On September 8, 2026, OpenAI announced a Lean-formalized proof of finite-time blow-up for the 3D incompressible Navier-Stokes equations with a smooth, compactly supported external force. The construction produces a self-similarly concentrating vortex that starts from rest, maintains uniformly bounded kinetic energy, and sends the velocity to infinity as t approaches a finite time T. This resolves Statements C and D of the Clay Millennium Prize formulation: breakdown of smooth solutions in the presence of a smooth force.
Statements A and B remain open. These ask whether smooth solutions exist globally when f = 0, starting from smooth, rapidly decaying, divergence-free initial data with finite energy. The question this document addresses: does the OpenAI construction contain a path from C/D to A/B?
The force f in the OpenAI construction is smooth and compactly supported in space and time. Its temporal support lies within [0, T) where T is the blow-up time. The force does not shut off at some earlier time and leave the solution to evolve freely. It is active throughout [0, T), continuously maintaining the self-similar collapse profile against viscous dissipation.
However, the construction achieves a critical property: the residual f and all its derivatives vanish faster than any power of the concentration parameter q as the singularity is approached. The force is not merely smooth at T; it is flat at T. Every derivative goes to zero as t approaches T. The force can be smoothly extended to be identically zero at and after T.
This flatness is the opening. A force that vanishes with all derivatives is, in a precise functional-analytic sense, the smallest possible nonzero perturbation. The question becomes: is a flat perturbation removable?
The naive argument proceeds: at any time before T, the solution is smooth, compactly supported, divergence-free, and has finite energy. If f were zero from that point onward, the solution would serve as initial data for unforced NS. If it still blows up at T, Statement A is resolved.
This fails because f is NOT zero on any subinterval before T. The force is active everywhere on [0, T). At every point, the remaining force is nonzero, however small. One cannot simply declare the force irrelevant because it is small. In a supercritical PDE, arbitrarily small perturbations can determine whether blow-up occurs. The NS equations are supercritical in 3D precisely because the scaling symmetry does not preserve the energy. Small forces at small scales carry disproportionate influence.
The constructive approach: take the blow-up solution at some time close to T as initial data for unforced NS. The perturbation w between the unforced and forced solutions satisfies a linear equation with growing coefficients (from the blow-up profile) and a flat source (from the force).
The forcing term on the right is -f, which by the flatness property is smaller than any power of (T - t). The linear operator has coefficients that GROW due to the blow-up profile. The question reduces to: does the decaying force dominate the growing linear operator, or vice versa?
In the OpenAI construction, the blow-up profile has velocity scaling with a small exponent h (0 < h < 1/100). The background shear scales accordingly. The force residual is flat: bounded by any power of (T-t).
Standard estimates for linear transport-diffusion equations with growing coefficients give a bound involving an exponential growth factor times a flat decay integral. For h < 1/100, the exponential grows slower than any inverse power of (T-t). The product of exponential growth and flat decay goes to zero.
This says the perturbation w STAYS SMALL. If removing the flat force produces a perturbation that remains bounded as t approaches T, then the unforced solution tracks the forced blow-up solution and also blows up.
The estimate in Section 5 is too crude. The Gronwall-type bound treats the linearized operator as a scalar. The actual operator has directional structure. The blow-up is axisymmetric with azimuthal velocity dominating. Perturbations in different directions experience different amplification rates.
The construction has two components: the background vortex (which concentrates) and the oscillatory pulses (which provide Reynolds stress). The force corrects the residual between these two components. Removing the force means the two components evolve under their mutual nonlinear interaction alone. The flat force was compensating for a MISMATCH between what the nonlinear term produces and what the self-similar profile requires. Without it, the profile drifts.
Whether this drift stabilizes or destabilizes the singularity depends on the sign of the linearized eigenvalues of the self-similar ODE around the blow-up profile. This is a spectral problem for the linearized self-similar operator.
