OpenAI Claims Navier–Stokes Proof - But a Dispute Over Credit Exposes AI’s Compute Power Imbalance

OpenAI claims an unreleased AI model has produced a machine-verified proof for the Navier–Stokes Millennium Prize Problem. But a fierce priority dispute with NYU and Anthropic mathematicians has transformed a historic scientific milestone into a battle over who controls discovery in the era of hyper-scale compute.
The timing of the events is central to the conflict. NYU mathematician Tristan Buckmaster and Anthropic mathematician Levent Alpöge had spent nearly a year developing a relatively unusual strategy for establishing singularity in the Navier–Stokes equations. Their work focused on approaches identified as options C and D in Charles Fefferman’s formulation of the problem, relying extensively on OpenAI’s Codex during their research.
On September 1, OpenAI mobilized a massive computational effort after, by its own account, hearing industry rumours that two Millennium Prize Problems had been solved. Just days later, it published its report, On the Navier–Stokes Millennium Prize Problem.
The resulting race exposes a structural crisis that extends well beyond a single equation: in frontier science, the researchers who find the conceptual route may no longer be the actors with the compute power required to finish the journey.
The mathematics is difficult. The power imbalance is harder
Navier–Stokes equations have been used for more than a century to describe the behaviour of fluids, from water and air to ocean currents and atmospheric systems. Their practical importance has never been in doubt. The unresolved question is whether physically reasonable, smooth initial conditions necessarily remain smooth indefinitely, or whether the equations can produce a singularity in finite time.
That question became one of the seven Millennium Prize Problems established by the Clay Mathematics Institute in 2000, with a $ 1 million prize for a satisfactory solution.
OpenAI's reported approach is notable not simply because it claims to have resolved the problem, but because of the industrial machinery deployed behind it. About 10,000 AI agents reportedly operated simultaneously during an approximately 88-hour run. They exchanged 2.7mn messages and generated roughly 130bn output tokens, with broader project figures reaching as high as 300bn. OpenAI valued the corresponding Astra compute at about $22.5mn.
The initial mathematical generation was followed by another 17 hours devoted to formalisation and machine verification in Lean, a proof assistant capable of checking each formal deductive step.
That changes the economics of mathematical research.
A university researcher can spend months identifying a promising conjecture or proof strategy. A technology company can potentially respond by deploying thousands of agents and tens of millions of dollars in compute almost immediately. Once the difficult conceptual bottleneck has been breached, the cost and speed of completing the work may become radically different for the two sides.
The traditional hierarchy of mathematical labour - intuition first, painstaking formalisation second - is therefore being disrupted by capital-intensive automation.
A priority dispute with unusual stakes
Buckmaster and Alpöge's work matters to the controversy because OpenAI did not begin its effort in a vacuum. The company acknowledged that rumours of academic progress prompted its September 1 mobilisation.
The mathematicians subsequently learned that information about their research had reached OpenAI. Buckmaster contacted the company to ask about the timing of its work and the extent of human involvement, but says he received evasive responses.
The dispute then became more serious. Buckmaster alleges that OpenAI mathematician Sébastien Bubeck sought a compromise under which Alpöge, whose professional affiliation is with Anthropic, would be removed from the credit. Buckmaster says that when he threatened to make the disagreement public, Bubeck responded with warnings about the effect on his career and said he would no longer be "nice" if Buckmaster did not cooperate.
OpenAI disputes the allegations. The company has categorically denied accessing Buckmaster's private data and says that its researchers and agents did not access specific user data before the work was publicly released. It has nevertheless acknowledged that it cannot completely exclude the possibility that de-identified telemetry or data derived from product use contributed to the capabilities of its underlying models.
That distinction is likely to become increasingly important.
There is no allegation in the briefing that OpenAI simply copied a private proof. OpenAI also says its mathematical proof differs substantially from Buckmaster's approach. The harder question is whether information generated during private scientific work can indirectly influence a commercial model's capabilities, and where responsibility lies when a user employs the same company's tools to conduct unpublished research.
