The scale of capital expenditure on AI infrastructure is staggering. Cloud providers and large enterprises are committing tens of billions to data centres and GPUs. The assumption is that demand for AI compute will continue to grow exponentially.
The disconnect sits between this spend and the revenue it generates. Enterprise AI adoption is happening, but the killer use case that justifies this specific hardware density is still forming. If the revenue does not materialise at the rate the spending implies, the correction will be sharp. We have seen this cycle before in telecoms and cloud.
The hyperscalers argue this is a generational build-out. They compare it to building the internet backbone. The difference is the cost of capital and the energy requirements. In the UK, grid connection delays are a practical check on ambition. The market is bifurcating between the few who can fund this and the many who are priced out.
The real test is next year. If the major cloud earnings calls show a plateau in AI services revenue against a rising capex bill, the mood shifts. The bubble does not burst overnight. It deflates as the cost of compute drops and the premium on raw access fades. The winners will be those who built for specific workloads, not general speculation.
Paul