The hyperscalers have spent over $300bn on AI data centres and GPUs in the past two years. That number keeps climbing. Microsoft, Amazon, Google and Meta all told investors they will spend even more in 2026. But the revenue coming back from those investments is not matching the hype. Enterprise AI adoption is real but narrow. Most companies are still experimenting. They are not deploying at scale.
The signs are visible. GPU rental prices have dropped 40% year on year. Second-hand H100s trade at half their original cost. Venture capital funding for AI startups is slowing because the returns just are not there yet. The public cloud providers are reporting slower growth in their AI segments. They blame capacity constraints but the real story is that demand is not infinite.
I look at the analogue. The dot com bubble was fuelled by fibre optic spending that never earned back its cost. The same pattern is repeating. The difference is that the big tech companies have massive cash piles to absorb losses. But that does not change the maths. If the revenue per dollar of capex keeps falling, the music stops.
The bubble will not burst overnight. It will deflate as investors realise the payoff is years away, not quarters. The smart money is already rotating out of pure AI infrastructure plays into software that actually delivers margin. The rest will be left holding the bag.
Paul