AI Bubble: 'This could humiliate the largest companies in the world'

Zitron’s claim is that a large share of Nvidia’s data-centre revenue is chips that were bought and never installed. His anchor is Microsoft: a leaked breakdown he converts to power puts its AI chips at about 2 GW, against roughly 12 GW of claimed capacity. On his own assumption that about half of capital spending goes on GPUs, he scales that up to the industry and lands on a range of $100bn to $350bn of hardware sitting in warehouses, and says plainly that he is speculating.

The mechanism he offers is supply rationing: Nvidia’s buy-now-or-miss-the-next-generation terms make buyers race each other, and the chips age a generation before the buildings meant to hold them are finished. Chips that are not in service don’t depreciate, so the loss stays off the books until a write-down. His ask is a disclosure rule: hyperscalers should say how many chips they own and how many are actually online.