01 / Reading the equation
Newton’s law, everywhere at once
Instead of tracking molecules, assign a velocity u(x, t) to every point. It answers: “How fast, and in what direction, is the fluid passing this location right now?” The evolving field is an Eulerian description. Following one marked parcel gives the complementary Lagrangian description.
We consider a Newtonian fluid with constant density ρ and constant positive kinematic viscosity ν. “Newtonian” means viscous stress is linear in the local rate of deformation. The incompressible equations express momentum balance plus conservation of volume. [1]
Tap a term to unpack it.
Change at a fixed location
∂ₜu measures how the velocity changes with time while you stay at the same point. A current meter bolted to a bridge measures this kind of change. It is only part of a moving parcel’s acceleration.
Incompressibility: a moving parcel preserves its volume, though its shape can change dramatically.
- u
- Velocity; units m/s. Three unknown components.
- p
- Pressure divided by density: p = P/ρ; units m²/s². This convention removes ρ from the equation.
- ν
- Kinematic viscosity μ/ρ; units m²/s. μ is dynamic viscosity.
- f
- External force per unit mass; units m/s².
- ∇²
- The Laplacian, applied to each velocity component.
Why two acceleration terms?
Let a parcel’s position be X(t), with dX/dt = u(X(t), t). The chain rule gives:
The second term is advection: a parcel encounters a different velocity because it moves somewhere else. A steady river can accelerate through a narrow passage even though every fixed current meter reports a time-independent reading.
In components, the x-acceleration from advection is ux∂xux + uy∂yux + uz∂zux. The field both chooses the direction of differentiation and is the thing being differentiated. That is the nonlinearity.
A steady flow can accelerate you
Local kinematic example: u = (x, −y), in dimensionless units. The arrows never change. Move the time slider to follow the orange parcel.
Here ∂ₜu = 0, but Dₜu = (x, y). Also ∇·u = 1 − 1 = 0. This unbounded affine field is an illustration, not admissible finite-energy initial data for the prize.
Pressure is a constraint force
Pressure is solved alongside velocity; in this incompressible model it is not specified by a separate density equation of state. It adjusts so that the acceleration preserves ∇·u = 0. Think of the tension that enforces a pendulum’s length constraint.
The pressure equation, if you want the algebra
Take the divergence of momentum balance. With ∂iui = 0 and repeated indices summed:
This is a Poisson equation. Solving it couples distant regions, even though the original equation is written with local derivatives. Pressure is determined up to a spatially constant function of time.
Viscosity smooths differences
Viscosity transfers momentum between neighboring parcels. It is not simply a drag −γu against an absolute rest frame. Uniform translation has zero Laplacian and suffers no viscous damping.
For a Fourier mode with wavenumber k, diffusion alone gives amplitude proportional to e−νk²t. Short wavelengths damp faster. The question is whether the nonlinear motion can nevertheless create structure at ever smaller scales quickly enough to become singular.
Where does ν∇²u come from?
The Cauchy stress is σ = −P I + μ(∇u + ∇uT). Momentum balance is ρDₜu = ∇·σ + ρf. Constant μ and ∇·u = 0 reduce the viscous stress divergence to μ∇²u. Dividing by ρ gives the displayed equation.
The velocity field moves itself around; pressure keeps the motion volume-preserving; viscosity smooths its gradients.
02 / The precise challenge
Can smooth motion cease to be smooth?
“Solve Navier–Stokes” here does not mean finding a formula for every flow. Many special solutions already exist. The challenge concerns whether the evolution remains mathematically well behaved for all finite times, starting from admissible smooth data.
