Claude helps settle a 78-year-old question about the six-sphere
A question that has stumped mathematicians since 1948 — and tripped up some of the field's most decorated names — appears to have finally fallen, with Claude helping a Harvard mathematician build the answer rather than search for it. In a year already full of AI math milestones, this one looks different in kind, not just degree. Also in the inbox: Nvidia's biggest customers start seeing the memory crunch show up as an actual price hike.
A Harvard mathematician working with Claude has apparently cracked the question of whether the six-dimensional sphere S⁶ carries a complex structure — open for 78 years — by constructing the object instead of arguing over its existence. Levent Alpöge, a junior fellow in Harvard's Society of Fellows who also researches at Anthropic, delivered a 108-page proof (hosted at alpo.ge/s6.pdf) that spells out every matrix, patch and gluing in the construction, then showed the resulting complex three-fold is genuinely the six-sphere. Mathematician Qiaochu Yuan scrutinized the argument with GPT-5.6 Sol: six minutes in, no flaw; fifteen more minutes, still none. If the proof holds, the answer is yes — a claim Fields Medalist Michael Atiyah staked in 2016 only to see his attempt collapse, and a problem the late Chen Xingshen also studied.
The important shift is in how the result was reached. Rather than hunting inside a known space of possible answers, the model helped manufacture a genuinely new mathematical object — turning the S⁶ question from "does one exist?" into "here it is." It's also the third frontier result for Anthropic's models in 35 days: on July 20 Alpöge used Claude Fable 5 to topple the 87-year-old Jacobian conjecture, and on August 10 a research Claude lifted the proven share of Riemann zeta zeros on the critical line to 67.2 percent (we covered that one — Claude lifts Riemann zeta zero bound to 67.2 percent). For the past three results you could still squint and call the model a powerful search engine; here the object didn't exist before it was made. That's the line where AI stops looking like a tool and starts looking like a collaborator.
Nvidia's largest customers have been told their AI servers are going up more than 15%, effective early 2027 — the moment the memory shortage finally turns into a bill. The notices, reported by Bloomberg and passed along by the contract manufacturers that assemble racks for Microsoft, Google and Oracle, cover the flagship Vera Rubin and Grace Blackwell platforms and land because memory — not the GPU — has gone vertical. DRAM contract prices roughly doubled in the first quarter, and memory makers Samsung, SK Hynix and Micron hold about 90% of the market and little spare capacity, so hyperscalers willing to pay are helping push the price up further.
What makes this worth watching is that the cost is moving down the chain to the actual box. Quanta, the largest AI-server contract maker, saw revenue double year over year while gross margin slid below 5% in the second quarter — squeezed between pricey memory and prices set elsewhere — and it has started shifting from buying components itself to consignment, making customers source the memory. One Vera Rubin NVL72 rack carries more than 20 terabytes of memory and runs into the nine-figure range per rack, so a 15% bump is far from rounding error. The AI boom's budget problem was always going to land somewhere; this quarter it lands on the people holding the OpenAI and Azure contracts.
What to watch: Whether the S⁶ proof survives peer review, and whether Anthropic confirms which model produced it — a detail conspicuously left open for all three recent results.
If AI can now build proofs instead of just searching for them, what — if anything — is left that only a human mathematician can do? Tell us in the comments.
