Hyperscalers depreciate GPUs over five to six years. Shorten that by two, and the math behind $119 billion in Nvidia supply commitments unravels.
A guest on CNBC's Closing Bell Overtime framed the AI chip bull case as a maturity mismatch: hyperscalers issuing 10-year bonds to buy assets that may depreciate in five years, a "Michael Burry bear case" that threatens the $119 billion in supply commitments Nvidia disclosed in its most recent 10-Q.
"If you're going to issue bonds that are 10 years out and the depreciation of the asset is five years, well, what do you do?" the guest said, arguing the AI memory hardware food chain is still trading as if there is no alternative.
Nvidia's Q1 FY27 revenue hit $81.61 billion, up 85.2% year over year, with its Data Center segment contributing $75.25 billion. The stock trades at a trailing P/E of 31 and a forward P/E of 23 — multiples that depend on hyperscaler capital expenditure remaining voracious. Advanced Micro Devices, up 171.25% year to date, trades at a trailing P/E of 192, while Intel, up 278.4%, carries a forward P/E of 152 despite a $3.73 billion GAAP net loss in its most recent quarter.
Polymarket traders assign only a 49.5% probability that Nvidia closes above $200 by end of July, and that crowd has been right on Nvidia 74.4% of the time across 195 resolved markets. The Burry critique only requires the depreciation schedule to be wrong by a couple of years for the earnings power investors are paying for to be rebuilt with different numbers.
The circularity question
The bull case for AI hardware rests on a food chain that increasingly appears to be buying its own output. Nvidia disclosed $30 billion in multi-year cloud service commitments alongside its $119 billion in supply commitments. On Reddit, a thread titled "AMD is literally funding a startup with $350M just so they can buy AMD chips" accumulated more than 1,600 upvotes, reflecting growing anxiety about circular financing in the AI chip ecosystem.
Taiwan Semiconductor Manufacturing Co., the foundry backbone for all three major GPU designers, posted May consolidated revenue of NT$416.98 billion, up 30.1% year over year. Chief Executive C.C. Wei guided to more than 30% full-year revenue growth. Yet the stock sold off 5.08% on June 23 on valuation and capital expenditure concerns, and overseas fab expansion in the US, Japan, and Germany represents a structural margin headwind.
The consumer canary
The end user ultimately funding this compute buildout is showing signs of strain. Nike posted fiscal Q1 2027 revenue down 1.1%, with Greater China falling 12% as reported and 17% on a currency-neutral basis. Lululemon also flagged North American weakness, a pattern the CNBC guest described as potentially structural. May job openings came in at 7.59 million versus the 7.3 million estimate, so the labor market is not obviously buckling. But AI has not visibly reduced corporate headcount, meaning the productivity payoff justifying $119 billion in supply commitments remains hypothetical.
Cheaper "good enough" models such as DeepSeek attack the demand side of the equation. If inference becomes efficient enough to require meaningfully fewer GPUs per unit of intelligence, the training-side capital expenditure commitments look extended. Frontier model OpenAI made a deal with Cerebras in January to use the startup's SRAM-based chips in its inference stack, and engineers are now seeing reduced inference costs, according to a report in The Information. Cerebras stock rose 19% on Monday and another 2% on Tuesday.
Nvidia shares, trading at 23 times forward earnings, have priced in uninterrupted hyperscaler spending. If GPU depreciation schedules shorten by even two years, the operating income at the buyers — Amazon, Microsoft, Google, Meta — tightens while the capital expenditure hole deepens. The food chain that has been trading as if there is no choice may discover there is one.
This article is for informational purposes only and does not constitute investment advice.