Anthropic published its S-1, and the thing I'm most relieved about is that we can finally stop arguing over leaked "annualised run rate" figures and look at actual GAAP quarterly revenue and a real P&L. The numbers line up with the growth everyone had been whispering about. After that it's a Rorschach test. If you worry about debt you'll find more than half a trillion dollars of compute leases and purchase commitments. If you worry about circularity, the filing itself points out that the hyperscalers are its customers, suppliers, shareholders and competitors all at once. And if you worry about governance, the seven founders keep 50.1% of the vote through a single share with no economic rights, which is Zuckerberg's trick done as a committee. None of it tells you much about where this business is in three years. It's a photo of a thing that's still moving.
The market spent the week looking for cracks and finding plenty of material. Google's search revenue grew 12% but missed, and management admitted AI Mode queries make less money than old ones. Meta is raising up to $95bn in equity, with Nvidia and AMD buying in, so the chip vendors now own a slice of their biggest customer. OpenAI cut its round from $30bn at $1.4tn to $20bn at $1.25tn and has reportedly only got about $11bn of it. Microsoft offered buyouts to 60,000 people, more than a quarter of its staff, to pay for data centres. The underlying worry is simple: capex now runs well past what these companies generate in cash, so they're borrowing and selling shares to fund it, and nobody knows whether the trillions produce a return, or whether the return goes to someone else. The quarterly numbers won't settle that. Meanwhile SK Hynix says memory is sold out through 2029, which is either very good news or exactly what Corning said about fibre in 1999. Elsewhere: North Korea walked off with $6.3bn from Binance, the largest theft of anything, ever, and Binance has frozen about half of it so far. So much for money the government can't take. And OpenAI and Anthropic both say their models disproved the Exponential Time Hypothesis, a 25-year-old assumption about how hard certain problems are, on the same day. The people who wrote the hypothesis are now checking the work. If it holds, a lot of things people stopped looking for faster solutions to, including some in cryptography, are suddenly worth another look.