Leadership2 publishers3 min readPublished
OpenAI's Navier-Stokes claim sparks dispute over whether mathematicians' data was accessed
OpenAI says no specific user data was accessed to solve the problem, and that it cannot rule out that de-identified data derived from two mathematicians' use of its products improved its models. A procurement team has to read both sentences.
The Board Room · Leadership desk

What happened
- OpenAI said on Tuesday that it had solved the Navier-Stokes Millennium Prize Problem, a day after Buckmaster and Alpoge posted related results of their own.
- A further concern Buckmaster raised is that the pair's work in progress was stored in OpenAI's Codex model, which he says potentially made it visible to the OpenAI team.
- Bubeck, at an OpenAI press briefing on Tuesday, denied that the company had used the pair's work or accessed material shared with OpenAI's servers.
- OpenAI says it spent millions of dollars on the effort, run on an internal system it describes as more powerful than its latest GPT-6 Astra model.
Compiled by The Board RoomSomething wrong?How this is made
Why it matters
- contradiction Buckmaster says Bubeck wanted Alpoge dropped as an author because he works at Anthropic; Bubeck says he never asked for that. Two named parties, no adjudicator, and a buyer assessing vendor conduct has only their competing accounts to weigh.
- exposure The exposure in the record is disclosure of timing, not a database read: a researcher mentioned a publication date to vendor staff. Contract language about training data does not reach conversations of that kind.
- constraint Any assurance a customer extracts that their content is not trained on now has to be tested against the de-identified-derivative residue OpenAI named itself, which limits how much comfort a standard clause can offer teams doing pre-publication work.
- decision Groups whose value rests on being first, and who used both OpenAI and Anthropic tools as Buckmaster and Alpoge did, now have to decide whether to keep in-progress work inside a tool sold by a vendor pursuing the same result.
The scope of OpenAI's two statements is where the procurement content sits. The company says that no specific user data was accessed in order to solve this problem, and in the same document that it cannot rule out that de-identified data derived from the pair's usage of its products helped improve its models [6]. One sentence covers a single problem and a single access path; the other covers the general capability that produced the result, and it is the vendor's own wording rather than an allegation. A buyer who asks only whether their chats were read gets a true answer to a narrower question than the one they meant to ask.
In both accounts, the information that moved first moved through people. Buckmaster says that after he told an OpenAI mathematician he and Alpoge planned to post their work, OpenAI staff including Sebastien Bubeck told him an internal model had produced a proof [5]. Bubeck says he was trying to coordinate the two releases, and that OpenAI took the problem up after viral rumours that Anthropic had resolved two Millennium problems [10]; OpenAI's own announcement gives the same origin story, citing rumours that two of the problems had been solved [23]. Both versions have OpenAI's timing responding to news of someone else's unpublished progress, arriving as talk. No retention clause governs that channel.
OpenAI's account of the run is about 10,000 agents working at once, 88 hours to a solution, and about 17 hours for GPT-6 Astra to verify it [13][14], which works out to roughly 880,000 agent-hours [17]. The Clay Mathematics Institute prize is $1m, and OpenAI says it does not intend to claim it [15]; set against a flotation that could value the company at around $1tn [16], the prize is about one part in a million of the valuation [18]. The return on the exercise is the announcement, which is why release timing rather than prize money is what the two accounts disagree about.
A skeptic will say this is a credit fight among mathematicians with no bearing on a company buying coding seats. The reply is to ask who else could occupy Buckmaster's position: any customer whose in-progress work is the asset, using a tool sold by a vendor that competes for the same result. Buckmaster is careful about the limits of his own account, writing that he has not seen OpenAI's proof, does not know whether their data was used, and is not accusing anyone [8]. Those questions stay open for a buyer as well, which is the practical difficulty: the contract has to be written against uncertainty that will not be settled before it is signed.
The trade is frontier capability against a residual the vendor itself says it cannot bound [6]. This quarter that falls to whoever owns the renewal, and the questions are what is retained, what "de-identified" and "derived" are taken to cover, whether a no-retention tier exists for teams doing pre-publication work, and which of the vendor's staff may hear about that work. Next quarter it falls to the research or product lead deciding whether the fastest available tool is the right one for work whose value depends on being first. Those two decisions usually sit with different people, and the paperwork is written before the discovery, not after.
What to watch
- Whether OpenAI publishes the proof and a provenance record that Buckmaster or the Clay Mathematics Institute can independently check.
- Whether enterprise agreements start naming de-identified derived data explicitly, rather than only promising that customer content is not trained on.
- Whether either side produces a record of the call in which Buckmaster says authorship and release timing were discussed.