The hyperscalers are spending hundreds of billions on data centres and GPUs. The numbers are staggering. This is not venture capital. This is cash from operating businesses. The assumption is that the demand for AI inference will eventually fill these server racks and justify the cost.
But look at the actual revenue from AI products. It is growing, but it is not growing exponentially in the way the infrastructure spending is. The gap between capital expenditure and operating income is yawning. The energy costs alone are a huge liability.
The FOMO is real. No one wants to be the cloud provider without capacity. But this creates a prisoner's dilemma. If everyone builds for peak demand, the market will be flooded with compute. Prices for inference will fall. The margins will shrink. The whole thesis of infinite demand for compute rests on models getting better and use cases exploding. That is a bet, not a certainty.
The music is still playing. The spending continues. But the underlying economics look brittle. A correction is coming. It might not be a crash, but a long, painful period of overcapacity and consolidation. Paul