Whole-space data must be smooth, divergence-free and rapidly decaying. Seek global smooth velocity and pressure with bounded energy. Alternatively, use a periodic box with opposite faces identified. [2]
The official challenge has an asymmetry. Positive alternatives fix zero forcing; negative alternatives allow smooth forcing. Establishing any one qualifies mathematically. [2]
| Alternative | Space | What to establish |
|---|---|---|
| A | All of ℝ³ | Global smoothness for all admissible initial data; f = 0. |
| B | Periodic box | Same global conclusion; f = 0. |
| C | All of ℝ³ | Some smooth data and allowed smooth force prevent a global admissible smooth solution. |
| D | Periodic box | Corresponding breakdown with allowed smooth forcing. |
Forcing must satisfy smoothness and decay requirements, including through the singular time. See pp. 1–2. [2]
C is not the logical negation of A. A theorem saying “this specially driven fluid breaks down” does not imply “an unforced fluid can break down.” It is consistent, logically, for the forced breakdown alternative and the unforced global-smoothness alternative both to be true.
Three words that carry a lot of weight
Existence
There is a velocity and pressure satisfying the equations and the specified data in the required class. Failure of smooth existence does not mean that no generalized continuation is possible.
Smoothness
All orders of derivatives exist and are continuous. A flow can remain smooth at every t < T yet lose this property as t approaches T. “Blowup” is failure at a finite endpoint, not an enormous but finite number.
Uniqueness
The same data determine only one solution within the chosen class. Classical smooth solutions are locally unique under the standard hypotheses. Generalized, or weak, solutions require a separate analysis; smoothness, existence and uniqueness are different questions.
What was already known?
Sufficiently regular three-dimensional initial data give a smooth solution for at least a short time. In two dimensions, smooth unforced solutions continue globally; suitable small-data results also hold in three dimensions. Leray’s global weak solutions provide another kind of existence: equations interpreted after integration against smooth test functions, with an energy inequality. Weak existence alone does not prove smoothness. [3]
So the difficult gap is between “the evolution gets started” and “nothing singular can happen, however long we wait.” A blowup construction closes that gap in the opposite direction for the data it constructs.
Announcement versus acceptance
On 8 September 2026, OpenAI announced a proof targeting C and D. Clay’s page still labels the problem unsolved at this snapshot. [4] [5] The prize rules require a qualifying publication, at least two years, and general acceptance before consideration. That institutional process is distinct from whether a new proof is correct. [6]
03 / What makes it interesting
The obvious energy argument almost works
Multiply the momentum equation by u, integrate, and use incompressibility. On all space with appropriate decay, or in a periodic box, surface terms disappear:
Advection redistributes kinetic energy but contributes no net total. Pressure also contributes no net work in this setting. With no forcing, kinetic energy can only decrease. With forcing, the final term is external power.
See the two cancellations
Integrating the viscous term by parts gives −ν∫|∇u|². These formal manipulations are justified for the smooth decaying or periodic solutions being discussed.
Why is that not the whole proof? Because an integral bound does not bound the maximum. Energy can concentrate in a vanishingly small volume.
A taller peak with less total energy
An exact scaling exercise for spatial fields, not a simulated time evolution. Start with a smooth localized divergence-free shape U. Define uℓ(x) = ℓ−1U(x/ℓ).
The dashed outline is the original region. The readouts, rather than arrow length, encode speed.
In three dimensions, a length reduction by ℓ gives volume ℓ³. Speed squared grows as ℓ−2. Thus E ∝ ℓ, which tends to zero while the peak speed tends to infinity. This is an algebraic possibility; making a sequence like this satisfy the evolution equation with admissible data is the hard part.
Why mathematicians call energy “supercritical”
Navier–Stokes has the scaling symmetry
Viscosity stays the same; a force transforms as fλ = λ³f(λx,λ²t). Energy becomes Eλ(t) = λ−1E(λ²t). Zooming toward small scales and larger speeds makes the energy smaller, so the energy estimate loses control in exactly the dangerous direction. In 2D the corresponding energy factor is λ⁰.
The extra freedom in three dimensions
Define vorticity ω = ∇ × u. It measures local rotation; for rigid rotation it equals twice the angular velocity. Taking the curl removes pressure:
The first term on the right is vortex stretching. A velocity gradient along a vortex can amplify its vorticity. The amplified vortex changes the velocity gradient in turn. This is a possible feedback loop, not by itself a proof that blowup occurs.
In a strictly two-dimensional flow, u = (ux,uy,0) is independent of z, while ω points along z. Hence (ω·∇)u = ωz∂zu = 0. That stretching mechanism is absent.
