Monologue: The AI Data Center Overbuild

In progressEd Zitron (Better Offline)podcast2026-09-25read planted ai llmsbusiness

Synopsis — AI-drafted from Dan's notes

An eleven-minute monologue that reads out the argument of Zitron’s newsletter “Where’re All The AI Chips?”. His claim is that a large share of the AI hardware sold so far is not running: it sits in warehouses or in data centres with no power yet.

He starts with Microsoft. In an investigation with the Guardian he reported about 2.2 million GPUs in service, which he now puts at roughly 2 gigawatts and $50 to 60 billion of chips. Bloomberg, he says, reported that only about 2 of Microsoft’s 12 gigawatts of capacity are for AI. Against more than $265 billion of capital spending since 2022, he estimates that up to $106 billion of GPUs and related hardware is uninstalled. He then looks for the same pattern elsewhere, in the “construction in progress” line of balance sheets, where unfinished buildings and uninstalled chips sit: about $374 billion across Google, Meta, Oracle, Amazon, SpaceX, Tesla and several neoclouds and colocation firms, and still growing. From this he guesses that about half of all AI hardware sold is waiting to be installed, so Nvidia’s recent sales are in effect pre-orders for 2028 and beyond.

The conclusions follow from that. Capacity announcements can’t be trusted. If Microsoft can’t bring its AI capacity online, nobody can. The apparent scarcity of compute is two labs, OpenAI and Anthropic, taking what little arrives, and half to three-quarters of the hyperscalers’ revenue backlogs are contracts with those two, which he says they cannot pay. OpenAI, he adds, was 70% of Microsoft’s AI revenue in its 2026 fiscal year. So he calls it an overbuild “magnitudes worse” than the dot-com bubble: like fibre that was never laid, with five to fifty times more capacity planned than demand. And there is no cheap afterlife for the hardware, because serving inference has an airline’s fixed costs. He ends on Dario Amodei telling Dwarkesh Patel in February that if he bought a trillion dollars of compute and revenue came in at $800 billion, there would be “no hedge on earth” against bankruptcy.

The anchor figures come from his own reporting, and Microsoft disputes the main one: it told the Guardian the estimates were “inaccurate” and drawn from incorrect assumptions. The Guardian investigation itself (August) and the 2.2 million figure check out, as do Satya Nadella’s remark in November 2025 about chips in inventory he couldn’t plug in, and the Amodei quote. The construction-in-progress totals, the backlog shares and the 70% are his, from his newsletter, and were not checked against filings. The move from construction in progress to warehoused chips is the weak step: that line also holds buildings, land and other projects, and the share he assigns to chips is his own estimate.