10,000 Agents, 88 Hours, One Millennium Problem

Yesterday Europe wrote a 3 billion euro check. This morning OpenAI published a Lean file that says a smooth fluid can blow up.

OpenAI’s writeup is dated September 8. An internal model significantly more capable than GPT-6 Astra — still training, not for sale — ran a swarm of about 10,000 concurrent agents for 88 hours. The agents produced an analytical proof that the three-dimensional incompressible Navier–Stokes equations can develop a singularity in finite time, with finite energy, from rest, under a smooth force. GPT-6 Astra then spent 17 hours turning that argument into Lean. The company says this establishes statements C and D in the Clay Mathematics Institute’s official formulation. It also says, in the same post, that it does not intend to claim the million-dollar prize.

That last sentence is doing more work than the announcement. A lab that had unambiguously closed a 90-year problem would take the check. OpenAI published the PDF and a public Lean repo instead. The kernel can check the kernel. That is the actual product.

The vortex that looks like spaghetti

The Clay problem is not “write down the Navier–Stokes equations.” Those have been on the wall since the nineteenth century. The question is whether a smooth 3D incompressible flow with viscosity can still develop unbounded speed in finite time. Viscosity is supposed to smear things out. Jean Leray proved weak solutions exist in 1934. Smoothness stayed open. Clay put a million dollars on it in 2000.

OpenAI’s construction is a vortex: a spinning swirl that spirals inward and stretches axially, “like spaghetti.” The core shrinks and speeds up while energy stays finite. The motion terms — acceleration, pressure, momentum transfer, viscosity — get large and cancel in a way that leaves a smooth external force even as velocity blows up. Because a real fluid cannot move infinitely fast, the result is a statement about the equations, not about water. Ven Chandrasekaran’s briefing line, as Nature has it: there exist fluids that start perfectly normal and, under these equations, achieve infinite speed in finite time.

Clay president Martin Bridson called it “an exciting day.” Princeton’s Charles Fefferman, who wrote Clay’s official problem statement, told Quanta he was thrilled the problem was solved — and that the heroes of the story are Diego Córdoba and Luis Martínez-Zoroa, whose strategy both AI-enabled teams followed. Martínez-Zoroa, asked by Nature, called the result “truly remarkable.”

The token bill is the other number.

If you are going to stare at a 130-billion-token math run on a box you actually own, 64GB of DDR5 and a 2TB NVMe still matter more than another SaaS seat. Lean does not care about your cloud invoice. It cares whether the file builds.

Priority, rumors, and a different Euler

OpenAI says it started training the new internal model on August 28. On September 1 it heard rumors that two Millennium problems had been resolved, pointed the swarm at all of them, and let the agents surprise it with an unforced Euler blowup first. That is the viscosity-zero cousin. Then it shifted the swarm onto Navier–Stokes and cross-pollinated groups with Codex.

The rumor was Levent Alpöge (Anthropic) and Tristan Buckmaster (NYU). After Lean verification on September 6, OpenAI reached out for a concurrent release. They had a forced Euler result, not Navier–Stokes. OpenAI congratulates them, recognizes their priority on forced Euler, and says the researchers and the agents did not see their work until it was public. It cannot rule out that de-identified product usage helped train the model. The proofs, it says, differ, and even the Euler statements differ (forced vs unforced).

Buckmaster, in Quanta’s telling, called the week a Deep Blue–Kasparov moment and argued Martínez-Zoroa deserves a Fields Medal. That is the grown-up version of the credit fight. OpenAI ran 10,000 agents at a rumor. Two humans with a stack of models published twelve hours earlier on a related problem. Clay has not awarded anything. The Lean repo is Apache-2.0 and uses Lean 4.34.0-rc2 plus Mathlib. Anyone with lake build can try to break it.

This is a different object from last week’s Fermat autoformalization. That was 13 million lines of Lean reconstructing a known theorem. This is a claimed resolution of an open Millennium problem, machine-checked, by a model that is not on the price list. Treat the kernel check as the evidence. Treat the prize as not claimed for a reason.

Distillation as a named campaign

While the math internet argued about vortices, three US agencies put names on a training method.

CISA, NSA, and the FBI published joint advisory AA26-251A on September 8. The claim: DeepSeek, Moonshot AI, Alibaba, MiniMax, StepFun, and Z.AI have run industrial-scale knowledge distillation against US frontier models — Claude, GPT, Gemini, Grok variants — since at least late 2024. Billions of tokens. Millions of exchanges. “The core, not merely a supplement” of how those companies build. Likely with Chinese government awareness.

Distillation is a real technique. Train a cheaper model on a stronger model’s outputs. The advisory’s argument is volume, targeting, ToS violation, and evasion: native APIs, cloud providers, aggregators, account pools, gray-market “transfer stations.” DeepSeek’s famous 5.6 million training-cost figure, the agencies say, excludes the cost of data acquired this way. Moonshot is accused of pulling Claude Fable 5 data into Kimi-K3. Alibaba of improving Qwen off Claude and GPT. Z.AI of billions of tokens from GPT-5.5 and Claude Opus by mid-2026. MiniMax of trying prompt injection against Claude Code.

China hit back Wednesday ahead of planned Trump–Xi talks. The document is an advisory, not an indictment. The named firms have not been convicted of anything in a court. What changed is that three letter agencies put the $5.6 million talking point in a PDF and told US labs to poison outputs when they suspect a siphon.

If your agents have API keys sitting in a repo, hardware keys beat a reused password, and an AIO cooler on the box you actually patch is more useful than a new policy PDF.

Qualcomm sold Amazon a decade of inference

The silicon print from Tuesday is the other stack story.

Qualcomm and AWS announced a multi-generation collaboration on custom inference silicon and optical connectivity up to 1.6T. The 8-K is the part that matters: Qualcomm issued Amazon a warrant for 25 million shares at 161.26 apiece, about 4 billion of notional equity, vesting against commercial milestones and up to 60 billion in purchases of server chips, technology, systems, and manufacturing services through September 3, 2036. 3.75 million shares vested on issuance against initial commitments.

This is Qualcomm’s first Western hyperscaler for custom AI inference silicon, after Meta’s Dragonfly C1000 commitment for 2028. Nvidia still owns training. Inference is where the phone-chip company thinks it can live. Same week, OpenAI’s Korea GM told a Seoul presser the company is deepening next-generation chip work with Samsung — the same Samsung that led yesterday’s Mistral round — without giving a node, a wafer start, or a date. Samsung would not confirm customer details. Jalapeño with Broadcom, TSMC to fab, remains the named inference part.

A 10,000-agent math swarm, a distillation advisory with six company names, and a 60-billion purchase ceiling on inference silicon. One of those is a Lean file you can build. One is a PDF from CISA. One is an 8-K. The unreleased model is still training. The prize is still unclaimed. The vortex is still spaghetti.

Want this in your inbox every morning? Subscribe to the SpaghettiStories newsletter.

Some links may be affiliate links. If you’re buying hardware to run local models, this affiliate link helps keep the lights on.