The race of timescales
For a feature of size L and velocity U, transport takes approximately L/U and viscous diffusion takes L²/ν. Their ratio gives the Reynolds number:
At small L, diffusion becomes very fast—but U may also grow. A single bulk Reynolds number does not settle what happens in a highly concentrated, anisotropic flow.
The challenge is to control local intensity using laws that mostly constrain global totals.
Why a physicist should care
This tests the internal reach of an extremely successful continuum theory. It also probes how nonlinear systems transfer activity across scales. Turbulence makes that transfer familiar, but turbulence and singularity are different: a turbulent solution can be smooth, and an engineered singularity need not be a typical turbulent flow.
Even a definitive counterexample would not supply a universal turbulence model or make ordinary engineering simulations useless. Its payoff is a precise account of a failure mechanism—and new mathematics for understanding nonlinear evolution.
04 / The result announced today
A carefully driven fluid, starting at rest
Scope as of 8 September 2026. OpenAI has released a paper and Lean formalization claiming forced three-dimensional Navier–Stokes breakdown. This guide explains that announcement; it is not an independent verification of the complete proof or a statement of Clay acceptance. [4]
The announced construction starts with zero velocity. A smooth external force drives it toward unbounded speed in finite time while kinetic energy stays bounded. [4] The public formalization describes whole-space and periodic results for every ν > 0, targeting alternatives C and D. [7]
“Starts at rest” makes the role of forcing especially clear. With u0 = 0 and f = 0, u = 0 remains a solution. External forcing first supplies energy; a later singularity cannot be attributed to an already singular initial velocity.
The construction in five moves
- Concentrate a vortex. Inward spiralling and axial outflow accompany a shrinking core.
- Identify the imbalance. Matching an exterior leaves a singular momentum residual in an annulus.
- Supply internal momentum transport. Oscillatory pulses provide averaged quadratic momentum fluxes cancelling that residual.
- Use growth, then damping. Shear amplifies tiny seeds; changing wavelengths later enable viscous damping.
- Correct and localize. Corrections control all residual derivatives; localization gives smooth compactly supported forcing. [8, §§2–3]
How can a zero-mean wiggle transport momentum?
Here is a general fluid-mechanics calculation that makes step 3 less mysterious. Imagine fluid crossing a horizontal interface between two layers. Upward fluid carries extra rightward momentum; downward fluid carries a rightward momentum deficit. The average upward velocity can vanish while the net upward transport of rightward momentum is positive.
Both mean components are zero. Their mean product is ab. A sinusoidal version, vy = a cos θ and vx = b cos θ, gives ⟨vyvx⟩ = ab/2.
This is the same basic reason that random variables can have zero means but nonzero covariance. In a fluid, that covariance is a momentum flux. A spatial variation in the flux changes the momentum of a region.
Connect that calculation to the nonlinear term
For an incompressible field, (u·∇)u = ∇·(u⊗u). Split velocity into mean U and a zero-mean fluctuation v, using an averaging operation compatible with the derivatives:
The mean momentum equation therefore contains ∇·⟨v⊗v⟩, often called a Reynolds-stress contribution (sign conventions vary). Zero mean velocity does not remove quadratic effects. This identity is an explanatory model, not a reconstruction of the paper’s pulse estimates.
Why choosing the force is not a shortcut to any answer
For any chosen divergence-free field and pressure one can formally write
That identity alone says nothing about whether the force is allowed. For a generic singular field, derivatives in this residual will also be singular. Smoothness for t < T is insufficient: the force must extend smoothly through T. A force proportional to 1/(T−t), for example, would simply insert the disaster from outside.
Nor can a spatially uniform acceleration be used to accelerate an infinite mass of fluid in the whole-space problem: the decay and energy hypotheses exclude that construction.
