Hyperscalers are hiding a massive secret about uninstalled GPUs | Ed Zitron

$200-300 billion worth of AI GPUs are estimated to sit uninstalled, and if borrowing costs tied to 10-year Treasuries reach 9-13%, interest payments—not accounting fraud or weak products—will likely burst the data-center buildout [39:10]. The installed-GPU count is inferred from capex and construction-in-progress, not audited working systems [16:40].

Key points

• Oracle issued a force majeure notice about its New Mexico project because it is not confident power will arrive by 2028, despite a hell-or-high-water clause that says it must pay if power fails [02:54]. That move is a legal shield, not a cancellation, and it shows the binding constraint is grid connection, not data-center construction [02:23]. Ed extends the point to other sites, including a $7 billion Port Washington power bond, a 6-month pipeline delay on Jupiter, and Bernstein’s 35% built-rate figure, which he thinks is still too high [11:56].

• Hyperscaler and neocloud filings show hundreds of billions in construction-in-progress, most of which Ed treats as data-center hardware rather than buildings, with a $374 billion total across SpaceX, Google, Meta, Amazon, Core, and Applied Digital [17:38]. He estimates Microsoft has only $50-60 billion of AI chips actually installed, against more than $265 billion of capex, implying roughly $120 billion of chips exist but $60-80 billion are not in service [18:13]. The figures rest on capex and construction-in-progress, not audited GPU counts [16:40].

• Oracle’s credit rating is one step above junk; if S&P downgrades it, institutional holders must sell, the yield rises, and it becomes a fallen angel, a fate Ed cites for Ford and nearly Boeing [14:33]. New AI debt may be priced at 9-13% as benchmark 10-year Treasury yields rise [32:41]. With best-case data-center margins of only 30-40% and construction taking two to four years, interest-only payments become the burst mechanism [32:58].

• OpenAI and Anthropic face about $200 billion of committed compute payments starting next year, then rising, which makes hyperscaler revenue forecasts dependent on lab solvency and project timing [34:12]. Anthropic needs $50-100 billion per year in debt if it goes public, but bond markets would price that debt at 10-14%, unlike the more lenient stock market [36:46]. If the labs stop paying, hyperscaler revenue fails; if they keep paying, debt markets punish them [35:35].