The man paid to prove Fermat by hand says Claude did it in 11 days

Anthropic published a proof of Fermat’s Last Theorem after dozens of Claude agents produced 13 million lines of Lean code and proved 30,300 intermediate theorems. The effort used 29,500 (additional) Lean steps/components (as reported in the article excerpt) and was completed in 11 days versus a stated five-year grant held by a mathematician. The mathematician quoted argues the result “tells us nothing about mathematics,” even as the AI workflow demonstrates large-scale automated theorem proving.

Analysis

This is less a math headline than a proof-of-capability for agentic workflows that require long-horizon planning, self-correction, and high token throughput. The first market read-through is not to software revenue, but to infrastructure demand: these workloads are exactly where inference intensity and orchestration costs compound, so the economic winner is the model/provider stack and the semiconductor supply chain, not the end-user app layer.

The second-order effect is on enterprise buying behavior. If AI can reliably attack structured, verification-heavy tasks, CIOs will push harder on premium-tier model usage, but they will also demand auditability, test harnesses, and human-in-the-loop controls, which raises the value of adjacent tooling around testing, code security, and verification. That is constructive for platform vendors with distribution and cloud attach, but it is a margin warning for labor-heavy IT services and generic SaaS where AI features become table stakes rather than pricing power.

The contrarian point is that this may be an unusually favorable benchmark for AI because the problem is closed-form and machine-checkable; it does not prove robustness in ambiguous, real-world enterprise codebases. The next 1-3 months matter more than the demo itself: what counts is whether hyperscalers and frontier labs translate these wins into larger enterprise commitments and visibly higher compute spend. If spend does not follow, the market will reclassify this as a research milestone rather than an earnings catalyst.

Key Decisions for Investors

• Maintain a tactical long bias in NVDA/SMH on pullbacks over the next 1-3 weeks; this read-through supports higher inference intensity, but only if hyperscaler capex guidance stays firm. Falsify if cloud capex comments roll over or AI spend shifts toward cheaper, smaller models.
• Prefer MSFT over the broader software basket (IGV) for a 3-6 month relative-value long/short: distribution plus model ownership should capture more of the monetization than generic SaaS. Stop if IGV begins showing evidence of AI-driven net retention expansion without margin dilution.
• No standalone long in AI application names on this headline alone; wait for enterprise rollout data and unit-economics disclosure before paying up for productivity beneficiaries. This is a watch item, not a conviction trade.