The geometry, with the paper’s scales
Let τ = 1−t and take fixed 0 < h < 0.01. For the leading core, up to constant factors:
| Quantity | Scale |
|---|---|
| Radius | τ1/2 |
| Axial extent | τ1/2−h |
| Swirl and axial speed | τ−1/2−h |
| Core energy | τ1/2−3h → 0 |
The core becomes more slender while shrinking in every direction. Radial viscosity remains in the leading balance. [8, §2.1]
Check the energy exponent yourself
Multiply speed squared by cylindrical volume: τ−1−2h × τ × τ1/2−h = τ1/2−3h. The allowed h makes the exponent positive. This is the concentration calculation from Chapter 3 with different radial and axial scales.
What has—and has not—been settled by this claim?
| Question | Consequence if the proof is correct |
|---|---|
| Can an admissibly forced viscous flow break down? | Yes: the claimed counterexample targets C and D. |
| Can an unforced viscous flow break down? | This construction does not answer it. |
| Does every flow blow up? | No such conclusion. An existence counterexample says “some,” not “all.” |
| Is a weak continuation impossible? | No. Smooth breakdown is a different claim. |
| Can we make a real fluid attain infinite speed? | No physical infinity follows from the continuum theorem. |
Several developments arrived together
Alpöge and Buckmaster separately released smooth-forcing blowup work for incompressible porous media, Boussinesq and three-dimensional Euler. Tao’s September 7 explanation discusses an iterative method that amplifies fine-scale disturbances while controlling the force, building on Córdoba and Martínez-Zoroa. Those equations and that announcement should not be conflated with the viscous Navier–Stokes claim. [9]
OpenAI’s repository also describes an unforced Euler result: singular growth of the velocity’s C¹ norm and a divergent time integral of maximum vorticity. Euler sets ν = 0, so that is a separate theorem from the forced viscous result. [7]
What does Lean verification add?
A proof assistant checks formal deductions from the definitions and assumptions encoded in it. That is stronger than a simulation showing a very large number or a language model judging a proof plausible. But the formal statement must match the intended mathematical problem, its assumptions must be inspected, and the checking process must actually succeed.
The release includes a Lean project and instructions for independent checking. I have not built or audited that formalization here. [7]
05 / What “explosive” means
An infinity at a finite deadline
In PDE language, blowup means a quantity needed to continue the smooth solution becomes unbounded as t approaches some finite T. It does not mean a blast wave, an outward explosion, or an infinite release of energy.
For intuition compare two dimensionless ordinary differential equations, both with y(0) = 1:
Exponential growth is finite at every finite t. The second law reaches arbitrarily large values before t = 1. Its rate accelerates with its current size strongly enough that the remaining time runs out.
Approach the deadline
Toy ODEs, not solutions of Navier–Stokes. The vertical axis is logarithmic so both curves remain readable.
The slider stops at 0.99. The infinity at 1 is a mathematical limit, not a value a finite plot can display.
How infinitely many doublings can fit
For y = 1/(1−t), the values 1, 2, 4, 8, … occur at t = 0, 1/2, 3/4, 7/8, …. The intervals halve each time and sum to 1. This is a useful picture of accelerating activity across scales; it is not a derivation of the Navier–Stokes construction.
What actually becomes infinite?
The announced viscous result concerns unbounded velocity magnitude, while total kinetic energy remains bounded. [4] The notation
says that arbitrarily large peak speeds occur arbitrarily close to T. It does not say the speed is monotone, that every location speeds up, or that a finite volume of fluid attains an actual infinite speed. The norm ‖u‖∞ is the spatial supremum of the speed.
How can an incompressible region shrink?
Mark a particular set of particles with dye. Its volume is preserved. Now mark the places where speed exceeds a threshold. That region need not follow those particles: fluid can enter and leave it, and its volume can shrink.
Think of the bright region in a moving spotlight: the region and the material passing through it are different objects. Similarly, a shrinking core of intense motion need not be a shrinking material parcel. There is no conflict with ∇·u = 0.
Does the mathematics predict a physical catastrophe?
It predicts breakdown of a smooth solution of an idealized system for particular data. A real fluid has molecular scales, compressibility, temperature-dependent material properties and other effects omitted here. Those effects may become important long before a mathematical singularity.
