What's this about?
A video essay on a collision course: leading labs' revenue has been growing ~10× per year while global lab compute only grows ~3× per year. If demand keeps outrunning supply, some combination of lab margins, compute prices, and inference's share of compute must rise — he reports Anthropic's inference margins already went from ~40% to ~80% in a year.
What exactly is "inference vs. training" compute?
Training is the one-time cost of building the model; inference is the ongoing cost of running it for users. Historically labs spent most compute on training; as products take off, inference eats the budget — and unlike training, you can't schedule it for cheap hours. It scales with every user request.
Why does it matter?
Released Aug 3, discussion still forming. If his arithmetic holds, "tokens get cheaper forever" — the assumption behind many AI business models — inverts within a couple of years, and compute-efficiency work (like today's Cursor token-efficiency push) becomes existential rather than nice-to-have.