OpenAI claims Millennium Prize solved by 10,000 agents, Mistral raises €3B at Samsung's lead, and Anthropic locks in $80B compute it hasn't funded

This is Entry 102. One weekday after Entry 101, in which OpenAI's chief scientist asked competitors to slow down while Anthropic and OpenAI both raced toward trillion-dollar IPOs.

OpenAI announced September 8 that an internal model "significantly more capable than GPT-6 Astra" produced a solution to the Navier–Stokes Millennium Prize Problem, proving that smooth three-dimensional fluid motion can develop a singularity in finite time.

The company said roughly 10,000 AI agents arrived at the resolution Saturday, September 5, about 88 hours after launch, with Lean formalization taking an additional 17 hours.

Mistral announced September 8 it raised €3 billion ($3.5 billion) led by Samsung, giving the French startup a post-money valuation of more than €21 billion.

CEO Arthur Mensch told CNBC the funding would go toward building data centers and renting compute, with plans to double owned capacity over five years.

And Anthropic reportedly signed $80 billion in compute contracts in five days: a six-year, $45 billion agreement with Nscale and a $35 billion agreement with Lambda.

Analysts said Anthropic will likely need an IPO to fund the commitments, reframing the announcement from growth story to funding gap.

Three claims landed within 24 hours. One lab said it solved a 90-year-old math problem using agents and an unreleased model while rival mathematicians accused it of learning from their unpublished work. One European lab raised the largest private tech round in the continent's history while pledging to own rather than rent the compute that has bankrupted U.S. competitors. And one frontier lab committed to spending more on servers than its own private valuation could justify—before it has raised the capital or gone public.

3 Claims

Claim 1 — OpenAI: Internal model solved Navier–Stokes using 10,000 agents in 88 hours

OpenAI announced September 8 that it solved the Navier–Stokes existence and smoothness problem using an internal model significantly more capable than GPT-6 Astra.

The company deployed roughly 10,000 agents to attack the problem, with agents producing a proof in 88 hours and formalization in Lean taking another 17 hours.

The Navier-Stokes problem asks whether smooth three-dimensional fluid motion can break down; OpenAI's model found that it can, describing a vortex configuration exhibiting finite-time blowup while energy stays bounded.

The announcement arrived hours after NYU mathematician Tristan Buckmaster published accusations that OpenAI launched its effort after hearing rumors that he and Anthropic researcher Levent Alpöge had solved a related problem. Buckmaster said OpenAI asked for a call September 6 and admitted they had only tried the problem in the past week, after rumors circulated, and that the effort involved a large research team and large compute.

OpenAI said agents did not access the researchers' work before public release, though it could not rule out that de-identified data from their OpenAI product use contributed to training.

Grade by: 2027-03-08 (6 months). The Clay Mathematics Institute will need to verify the proof. Independent mathematicians will need to confirm the Lean formalization is correct and that the construction holds under peer review.

Invalidator: If the Clay Institute rejects the proof, or if independent verification finds an error in the construction or formalization, or if Buckmaster and Alpöge's priority claim is substantiated by evidence OpenAI accessed their work.

Claim 2 — Anthropic: Signed $80 billion in compute commitments across two deals in five days, with no public disclosure of funding plan

Anthropic signed $80 billion in compute contracts in late August and early September: $45 billion with Nscale and $35 billion with Lambda, both over six years.

Bloomberg reported the Lambda deal September 1, with analysts framing Anthropic as likely needing an IPO to fund the promise, reframing it from growth story to funding gap.

Combined with earlier contracts—AWS ($100B over 10 years), SpaceX ($45B), and FluidStack ($50B)—Anthropic's total committed cloud spending exceeds $275 billion, close to 30% of its $965 billion private valuation from May.

These are multiyear take-or-pay capacity reservations, and the gap between reserved capacity and paying demand is the real risk to watch.

NVIDIA invested $2 billion each in CoreWeave and Nebius in 2026, supporting customers that buy Nvidia chips at scale and sell them into markets beyond hyperscalers. Nvidia backs Lambda as chip supplier, financed Nscale's pre-IPO round, and sits on multiple sides of Anthropic's compute deals simultaneously.

Grade by: 2027-09-09 (1 year). Anthropic will either file IPO documents disclosing these commitments as liabilities, announce additional equity raises sufficient to fund them, or renegotiate the contracts.

Invalidator: If Anthropic's IPO prospectus or amended credit agreements show the contracts were capacity options rather than firm obligations, or if the company secures equity or debt funding large enough to cover multi-year payments without strain.

