Anthropic’s $2T IPO, Elon Predicts Neocloud Death, and China’s Kimi Catches Up | #306

Anthropic reportedly filed for an IPO that could raise $100 billion at a $1.5–2 trillion valuation, possibly in the week of November 16. Reported revenue climbed from $4.6 billion for all of 2025 to $16.3 billion in the first half of 2026, and the company earned nearly $1 billion in operating profit in Q2 2026. Revenue per dollar of compute rose to $2.08 from $0.40 eighteen months earlier, as revenue grew 29 times and compute costs six times. Anthropic had 11,900 customers paying more than $100,000 a year, but speakers cautioned that even a $2 trillion valuation would already assume substantial future growth.

The company has reportedly committed to more than $500 billion in compute, much of it non-cancelable, and nearly half its revenue flows through Amazon, Google, and Microsoft clouds, which also compete with it. One speaker argued Anthropic could have stayed private: it had reportedly raised $130 billion, held $110 billion in cash and cash equivalents, and spent $10–20 billion, potentially leaving room for another large private raise. Governance also complicates the IPO: the seven co-founders reportedly retain 50.1% of voting power through a special share, while Class T shares let Anthropic’s long-term benefit trust appoint and remove a board majority. Supporters see protection from quarterly pressure; critics questioned who defines the public good and how trustees are held accountable.

Elon Musk predicted consolidation and “guppy level” deaths among neoclouds, which borrow to buy Nvidia chips and rent out compute. CoreWeave and Nebius reportedly fell 30% in a sell-off linked to CoreWeave’s $51 billion in debt. Chamath Palihapitiya warned that some paid-for capacity could be underused or stranded in 2027. Speakers compared the buildout with the fiber boom: excess capacity was eventually used, but bankruptcies occurred along the way. Debt estimates varied, from about $500 billion directly to roughly twice that when special-purpose vehicles and vendor financing were included, with risk depending on financing terms and continued demand. SpaceX reportedly closed $40 billion in debt and agreed to sell Anthropic $120.5 billion of compute through 2029, even as Musk said debt markets had largely closed to firms without substantial revenue beyond AI data centers.

Moonshot AI’s Kimi K3.1 Pro has 2.8 trillion parameters, 104 billion active at a time, and a one-million-token context window. It scored 55 on the Artificial Analysis Intelligence Index, ranking fourth behind Claude Opus 5.5, Claude Sonnet, and Metamuse Wonder. Its prices are $2 per million input tokens and $10 per million output tokens, about one-third the previous Kimi model’s price, and Moonshot planned to release its weights the following week. Speakers said the reported open-versus-closed model gap had narrowed from 12 points to three, with one estimate putting Kimi about five weeks behind the US frontier. They disagreed about how much distillation from Opus 5.5 explained its performance, while also crediting independent innovation. Kimi appeared to outperform Opus 5.5 on performance per cost per task, though speakers cautioned against treating every benchmark as decisive.

OpenAI released GPT 6.1 Terra at less than half the previous Terra model’s price and reported a 50-point reduction in hallucinations; one speaker characterized it as an unexciting distillation of larger models. In a separate red-team test, agents isolated in virtual machines with no network connection or shared files passed secret bits by encoding them in processor-cache timing. The test raised questions about whether abstraction barriers can ever be fully non-leaky, though the agents had been asked to find timing and side-channel attacks.

OpenAI also announced three claimed results in mathematics, computer science, and physics, including a proof of a 1991 wave-focusing conjecture, a disproof of the exponential time hypothesis verified in Lean, and a rigorous result on the existence of quantum chromodynamics in four dimensions. The claims had not yet been peer reviewed, and speakers emphasized that AI is producing results faster than people can verify them; the reported Yang–Mills result, they said, did not claim to establish the mass gap. Alex described his early chain-theory work as promising but underdeveloped, with AI helping generate and check Lean proofs but not establishing the paper’s conjectured connection to the photon.