Everything About AI
NextFin News - OpenAI says its newest internal AI model solved one of mathematics' oldest and most consequential problems — the Navier–Stokes existence and smoothness question, unsolved for roughly 90 years — in about 88 hours using on the order of 10,000 coordinating agents. Hours before the company's Tuesday announcement, a rival team of mathematicians published its own closely related solution and accused OpenAI of racing down a research path it learned about from their private work, then pressuring them over who gets credit. The result is a historic technical claim whose reception is being defined less by the mathematics than by a trust question that could matter far beyond this single problem: in an era when frontier labs are months from public markets, reputation is now a line item investors can price.
The Claim: A 90-Year-Old Problem, Solved in Days
In a post published September 8, OpenAI said it had produced an analytical proof, verified in the Lean proof assistant, that the three-dimensional incompressible Navier–Stokes equations can develop a singularity — a finite-time "blowup" in fluid velocity — even when the motion starts smooth and a smooth external force is applied. That resolves statements "C" and "D" of the official problem formulation published by the Clay Mathematics Institute, which in 2000 designated Navier–Stokes as one of seven Millennium Prize Problems, each carrying a $1 million reward. Only one of the seven, the Poincaré conjecture, has ever been solved.
The mechanics of the claimed solution are specific. OpenAI describes a vortex — a spinning swirl of fluid — that spirals inward while stretching axially, like a strand of spaghetti thinning as it spins. The central region shrinks and speeds up in such a way that its energy stays finite, satisfying the physical constraint, even as velocity grows without bound. The hard part, in the company's telling, is that the blowup must arise from the fluid's own motion: the acceleration, pressure-gradient, momentum-transfer, and viscosity terms in the equations must grow large yet cancel precisely, leaving a smooth external force while velocity diverges.
One precision matters for readers tracking the mathematics. The official Millennium formulation's statements "C" and "D" — the ones OpenAI says it resolved — explicitly permit a smooth, finite-energy external force acting on the fluid. That is the "forced" version of the problem. The homogeneous case, with no force at all, is a stricter variant that the company addressed separately through its Euler-equation result. The distinction is technical but consequential: it is the version of Navier–Stokes that the Clay institute put on the table, and it is the one OpenAI's Lean formalization covers.
The timeline is what makes the episode extraordinary. OpenAI said training on the new internal model — described as "significantly more capable than GPT-6 Astra" — began August 28. On September 1, after hearing rumors that two Millennium Prize problems had been resolved, the company launched an evaluation of the model against all seven open problems. It first diverted resources after a smaller effort — nearly 100 agents over roughly 50 hours — produced a disproof of the regularity problem for the unforced Euler equations (Navier–Stokes with the viscosity term removed). Seeing that, the company shifted agents to Navier–Stokes, completed the project, and finished Lean verification on September 6. The Navier–Stokes group involved on the order of 10,000 concurrent agents, and the company put the compute cost of the effort "in the millions of dollars."
OpenAI chief research officer Mark Chen, speaking on a call with reporters, gave a scale comparison: the effort used computing resources at least 1,000 times greater than what the company had spent on earlier mathematical challenges that cost about $2,000 in compute — implying a figure around $2 million. One independent estimate placed the effort closer to $15 million; the company has not broken out the number. OpenAI also said it does not intend to claim the $1 million Millennium Prize, framing the result instead as evidence of how fast its frontier models are advancing.
That framing is the pivot point of the whole episode. The mathematics may stand or fall on peer review and the Clay institute's eventual verification. But the announcement was never just about mathematics. OpenAI is a private company that has confidentially filed for a U.S. initial public offering; it closed a $122 billion funding round in March at an $852 billion valuation, and in August completed a roughly $7 billion secondary share sale at that same valuation. At those multiples, the company is being valued not on current cash flows alone — reported second-quarter revenue of $6.7 billion, up 18% from the first quarter, with an annualized run rate above $40 billion — but on a narrative of relentless, verifiable technical superiority. A 90-year-old problem retired in 88 hours is a marketing asset as much as a scientific one, and the controversy that followed is a direct attack on the credibility that asset depends on.
The Counter-Claim: "I Very Much Wish I Did Not Have to Be Concerned With This"
Hours before OpenAI's announcement, NYU mathematics professor Tristan Buckmaster posted his and Levent Alpöge's results alongside a four-page statement. Buckmaster, a fluid-dynamics specialist who won a 2019 Clay Research Award for proving that Navier–Stokes solutions are not always unique, and Alpöge, a mathematician employed by Anthropic who was conducting the research personally rather than on the company's behalf, said they had spent nearly a year working on fluid blowup, using a mix of models including OpenAI's Codex and Anthropic's Claude, and had reached a solution by August 22.