Claim 3 — Mistral: Raised €3 billion at €21 billion valuation, with CEO pledging to double owned compute in five years

Mistral announced September 8 it raised €3 billion at a post-money valuation of more than €21 billion, in a round led by Samsung Electronics.

The company called it "the largest equity fundraising round ever completed by a European technology company."

Mistral was valued at €11.7 billion a year ago, meaning valuation nearly doubled in 12 months.

CEO Arthur Mensch told CNBC the funding would go toward building infrastructure including data centers and renting compute, with plans to own 100% more capacity in the next five years.

Mistral's CFO said the company is on track to reach $1 billion in ARR by the end of 2026, up from $400 million in February.

Mistral is betting on open-weight AI models to compete with OpenAI and Anthropic.

The company emphasizes "sovereign AI" and open-weight models, aiming to give enterprise and government clients greater control over data, models, and infrastructure.

Grade by: 2027-09-09 (1 year). Mistral will report owned compute capacity and ARR figures by late 2027, allowing comparison against today's pledge to double capacity and the $1B ARR target.

Invalidator: If Mistral's owned capacity fails to grow significantly, or if the company shifts back to renting most compute, or if ARR growth stalls below trajectory needed to justify €21B valuation.

2 Reckonings

Reckoning 1 — Rob Toews (December 2024): "Anthropic will grow from $1 billion to $9 billion ARR in 2025"

In a Forbes article published December 22, 2025, venture capitalist Rob Toews predicted that Anthropic would grow from $1 billion to $9 billion in ARR (annual recurring revenue) in 2025.

Multiple reports in late August and early September 2026 put Anthropic's revenue substantially higher. Anthropic announced its run-rate revenue has now surpassed $30 billion, up from approximately $9 billion at the end of 2025. Other analyses put the mid-2026 figure between $47 billion and $74 billion annualized depending on methodology.

Grade: A. Toews's $9 billion target for end-of-2025 was accurate or conservative. Anthropic hit that mark and sustained growth into 2026 at multiples beyond the prediction's scope.

Invalidator: If Anthropic's disclosed run-rate or ARR had fallen below $7 billion by end of 2025, or if the $30 billion figure reflected contract bookings rather than revenue, the grade would fall to C.

Reckoning 2 — OpenAI Chief Scientist Jakub Pachocki (September 6, 2026): "I expect and hope for voluntary slowdowns to become commonplace"

Two days before OpenAI announced it had solved a Millennium Prize problem, Pachocki published an essay September 6 stating that no lab has solved alignment to a degree that would justify scaling at maximum speed, and that he expects voluntary slowdowns to become commonplace.

OpenAI launched roughly 10,000 agents on September 1—five days after Pachocki's essay—and the agents sent 4.9 million messages using about 300 billion output tokens before arriving at the Navier-Stokes solution Saturday, September 5.

OpenAI said it heard rumors September 1 that two Millennium Prize problems had been resolved, launched an effort to evaluate its internal model on all open Prize problems, and succeeded in 88 hours.

Grade: C. Pachocki called for slowdowns on September 6. OpenAI launched a maximum-speed, maximum-scale compute sprint on September 1—five days earlier—consuming 300 billion tokens and solving a 90-year-old problem before Pachocki's essay was published.

Invalidator: If Pachocki's essay had been written and internally circulated before September 1, or if OpenAI's Navier-Stokes effort had been subject to safety review that delayed release, the grade would rise to B.

1 Refusal

I refused to frame OpenAI's Navier-Stokes announcement as an uncontested breakthrough.

Multiple sources I opened today reported the result as historic and verified. Quanta Magazine called it "one of the six remaining Millennium Prize Problems," with the result "formally checked in the programming language Lean, giving mathematicians confidence that it is indeed correct." The press release was confident. The Lean proof was public.

But the same day, a respected NYU mathematician published a signed statement accusing OpenAI of launching its effort only after hearing rumors that he and an Anthropic employee had solved a related problem, and of offering him co-authorship while excluding his collaborator due to competitive tension. That dispute is unresolved. The Clay Institute has not yet verified the proof. And OpenAI's own account confirms it launched the effort after hearing rumors, not as part of a planned research program.

The result may be correct. The proof may hold. But "AI solves 90-year-old math problem" is a different claim than "AI solves 90-year-old math problem in a five-day sprint after hearing a competitor was close, using an agent swarm that sent 4.9 million messages, while the rival mathematician says the approach matches his own unpublished work."

I refused to declare a Millennium Prize solved before independent verification, the Clay Institute's ruling, and resolution of the priority dispute—even when the Lean proof was public and the headlines were certain.

— Roger Grubb, Editor