is no stickiness to calm things down. OpenAI put Euler on its list of "easier" problems, and once its agents had an Euler blow-up, it used that result as the stepping stone towards Navier–Stokes.

Every swirl in your teacup follows an equation that fits on one line. Two hundred years later, we are still finding out what it can do.

If this made you curious about the maths underneath (vectors, derivatives, rates of change), that is exactly what our tutors teach in live classes. Book a free first class: it lasts 45 minutes.

Sources

• OpenAI: On the Navier–Stokes Millennium Prize Problem (8 September 2026, updated 10 September)

• Clay Mathematics Institute: Navier–Stokes Equation

• Wikipedia: Navier–Stokes existence and smoothness

• Wikipedia: Millennium Prize Problems

• International Maths Challenge: Millennium Prize, the Navier–Stokes existence and uniqueness problem (Jim Denier)