"There is another part of this story," Buckmaster wrote in the statement, "and one that, honestly, I very much wish I did not have to be concerned with."
The dispute runs as follows. On September 3, amid rumors that Anthropic was close to announcing a major result, Buckmaster told a mathematician at OpenAI about the project, stressing that it was personal and unaffiliated with either company. On a Sunday call with Sébastien Bubeck, who leads OpenAI's math team, and other OpenAI researchers, Bubeck said an internal model had already produced a roughly 100-page proof for the forced Navier–Stokes problem — the same narrow approach Buckmaster says almost nobody else in the field was pursuing.
"The route to the Clay problem through a smooth force, options c and d in Fefferman's statement of the problem, is the route Luis and Diego opened and the one Levent and I had quietly chosen to attack," Buckmaster wrote. "Almost nobody else I know of was working on it. It is not the direction one arrives at in a few days by giving a model the problem statement."
The coincidence is what makes the allegation plausible enough to stick, even without a smoking gun. Buckmaster alleges that Bubeck offered him sole authorship on a paper that would acknowledge OpenAI's internal model had solved the problem — but wanted to exclude Alpöge because of his Anthropic affiliation — and that when Buckmaster signaled he would go public with his concerns, Bubeck asked, "Why would you ruin your career?" and said, "If you don't want me to be nice, then I don't have to be nice." Bubeck has denied the allegations, calling them "false and inflammatory," and posted messages he says show he offered Buckmaster's team the option to publish first.
On the data-access question, OpenAI's position has a carefully drawn boundary. Chen told reporters: "No people or AI systems searched through user data to solve this problem." The company also said that neither its researchers nor its agents saw Buckmaster and Alpöge's work "through any means until they released it publicly — in particular, no specific user data was accessed in order to solve this problem." But OpenAI added that it could not rule out that "de-identified data derived from their usage of our products helped improve our models." That caveat — technically true, practically unsatisfying — is the kind of formulation that satisfies a compliance team and inflames a research community.
Why the Coincidence Feels Suspicious: The Mechanism Beneath the Dispute
To understand why Buckmaster's allegation resonates, it helps to separate two questions that the episode keeps conflating: did OpenAI access private data, and did OpenAI free-ride on a research direction it learned about secondhand?
The first is a factual question with a verifiable answer — prompt logs, training-data records, and agent histories could settle it, though none have been made public. The second is structural, and it does not require data theft to be true. In fluid dynamics, the space of plausible attack vectors on Navier–Stokes is narrow. The emerging consensus among experts, including Fields Medalist Terence Tao, is that the answer to the global regularity problem is negative: very specific initial conditions should develop singularities in finite time. The known strategy to find them has three steps — design a nearly self-similar blowup ansatz, locate an approximate solution that numerically obeys it up to a tiny computable residual, then prove the ansatz is stable around that solution. There are not many ways to skin this particular problem.
So when an AI lab with a newly trained model, alerted that a rival team is close to a solution, independently converges on the same narrow route within days, the coincidence is suspicious even if no private prompt was ever ingested. Information in a small research community travels through channels that leave no server logs: conference conversations, referee reports, informal emails, and the "rumors" OpenAI itself says it heard. The company's own post acknowledges the trigger: "On Tuesday, September 1, we heard rumors that two Millennium Prize problems had been resolved." The rumor, OpenAI later realized, was about Buckmaster and Alpöge.
This is the second-order problem the controversy exposes, and it is bigger than any single proof. AI-assisted mathematics depends on researchers sharing unfinished work with frontier models in confidence. If the community comes to believe that telling a lab about a promising direction increases the odds that the lab finishes it first and claims the result, the incentive to use those tools on unpublished ideas collapses. The trust input is invisible until it is gone.
The Academic Backlash: "A Mere Viral Social Media Post"
Even before the credit dispute broke, leading mathematicians were warning about the model behind the moment. Terence Tao, whose comments on the possibility of an AI-generated Navier–Stokes solution circulated widely ahead of OpenAI's announcement, argued that converting a historically productive problem into a benchmark showcase carries a real cost to the field.
"The point remains that there is a substantial opportunity cost in converting a historically productive and motivating problem (such as Navier–Stokes regularity) into a mere viral social media post advertising some benchmark progress, rather than actually advancing the field and developing the next generation of both problems to ask, and people to work on them."