A theorem of this kind does not decide which additional model takes over. Possibilities include compressible hydrodynamics, thermal effects or kinetic descriptions; one need not automatically switch to tracking every molecule.
It also does not show that the constructed forcing is easy to realize, that blowup is stable under experimental errors, or that ordinary stirred water commonly follows this trajectory. Those are separate physical and mathematical questions.
Finite energy is not a universal speed limit. It constrains the integral of speed squared, not the largest speed at a shrinking set. That distinction—not detonation—is the heart of the “explosion.”
Was “explosive” also about the news controversy?
Possibly. Buckmaster’s accompanying statement raises allegations about credit, communications and possible use of unpublished work. OpenAI disputes improper access and says its proof was developed independently. Those are contested provenance claims, separate from the mathematical meaning of blowup; this guide does not adjudicate them. [10] [4]
06 / Check your understanding
Five questions to take away
Try an answer before opening each panel. These are conceptual checks, not a test of fluid-mechanics vocabulary.
1. If ∂ₜu = 0 everywhere, must every parcel have zero acceleration?
No. A steady field can vary in space, so a moving parcel can have nonzero (u·∇)u. The current meter is stationary; the parcel is not.
2. Why does a bound on kinetic energy not rule out unbounded speed?
Because E integrates |u|² over volume. A peak can grow while its occupied volume shrinks faster. The Chapter 3 example has peak ∝ 1/ℓ and energy ∝ ℓ.
3. Why is a smooth force that drives blowup a nontrivial construction?
Choosing a singular velocity and defining its residual as a force usually produces a singular force. The cancellation must leave every derivative of the force smooth through the singular time. An externally imposed infinite force would evade the problem.
4. Would forced Navier–Stokes blowup prove unforced Navier–Stokes blowup?
No. Those are different quantifiers over different input classes. The official negative alternatives permit forcing, while A and B ask about zero forcing.
5. Can oscillations with zero mean velocity affect the mean flow?
Yes. ⟨v⟩ = 0 does not imply ⟨v⊗v⟩ = 0. Correlated fluctuations transport momentum. Spatial differences in that transport affect the mean motion.
A compact map of the ideas
Equation: Newton plus the material derivative. Difficulty: nonlinearity creates fine scales while diffusion erases them. Obstacle to a simple proof: energy does not control peaks. New claim: a smooth forcing can drive viscous finite-time breakdown. “Explosion”: a local mathematical divergence, with bounded total energy.
Sources and further reading
Current-result claims are linked near the text they support. The toy examples and algebraic exercises are explanatory calculations, not numerical reproductions or independent proof verification. Source status was checked on 8 September 2026.
- David Tong — Fluid Mechanics
Undergraduate lecture notes. Chapters 1–2 are a natural next step for a physics reader. For another derivation, see Tao’s physical introduction.
- Charles Fefferman — Official problem description
The authoritative hypotheses and alternatives A–D, especially pp. 1–2. Its historical discussion predates today’s claims.
- Terence Tao — Weak solutions of the Navier–Stokes equations
Background on energy estimates and what weak existence means. Fefferman’s official description also summarizes the classical partial results.
- OpenAI — On the Navier–Stokes Millennium Prize Problem
8 September 2026. The announcement’s claims, research account and response concerning concurrent work.
- Clay Mathematics Institute — Navier–Stokes Equation
Official status page; displayed “Unsolved” when consulted.
- Clay Mathematics Institute — Prize rules
Acceptance and publication requirements. A same-day announcement is not a prize adjudication.
- OpenAI — NavierStokesAndEuler
Public Lean formalizations, stated theorem scope and checking instructions. Not independently built for this explainer.
- OpenAI — Finite time blowup for Navier–Stokes
Theorem 1.1 and §§2–3 explain the construction; this guide is not a proof audit.
- Terence Tao — September 7 discussion of Alpöge–Buckmaster
Related smooth-forcing results and their connection to earlier Córdoba–Martínez-Zoroa work. Read its equation-specific qualifications carefully.
- Tristan Buckmaster — Accompanying statement
A firsthand account of the separate results and disputed events. Claims concerning other parties are allegations, not findings of this guide.