Microsoft Slashes Spending on Anthropic AI Claude - AI is Too Expensive

A 26-minute episode, uploaded on 8 October 2026, in which Eli reads through The Information’s report of 5 October and argues with it as he goes. The report: Microsoft had projected at least $1 billion a year of internal spending on Anthropic and has cut that by more than a third. At Meta, Claude Code users fell from about 60,000 to about 30,000, partly from the spring layoffs and mostly because Meta pushed its own tools, MetaCode inside and Muse Code, released publicly in August. Meta had spent more than $105 million on Claude Code in a single 28-day stretch while staff competed on a token leaderboard, until a memo in June brought in budgets and caps. Microsoft’s Cloud and AI group, more than 60,000 people, cut each engineer’s monthly AI allowance from $100,000 to about $10,000 in most cases, and Jay Parikh told staff that moving work to OpenAI models gave better value.

Eli’s case is about architecture, not personalities. He never believed one or two labs would own the “intelligence layer” for everyone, because systems are not built that way. Companies with the means build their own; everyone else mixes in open-weight and Chinese models. He offers Reflection AI’s new open model, which its makers say competes with Z.ai’s. $100,000 a month per engineer, $1.2 million a year, was never sustainable, and he blames agents built to loop as fast as the hardware allows, several of them scanning the same repository for different people, when a delay of minutes or hours would do. He thinks the layoffs blamed on AI came from interest rates. And he uses NoSQL to warn against confusing a technology with a stock: NoSQL changed the world, and that was no reason to buy Twitter. Big pharma will use AI, and build its own.

He leans on two claims he doesn’t source. Anthropic is aiming for a November listing at about $2 trillion, which is reported. But Michael Burry’s argument, as Eli gives it, that Anthropic must list before third-quarter numbers expose a run-rate inflated by token-maxing, does not appear in anything Burry is reported to have said. Burry has said token-maxing won’t last and that markets should sink both IPOs. OpenAI’s new $500 Pro plan is real; the $200 plan kept its price but halved the allowance for new subscribers.