I was sitting in a pub with an old mate who works at one of the big cloud providers. He was telling me about the numbers they're seeing. Data centre buildouts doubling every six months. Power grid upgrades being fast-tracked for GPU clusters nobody has actually proven they need yet. He said it feels like 1999 but nobody wants to say it out loud because the money is too good.
The thing is, there's a difference between building something useful and building something because everyone else is. Right now the hyperscalers are spending like there's a gun to their head. Miss the AI wave and your company dies. That's the logic. But that same logic drove people to buy pets.com at 80 times revenue.
The real question is whether the demand actually shows up. Inference costs are dropping fast. Models are getting smaller and more efficient. You don't need a million dollars worth of H100s to run a decent LLM anymore. That means the need for giant clusters might peak sooner than the capex schedules assume.
I'm not saying the whole thing collapses overnight. But there are already signs. Some startups that raised huge rounds for cloud compute are burning cash faster than they can raise the next round. The secondary market for GPU reservations is softening. If the growth rate of AI usage slows even a little, all that capacity gets repriced.
People keep telling themselves this time is different because AI is a productivity revolution. Maybe. But revolutions still have hangovers. The hangover comes when you realise you overpaid for the party.
Paul