The scale of AI infrastructure spending is staggering. Microsoft, Google, Amazon and Meta will collectively spend over $200bn this year on data centres and GPUs. That is more than the entire global semiconductor industry spent on fabrication plants in 2023. The assumption behind it is that AI demand will grow exponentially forever.
But the cracks are showing. Microsoft reported slower Azure AI growth last quarter. Google’s cloud division missed revenue expectations. Amazon’s AWS growth has plateaued. The hyperscalers are still spending, but their tone has shifted. CFOs are now talking about “efficiency” and “rightsizing”. That is code for we spent too fast.
The GPU glut is real. Nvidia’s lead times have collapsed from 12 months to a few weeks. Second-hand H100s are trading below list price. Startups that bought capacity on multi-year leases are trying to sublet it. The market is long compute and short applications.
The bubble is not popping in a single day. It is deflating slowly. The buildout will continue for another year because the orders are already placed. But new commitments are getting harder to justify. The ROI on AI inference is still unclear. Most enterprise use cases remain experimental.
When the music stops, the companies with the weakest balance sheets will get crushed. Some data centre REITs are already down 30% from their highs. The next 12 months will separate the real demand from the speculative frenzy.
Paul