$9 trillion. That's the figure now being attached to global AI data center investment commitments, and the Financial Times is asking the question that many in the industry have been avoiding: what if demand never catches up?
The numbers are staggering even by Big Tech standards. Microsoft, Google, Amazon, and Meta have collectively pledged hundreds of billions in capital expenditure on AI infrastructure over the next several years. Microsoft alone plans to spend roughly $80 billion in its current fiscal year on data centers, most of it AI-focused. These facilities need massive amounts of power, specialized cooling, and thousands of the most expensive chips ever manufactured.
The Telecom Parallel
The comparison that keeps surfacing is the late-1990s telecom boom, when companies laid enough fiber optic cable to circle the earth thousands of times over. Demand eventually arrived - but not before dozens of companies went bankrupt and investors lost trillions. The infrastructure got used, just not by the people who paid for it.
AI infrastructure carries similar dynamics. Building a data center takes 2-3 years. Demand forecasts for AI compute are based on growth curves that assume enterprise adoption accelerates continuously. But most companies are still in the "experimenting with AI" phase, running pilots and proofs of concept rather than deploying at the scale these facilities are built to serve.
The Bull Case Has Cracks
Defenders of the spending spree argue that AI compute demand is growing so fast that today's capacity will look inadequate within a few years. There's some truth to this - training runs for frontier models keep getting larger, and inference costs (the compute needed every time someone uses an AI model) add up quickly at scale.
But three things could puncture that logic. First, model efficiency is improving rapidly. Techniques like distillation, quantization, and better architectures mean newer models can match older ones using a fraction of the compute. Second, open-source models are getting good enough that not every company needs to rent capacity from hyperscalers. Third, the actual revenue AI generates for most businesses is still modest compared to the infrastructure cost required to deliver it.
Who Bears the Risk?
The tech giants can absorb write-downs that would destroy smaller companies. But even they aren't immune to shareholder pressure if returns take longer than expected. The more interesting risk sits with the secondary players - the data center REITs, power companies that signed long-term contracts, chip suppliers beyond Nvidia, and the construction firms that staffed up for a decade-long building boom.
For anyone using AI tools day-to-day, the practical takeaway is less dramatic. Overcapacity, if it materializes, tends to push prices down. The telecom bust gave us cheap bandwidth that enabled YouTube, Netflix, and cloud computing. An AI infrastructure glut could mean cheaper API calls and more accessible compute. The builders might suffer, but the users rarely do.