Codex turns a research tool into a potential conflict of interest
For academics, AI coding and reasoning systems are becoming ordinary research instruments. Yet the relationship is fundamentally different when the provider of the instrument is also competing to make the scientific breakthrough.
A mathematician can reasonably regard work conducted inside a commercial development environment as part of an unpublished research process. The AI company, meanwhile, operates the infrastructure, controls the model, and potentially collects telemetry associated with product use. Even where individual user data is not directly exposed to researchers, the boundary between product interaction and corporate intelligence becomes difficult to explain from the user's perspective.
That creates a new category of research risk.
The issue is not confined to mathematics. If an academic researcher uses a frontier model while developing a new algorithm, scientific method or engineering concept, the commercial provider may possess an informational position unavailable to competing universities.
The Navier–Stokes episode makes that asymmetry unusually visible because the company and the researchers appear to have converged on the same frontier problem within days.
The presence of Alpöge adds another layer. His affiliation with Anthropic placed the work across the boundary between two competing AI companies. For OpenAI, claiming independent priority would have strategic value beyond academic recognition. For researchers, collaboration across institutional lines can create opportunities for expertise and resources - but also potentially expose them to corporate conflicts that traditional academic norms were never designed to manage.
Lean may settle the proof, but not the priority
There is an important distinction between proving a mathematical proposition and establishing that a scientific community recognises a particular solution.
Formal verification addresses the first problem. A proof encoded in Lean and accepted without errors provides an extraordinarily powerful check that the formal logical deductions are valid. It can eliminate entire classes of mistakes that might survive conventional manuscript review.
But formal correctness does not automatically settle questions of discovery, attribution or priority.
The Clay Mathematics Institute's process still requires publication in a major peer-reviewed journal followed by a two-year period in which the solution is scrutinised by the mathematical community before a prize can be awarded.
OpenAI has published its proof and Lean repository for peer review, but says it will not seek the $1mn prize. The decision removes the obvious financial incentive from the dispute while leaving the more consequential question intact: who gets recognised for discovering the route to the result?
That question becomes particularly awkward when an AI system can generate and formalise enormous volumes of mathematical material faster than humans can evaluate it.
Only one of the Millennium Prize Problems had previously been solved: Grigori Perelman's work on the Poincaré Conjecture in 2002–03. That episode belonged to an intellectual economy in which the mathematician, rather than the computing infrastructure, remained the unmistakable centre of the discovery.
The new model is different.
The arrival of compute-snatching
The emerging danger for independent researchers is not necessarily that AI will replace mathematicians. It is that mathematicians may still be needed to identify the ideas, while becoming unable to compete for ownership of the final result.
Buckmaster and Alpöge's year-long effort illustrates the value of conceptual research into a less common route to singularity. OpenAI's subsequent deployment illustrates the value of scale.
If the decisive breakthrough is identifying an obscure but viable strategy, while the remaining task can be attacked by thousands of coordinated agents, the economic advantage moves sharply towards whoever controls the compute.
This creates what could become a form of "compute-snatching": researchers absorb the intellectual and career risk of exploring an idea, while a company with vastly greater resources can move faster once the direction becomes visible.
Formal proof systems may intensify that shift rather than prevent it. Lean, Coq and Isabelle can provide exceptionally strong verification, but verification itself becomes another stage that can be industrialised.
The result is a new scientific bottleneck. It may no longer be enough to have the best idea. Researchers may need to protect the idea until they can afford to prove it.
That has implications for universities and research institutions already encouraging the use of commercial AI systems. Safeguards around unpublished work, telemetry, model training, and research confidentiality were designed for a world in which the provider of a tool was not necessarily a direct competitor for the discovery being made with it.
Frontier AI has collapsed that separation.
The most consequential question raised by OpenAI's Navier–Stokes claim may therefore have little to do with whether machines can do elite mathematics. If the proof survives peer review, that question will have a powerful answer.
The harder question is whether a scientific system built around open collaboration can remain genuinely open when one class of participant controls the models, the developer infrastructure, the verification machinery - and enough compute to turn a rumour about someone else's idea into an 88-hour research operation.
Sources: Techcrunch, Xeno Spectrum, OpenAI.com
