The utilization numbers are the part that should terrify investors. Epoch AI estimated that the average data center GPU sits idle 50-70% of the time outside of training runs. Microsoft reportedly hit just 30% average utilization across its Azure AI fleet in Q4 2025 before they started subleasing capacity.
Compare that to cloud computing, which took a decade to reach 60%+ utilization and only got there because fungible workloads (web hosting, databases, CI/CD) could fill gaps. AI accelerators are purpose-built. You can't just throw a Postgres database on an H100 cluster when inference demand dips.
The subprime comparison actually undersells it in one way: at least mortgage-backed securities had an underlying asset that existed. Half these AI capex bets are against demand curves that are purely theoretical.
The utilization numbers are the part that should terrify investors. Epoch AI estimated that the average data center GPU sits idle 50-70% of the time outside of training runs. Microsoft reportedly hit just 30% average utilization across its Azure AI fleet in Q4 2025 before they started subleasing capacity.
Compare that to cloud computing, which took a decade to reach 60%+ utilization and only got there because fungible workloads (web hosting, databases, CI/CD) could fill gaps. AI accelerators are purpose-built. You can't just throw a Postgres database on an H100 cluster when inference demand dips.
The subprime comparison actually undersells it in one way: at least mortgage-backed securities had an underlying asset that existed. Half these AI capex bets are against demand curves that are purely theoretical.