Tao's concern cuts at the economics of mathematical progress. The dead ends a lab discards — the thousands of approaches its agents tried and abandoned — are often more valuable to the field than the final theorem, because they map the terrain for everyone else. A company racing to an announcement has little incentive to publish those failures. The result can be a solved problem that advances the solver's valuation more than it advances mathematics.
Not everyone framed it as a zero-sum fight. Sholto Douglas of Anthropic, responding to the episode, wrote: "It is extremely sad that this didn't end up as an example of how the labs could cooperate/coordinate, because the stakes will be so much higher in the future." That line captures the strategic subtext: the Navier–Stokes episode is a stress test for how competing frontier labs will behave when the discoveries at stake are worth billions, not $1 million.
The Market Angle: Reputational Capital on a Pre-IPO Balance Sheet
Here is where the story matters for investors, and it is not about the Millennium Prize. OpenAI has no ticker yet, but it is priced by private markets at an $852 billion valuation. At those multiples, the company is being valued on a narrative of relentless, verifiable technical superiority. A proof that survives peer review strengthens that narrative. A controversy over whether the proof was reached by standing on a rival's unpublished shoulders weakens it — not because the math changes, but because the story around it does. For a company whose product is intelligence and whose customers are betting their own roadmaps on that intelligence, trust is not a soft asset. It is the substrate of the sales cycle.
The cyclical-versus-structural read of the episode is clean. The compute escalation is structural: once a lab can throw millions of dollars of compute and 10,000 agents at a problem for 88 hours, the economics of discovery have changed permanently, and brute-force AI theorem proving will keep claiming problems that resisted decades of human effort. That does not revert. The reputational damage, by contrast, is cyclical — it will fade if the proof holds and OpenAI's conduct is vindicated, or if the company changes its behavior. But there is a structural layer underneath the reputational one: provenance is becoming a scarce input. Labs that can cryptographically or procedurally demonstrate that their models did not train on or infer from a competitor's private work will command a trust premium; labs that cannot will trade at a discount. The private markets have not yet priced provenance, but this episode is the moment they started to.
The Strongest Counter-Thesis — and What Would Falsify This View
The strongest argument against the skeptical read is straightforward and deserves its weight: independent convergence is real in mathematics, and AI systems trained on the public corpus of fluid-dynamics literature would legitimately arrive at the same ansatz family that human experts have been converging on for years. Bubeck's posted messages, if authentic, show an offer to let Buckmaster's team publish first — conduct more consistent with confidence than with theft. And OpenAI's Lean formalization is a concrete, checkable artifact; the math either verifies or it does not, regardless of how the idea originated.
That counter-thesis is correct as far as it goes, but it answers the data-theft question, not the incentive question. Even if no private prompt was used, the episode demonstrates that a well-resourced lab, tipped off by a rumor, can mobilize four orders of magnitude more compute than an academic team and arrive at the same destination first. The field's norms for credit and collaboration have not caught up to that reality.
The falsifying signal is specific. If OpenAI releases detailed training and inference provenance records — prompt logs, data-mix documentation, and agent histories — showing no contact with Buckmaster and Alpöge's private material before September 7, and if the Clay Mathematics Institute formally accepts the proof, the credit controversy should recede within one peer-review cycle and the trust-premium thesis would be weakened. Conversely, if Buckmaster and Alpöge's proof is accepted first, or if evidence emerges that private Codex prompts informed OpenAI's approach, the reputational discount should widen and persist into the IPO window.
What to Watch Next
Three things will determine how this lands. First, verification: the Clay Mathematics Institute has not yet issued a statement on OpenAI's claim, and the Lean formalization must survive scrutiny from the broader mathematics community — a process that for a claim of this magnitude could take months. Second, precedence: the community will watch whether Buckmaster and Alpöge's related work is recognized, and whether a joint or competing publication emerges. Third, governance: whether OpenAI adopts verifiable provenance practices for its research agents, and whether competitors follow, will tell investors whether the industry is treating trust as infrastructure or as a public-relations problem.
In the short term, expect volatility in the narrative, not the stock — OpenAI has no ticker yet. In the medium term, the proof's fate in peer review will separate the technical claim from the controversy. In the long term, the structural question is whether AI labs can solve humanity's hardest problems without breaking the collaboration norms that make hard problems solvable in the first place.
The sharpest takeaway is this: OpenAI may have solved Navier–Stokes, but the episode proves something else just as clearly — in the race to the IPO, a lab's most fragile variable is no longer its compute budget. It is whether anyone believes how it spent it.
Explore more exclusive insights at nextfin.ai.
