Dead Money

This argues that the recent hyperscaler spending spree on GPUs and data centers is largely speculative and wasteful: hundreds of billions of dollars’ worth of NVIDIA (and Broadcom) chips are sitting uninstalled because of power, cooling, labor and other physical-execution bottlenecks, and research suggests more than half of GPU servers sold between 2026 and 2028 may have nowhere to be plugged in. Big cloud providers have been hoarding capacity, creating an illusion of broad demand when most compute consumption is concentrated in a few AI labs. Rough tallies show roughly $200-$300 billion of GPU sales unutilized, around $217 billion funneled to the largest AI labs in nine months, and two companies (Anthropic and OpenAI) accounting for roughly 60-80% of many hyperscalers’ AI revenue. Much of the remaining “demand” comes from VC-funded startups that will evaporate if funding dries up.

The financial punchline is stark: to justify recent capex, hyperscalers need vastly more AI revenue than currently exists. Estimates put required annual AI revenue at about $308 billion just to break even on 2026-27 capex (and $417 billion for a 10% ROIC), rising to $2-3 trillion by 2030 to cover broader investments. Current combined AI revenue for major hyperscalers is roughly $183 billion, leaving a gap of $125-$243 billion versus break-even - evidence that large parts of the AI buildout risk becoming dead money unless demand scales dramatically.