In accounting, depreciation spreads the cost of equipment over its “expected useful life,” so extending that lifespan lowers the annual expense recorded against earnings and, potentially, increases reported profits.
Burry has estimated that between 2026 and 2028, hyperscalers will understate depreciation by more than $175 billion, according to Business Insider. And while Eisman accepts the accounting logic, he thinks Burry is putting too much weight behind it.
“With all due respect to Michael, I think his argument is too academic,” Eisman said, claiming that older AI chips aren’t losing their value nearly as quickly as Burry assumes. Demand for computing power remains so strong that previous generations of GPUs are still finding customers, even as newer and more powerful chips reach the market, often with long order backlogs.
As Business Insider reported, CoreWeave offers a striking example. The AI infrastructure company recently disclosed a contract for Nvidia A100 GPUs that runs through 2029, even though the chips were introduced six years ago. CoreWeave CEO Michael Intrator said pricing for some older-generation hardware remains at or above levels seen years ago.
IREN, another crypto miner-turned-AI data center operator, has made a similar observation. On a recent earnings call, management said older GPUs remain in strong demand because the market is still undersupplied.
For Eisman, chip depreciation is a secondary concern. If AI demand keeps growing, he argues that stretching a chip’s useful life from three or four years to five or six isn’t a serious problem. But that doesn’t mean Eisman thinks the AI trade is safe. In fact, he sees a much bigger vulnerability elsewhere.
Two companies underpin the entire AI boom
Major tech companies are forecast to spend as much as $750 billion on AI this year, a figure that could nearly double to $1.4 trillion in 2027, according to Goldman Sachs estimates. But Eisman believes much of that spending ultimately depends on two companies: OpenAI and Anthropic.
Much of that dependence can be seen throughout the AI supply chain. At Nvidia, which dominates the market for AI chips, Eisman pointed to a company filing showing that 70% of its accounts receivable came from just five customers. He called that a “warning flag.”
The concentration risk extends to hyperscalers buying Nvidia’s chips. Eisman said roughly 70% of the AI revenue generated by Microsoft, Meta, Amazon and Alphabet comes from OpenAI and Anthropic, equivalent to about 25% to 30% of their overall cloud revenue. At Oracle, he said roughly half of the company’s remaining performance obligations, a measure of contracted future revenue, are tied to OpenAI.
“The entire chain from Nvidia to the hyperscalers all the way down […] rests on the future health and growth of Anthropic and OpenAI,” Eisman warned.
