Anthropic to reveal first Q3 numbers at pre-IPO investor day
The private AI market's most important number lands Wednesday: Anthropic is handing institutional investors its third-quarter financials for the first time — and after last week's OpenAI revenue whiplash, every decimal will be read as a signal about the whole AI trade.
Anthropic will publish third-quarter financials to institutional investors at a meeting at its San Francisco headquarters on Wednesday, according to two people familiar with the matter — the first time the lab's Q3 numbers have been seen by anyone outside the company. CNBC reported the disclosure; Bloomberg first reported that the investor day was happening. The timing is not accidental: Anthropic is marketing what could be the largest IPO in history, at a potential $2 trillion valuation that would put even SpaceX's debut in the shade, and people familiar with the plans say revenue could top $100 billion by the end of the year. The critical question is accounting, not top-line — OpenAI's updated $50 billion annualized figure, which stripped out cloud-partner revenue and landed $20 billion below what had been reported, was enough to drag down semiconductor stocks and the Nasdaq last week. It is unclear whether Anthropic will count partner revenue the same way. As we covered before the roadshow — Anthropic's $2 trillion IPO turns its trustee experiment into a public-market question. Wednesday's reaction is now a live test of investor appetite, with Hock Tan's warning that climbing rates make capital access the real constraint on AI still ringing through the sector.
China launched the Yau Laboratory Alliance for Mathematics and Artificial Intelligence at Tsinghua University on Sunday — while the West's math establishment keeps warning about the same technology. The alliance unites eight institutions and aims to network around a thousand researchers around specialized math models, machine-verifiable proofs and transparent reasoning; Fields Medal winner Shing-Tung Yau, who fronted the launch, warned that China is still falling behind the United States. The contrast is the story: the same week China organized to build, more than two dozen Western Fields Medalists — Terence Tao and Peter Scholze among them — signed a declaration arguing that AI-generated proofs risk eroding mathematical understanding itself, with parts of the community now calling for a boycott of AI-written results. We covered that declaration — 25 Fields Medalists declare war on the benchmark race. One technology, two postures: build faster, or slow down and protect the craft.
Apple's research team says the fix for data-starved language models is mixing in English text, not tuning hyperparameters. In a study of roughly 1,000 pre-training runs across four model scales from 150 million to 1.43 billion parameters — Arabic as the target language, English as the auxiliary — mixing high-resource data beat aggressive hyperparameter tuning on both validation loss and downstream accuracy, and the gap grew with model size. The authors quantify the boost as equivalent to two to three times the unique target data on validation loss, and 2 to 13 times on downstream tasks. The sharper finding is a warning for anyone who evaluates multilingual models: target-language validation loss systematically underestimates what data mixing buys, because it only captures the regularization effect and not the knowledge the repeated target corpus can never supply.
What to watch: whether Anthropic counts cloud-partner revenue the way OpenAI just did — the same business can read $20 billion differently depending on the accounting.
Should private AI labs be required to disclose audited financials before shopping a trillion-dollar valuation? Tell us in the comments.
