Monologue: The AI Data Center Overbuild

Microsoft has spent over $265 billion on capital expenditure since the start of 2022, but only about $50 billion worth of GPUs have actually been put into service. The rest—perhaps as much as $106 billion in chips and associated hardware—is sitting in warehouses or in powered data centers that have not been brought online. This is the central finding of an investigation by Better Offline host Ed Zitron and The Guardian, and it points to a much larger problem across the entire AI industry.

Zitron estimates that roughly 50 percent of all AI hardware and chips sold have been warehoused, with Nvidia making somewhere between $200 and $300 billion in revenue on products effectively sold twenty-four to thirty-six months in advance. The apparent scarcity of AI compute is not a sign of insatiable demand, but rather a consequence of most available capacity being funneled directly to two companies: OpenAI and Anthropic. The rest of the world is left with very little, while the capacity that does exist takes agonizingly long to build.

The Construction in Progress Problem