THE NAVIER–STOKES PROBLEM
A flow at the
edge of infinity.
Stir a cup of water. For a moment, a little galaxy turns inside it. Hidden in that ordinary swirl is a question mathematics has struggled with for almost a century.
Can the equations of fluid motion begin with something perfectly smooth—and lead to something infinitely fast?
On September 8, OpenAI published a proof reporting that, under carefully constructed conditions, they can. [1]
Understand the result ↓Published proof + Lean formalization
Explainer checked September 8, 2026. Independent review is distinct from publication.
3D is unavailable here. The equation, scale explorer, and explanation still work.
Camera follows the shrinking core. Switch to fixed scale to see it contract. Explore the scaling laws ↓
Follow a trace of fluid as it spirals inward and travels along the axis. Teal traces show the outer flow; warm traces highlight the inner region. Colors identify regions, not measured speeds.
An ordinary rule.
An extraordinary limit.
Picture a tiny arrow at every point in a river. Each arrow tells you which way the water is moving, and how fast. All those arrows together are a velocity field. Navier–Stokes tells us how that field changes.
“Smooth” means neighboring arrows change continuously, with well-behaved derivatives. A singularity is a place and time where that smooth description fails. Here, the reported failure is unbounded speed.
For any positive viscosity, there is a specially chosen smooth force that drives an initially motionless, three-dimensional incompressible fluid to unbounded speed in finite time, while its total kinetic energy stays bounded. [2]
It is an existence result: one carefully built example is enough. It does not say that every swirl, storm, or cup of coffee must do this.
Newton’s law, written in water.
Select a term. The symbols have ordinary jobs.
Watch one place in the river.
Stand on a bridge. As the water beneath you speeds up or changes direction, its velocity at that fixed place changes. This term records that change over time.
∇ · u = 0 adds one more rule: a moving parcel keeps its volume. If it narrows across, it must stretch along. Here u is velocity, p is pressure divided by constant density, ν is kinematic viscosity, and f is force per unit mass. [4]
Faster motion. Less space.
Move the slider. Watch the numbers as the core contracts.
Relative scale, logarithmic axes.
Each horizontal step gets ten times closer in remaining time.
The 3D camera follows the shrinking core by default; switch to a fixed camera to see it disappear from view. These are normalized scaling illustrations from §2.1, with h = 0.008 and scale constants set to 1, not exact values of the constructed solution. The slider never reaches the singular time. [2]
How can infinite speed cost finite energy?
Energy depends on how much fluid is moving, as well as how fast. Imagine an ever-taller spike drawn on paper. Its height can grow without limit while its area stays small, if its width shrinks quickly enough.
For fluid, we add up speed squared over volume. A sufficiently small region can contain extremely high speeds without making that total infinite.
Stretching is not the same as getting bigger.
Track a parcel of fluid and it stretches along the axis. Track the region of intense flow and both its radius and height shrink. The radius shrinks faster, so the region becomes relatively more slender. Those are different things to follow. [2]
See the small amount of math behind the sliders
Let τ = T − t be the time remaining. The leading core scales in the paper are ℓr ≍ τ1/2, ℓz ≍ τ1/2 − h, and U ≍ τ−1/2 − h, where 0 < h < 0.01. The symbol ≍ means comparable up to fixed positive factors.
Volume scales like ℓr²ℓz. Multiplying by U² gives Ecore of order τ1/2 − 3h. Its exponent is positive, so this scale tends to zero while U grows. This is the leading core energy scale, not a calculation of the total energy of the full flow. The starting slider position is a normalized reference within the collapse, not the fluid’s initial state of rest.
A vortex alone is not enough.
Making a computer draw an accelerating swirl is easy. Showing that a flow obeys every condition of the equation is the hard part.
- 01
Build a collapsing core.
The paper begins with an inward-spiraling flow and axial outflow. The surrounding fluid must fit that motion consistently.
- 02
Find the imbalance.
Joining the core to its surroundings leaves an error in the momentum balance. Without further work, the required outside force would itself become singular.
- 03
Let small motions carry momentum.
Oscillating pulses around the core supply missing momentum transport. Successive corrections cancel the singular errors, leaving smooth forcing. Use “Reveal the pulses” above for a schematic view. [2]
Think of two people moving objects back and forth across a doorway. Even if neither person travels one way on average, they can still transport a load. Likewise, fluctuations can carry momentum even when their average velocity is zero. That analogy explains a mechanism; it is not the proof.
A paper.
A formal proof.
Work to examine.
OpenAI released a 166-page analytical paper and a public Lean formalization. Lean is a proof assistant: it checks a precisely stated mathematical argument against formal rules. Checking the argument and checking that the formal statement matches the intended problem are both essential.
The repository identifies the results as Clay’s alternatives C (all of three-dimensional space) and D (a periodic setting, like a box whose opposite faces connect). Both permit smooth external forcing. This result does not establish unforced Navier–Stokes blowup. The repository’s separate unforced result concerns the Euler equations, which omit viscosity. [3]
How the AI work happened
OpenAI reports using an unreleased internal model, more capable than GPT-6 Astra, with roughly 10,000 coordinating agents in the successful group. It reports reaching the result in 88 hours, followed by 17 hours of Lean formalization and verification using Astra. These are OpenAI’s reported figures. [1]
OpenAI also acknowledges the priority of Levent Alpöge and Tristan Buckmaster’s concurrent work on forced Euler. The results should be distinguished when crediting the surrounding progress. [1]
Status on September 8, 2026: a published proof claim with a formalization available for scrutiny. This page does not independently certify the complete proof. OpenAI says it will not claim the prize; Clay’s process requires qualifying publication, a minimum two-year interval, and general mathematical acceptance. [1] [5]
The map has an edge.
The water does not.
A mathematical fluid fills space continuously. Real water is made of molecules. An infinity in a continuum equation does not mean a real drop of water reaches infinite speed. It tells us to examine the limits of that description.
This is not a new weather forecast, an explanation of every turbulent flow, or a reason to discard fluid engineering. It is a precise answer to a precise question about what these equations can permit.
The wonder is that a familiar swirl can take us to the boundary between a model of nature and nature itself.
Read it. Check it. Keep asking.
- OpenAI — On the Navier–Stokes Millennium Prize Problem ↗September 8, 2026 · announcement, reported process, concurrent work, and prize intent.
- OpenAI — Finite Time Blowup for Navier–Stokes (PDF) ↗Theorem 1.1; §2, physical explanation; §2.1, scales; §2.2, oscillatory pulses.
- OpenAI — NavierStokesAndEuler, Lean 4 formalization ↗Formal statements, source code, build instructions, and independent-checking instructions.
- Charles Fefferman / Clay — Official problem description (PDF) ↗The governing equations and the four accepted alternatives A–D.
- Clay Mathematics Institute — Millennium Prize rules ↗Requirements for considering a proposed solution.
About these visualizations
The rotatable scene is an original 3D teaching illustration. Its paths use a simple spiral-and-strain field; the rings and arrows explain regions and mechanisms. It does not numerically solve Navier–Stokes or reconstruct OpenAI’s full solution. Path timing, pulse spacing, pressure arrows, and colors are illustrative. The equation animations are qualitative. Only the scale explorer evaluates the displayed power laws, with a representative h and normalized constants.
No finite animation reaches or proves infinity. The written explanation remains available without JavaScript; a static schematic appears when 3D graphics are unavailable. Motion can be paused, and reduced-motion preferences are respected.