Is AI infrastructure spending a bubble?
Bears
- Michael Burry — an accounting argument, not a valuation one: "Depreciation cannot move two directions at once if it were a measurement of something. It is an economic lever, not a physical measurement." Cites Amazon shortening server lives while Meta extends them on the same hardware.
- Jim Chanos — neoclouds are "financial conduits, not technology companies"; Hopper rental rates down ~28% y/y
- Jeremy Grantham — "the most expensive market in American history," warns of a drawdown "closer to 70%"
- David Einhorn — calls SpaceX's listing "the meme-ification of the market at scale"
Bulls
- Jensen Huang — declared AGI had arrived, ~9 Sep; at GTC promised "$1 trillion in cumulative revenue from Blackwell and Vera Rubin through end of 2027… we are going to be short"
- Satya Nadella — the 2014 cloud analogy: "We knew there was going to be margin in that area, and we kept building"
- Masayoshi Son — $40–65B of OpenAI commitments
- Larry Fink — "own assets," AI infrastructure
Burry's is the harder and more falsifiable claim: not that AI is overvalued, but that hyperscaler earnings are overstated. It landed hard enough that Nvidia emailed sell-side analysts a memo rebutting him — an issuer lobbying analysts against a short seller is itself a tell.
Howard Marks occupies the only interesting middle ground, with a distinction the bears skip: "More money is going into inference capex these days than training capex… inference capex is taking place in response to actual demand." His advice — "no one should go all-in without acknowledging risk of ruin" — is the most defensible position in the argument.
Telling detail: when Huang declared AGI on 9 September, Nvidia fell 2%. Gary Marcus asked publicly whether Huang "had a financial incentive to declare AGI." The market appears to agree with Marcus.