Where Are All the AI Chips? The Case for a Data Center Overbuild

Summary

The article argues that the AI infrastructure boom may be substantially less operational than company announcements and investor narratives suggest. Its central focus is Microsoft: the author contrasts Microsoft’s repeated claims about adding gigawatts of capacity with reporting that only about 2 gigawatts were specifically tied to AI chips, and estimates that roughly $50 billion of GPUs were actually in service. Using Microsoft’s capital expenditures, asset classifications, earnings-call language, and outside estimates, the author calculates that tens of billions of dollars, potentially $50 billion to $100 billion, of Microsoft’s GPUs may remain uninstalled or lack connected power. The analysis extends this concern to other hyperscalers, neoclouds, and colocation providers, citing more than $374 billion in reported construction-in-progress assets and estimating that roughly $200 billion or more in GPUs and other AI accelerators could be sitting in warehouses or unfinished facilities. It also compares this estimate with about $561.5 billion in AI chips and related hardware sold by NVIDIA and Broadcom since the beginning of 2023, arguing that a large share of sales may represent advance purchases rather than immediately productive compute. The author says companies often fail to distinguish total power, secured power, IT load, active AI capacity, and revenue-generating infrastructure. Examples involving CoreWeave, Oracle, Amazon, OpenAI, and Stargate Abilene are presented as evidence that public capacity claims can describe partial campuses, future commitments, or ambiguous forms of power. The article further argues that slow deployment, long depreciation schedules, large project-financed debts, and demand concentrated among OpenAI and Anthropic could create an overbuild scenario. These are the author’s interpretations and estimates, not independently verified industry totals. The proposed remedy is more precise disclosure of installed GPUs, powered capacity, storage, deployment timelines, and the proportion of sold chips generating revenue.