The OpenAI blow-up profile is a stationary point of a renormalized evolution equation in self-similar coordinates. Linearizing around the stationary point gives a linear operator L whose spectrum determines stability.
If L has no unstable eigenvalues (all eigenvalues have negative real part): the blow-up profile is an attractor. Nearby unforced solutions are pulled toward it and also blow up. Removing the flat force produces a perturbation that decays in self-similar variables. Statement A falls.
If L has an unstable eigenvalue: the profile is a saddle point in function space. The forced solution sits on the stable manifold (the force holds it there). Remove the force, and the solution falls off the stable manifold and the singularity may not persist. Statements A and B remain open.
The OpenAI construction does not compute or discuss the spectrum of L. The proof works by construction rather than by stability analysis. This is the gap.
For the inviscid Euler equations, self-similar blow-up stability has been studied extensively. Merle, Raphael, Rodnianski, and Szeftel proved stable self-similar blow-up for compressible Euler. For Navier-Stokes specifically, Necas, Ruzicka, and Sverak proved there are no nontrivial self-similar blow-up solutions with finite energy in the classical scaling. The OpenAI construction evades this by using a different scaling with the small exponent h.
Tao's 2016 averaged NS blow-up used a similar strategy: construct a blow-up for a modified system, then ask whether it transfers. The transfer failed for averaged-to-true NS because the nonlinear term's structure changed. Here, the transfer would be forced-to-unforced, where the equation structure is identical but the source term changes. This is a strictly easier perturbation than Tao's case.
Resolving Statements A and B via the OpenAI construction requires one result: compute the spectrum of the linearized self-similar operator L around the OpenAI blow-up profile in the modified scaling.
If the spectrum is stable: write the perturbation equation in self-similar variables. The flat force maps to a source that decays exponentially in self-similar time. A decaying source with a stable linear operator gives a decaying perturbation. The unforced solution tracks the blow-up profile. Statement A falls.
If the spectrum is unstable: characterize the unstable manifold. If it has finite codimension, blow-up still occurs from a codimension-k set of initial data. This would give a weaker result: specific blow-up exists but generic blow-up fails.
The spectral computation is concrete, well-posed, and can be performed numerically on OpenAI's explicitly constructed profile, or analytically via the known asymptotic behavior of the profile.
Output type: Proof strategy. Outcome: PARTIAL. The forced-to-unforced reduction is identified as a spectral stability problem for the self-similar operator around the OpenAI blow-up profile. The flatness of the force makes this a best-case scenario for perturbation theory. What prevents completion is that the spectrum of L has not been computed. If it were computed and found to be stable, the argument would close to a proof of unforced blow-up.
[1] OpenAI. "Finite Time Blowup for Navier-Stokes." September 8, 2026. link
[2] Fefferman, C. "Existence and Smoothness of the Navier-Stokes Equation." Clay Mathematics Institute, 2000. link
[3] Merle, F., Raphael, P., Rodnianski, I., Szeftel, J. "On the implosion of a three dimensional compressible fluid." Annals of Mathematics, 2022.
[4] Elgindi, T. "Finite-time singularity formation for strong solutions to the axi-symmetric 3D Euler equations." Annals of PDE, 2021.
[5] Necas, J., Ruzicka, M., Sverak, V. "On Leray's self-similar solutions of the Navier-Stokes equations." Acta Mathematica, 1996.
[6] Tao, T. "Finite time blowup for an averaged three-dimensional Navier-Stokes equation." JAMS, 2016.
[7] Ultrathink. "OpenAI Says 10,000 Agents Found a Finite-Time Navier-Stokes Blowup." September 2026. link
[8] CellCog. "Navier-Stokes: OpenAI's 10,000-Agent Proof and the Dispute." September 2026. link
[9] Jones, N. "AI cracked the Navier-Stokes challenge." Nature, September 18, 2026. link
[10] Castelvecchi, D. "OpenAI claims huge maths breakthrough." Nature, September 8, 2026. link