The hidden ticking bomb behind the AI boom

Michael Burry accuses Big Tech of underestimating GPU depreciation.

On Wall Street, Michael Burry is famous for spotting the subprime crisis long before anyone else. Immortalized on screen in The Big Short, the investor is making headlines again. This time, he’s betting on the bursting of a bubble surrounding generative artificial intelligence, amid colossal investments and sky-high valuations. More specifically, he’s shorting two flagship stocks of the sector: Nvidia, the graphics-card leader, and Palantir, a specialist in data analysis.

Beyond his market wager, Michael Burry is raising a crucial issue: the depreciation of AI chips, purchased in massive volumes over the past two years by U.S. tech giants to boost their computing capacity, essential for training and running AI models. He accuses these companies of underestimating the depreciation of these assets in their financial statements. This practice would allow them to artificially — and temporarily — limit the impact of their capital expenditures on their profits.

Meta Saves $2.9 Billion

Depreciation is an accounting principle that spreads the cost of an investment over time. Take a GPU costing $20,000, with an expected useful life of four years. Its cost won’t be recorded in full during the first year; instead, the company will account for $5,000 per year over four fiscal periods. If the GPU’s lifespan is extended to six years, the annual charge falls to $3,333 for two additional years.

Each company is free to determine the depreciation period of its investments. Given today’s colossal spending levels, even a small adjustment can have a significant impact. Earlier this year, for example, Meta extended the estimated useful life of its AI chips from four and a half years to five and a half. The result: its charges will drop by $2.9 billion in 2025 compared with what they would have been without this change — and its profits will rise accordingly.

$176 Billion Over Three Years

Facebook and Instagram’s parent company isn’t the only one adopting this strategy. Google and Microsoft also now depreciate their chips over six years — double the period used in 2020. Oracle, which has pledged massive investments to meet OpenAI’s needs under a $300 billion contract, and CoreWeave, the largest “neo-cloud,” use the same duration. One notable exception: Amazon, which also extended its depreciation period from five to six years in 2024, but reversed course this year.

According to Michael Burry, these estimates aren’t realistic, since graphics cards become obsolete increasingly quickly, replaced each year by more powerful models from Nvidia and its rivals. He believes the actual useful life is no more than two or three years. He estimates that Google, Microsoft, Amazon, Meta, and Oracle will understate GPU depreciation by $176 billion between 2026 and 2028. Oracle’s profits would be overstated by 27% and Meta’s by 21% in 2028.

“Far From Obsolete”

In March, Jensen Huang added fuel to the debate during the launch of the Blackwell architecture: “There are circumstances where Hopper, [the previous generation] is fin… but not many.” Speaking to Cafétech in June, Nvidia’s CEO softened his tone: “The older models are far from obsolete, but I recommend buying the latest because the performance per dollar is much better.” Indeed, while older GPUs may no longer suffice for the most complex tasks, they remain useful for lighter computations such as inference.

CoreWeave reports that demand for its A100 chips, released in 2020, is currently at full capacity. Demand for H100-based machines, launched two years later, is at 95%. But this level of demand alone doesn’t justify extending depreciation periods: revenues generated by these GPUs also matter, and the prices billed to customers tend to decrease over time. According to The Information, Oracle even reports negative margins on some older models.

A Risky Strategy

To be profitable, GPUs must generate revenue exceeding their depreciation. If a company opts for linear depreciation — the same annual amount every year — this equation becomes harder to meet over time. In short, the strategy boosts short-term profits but could result in accounting losses in coming years on older GPUs that haven’t yet been fully depreciated.

To offset these potential losses, companies are betting on a sharp rise in AI-related revenues. This will be all the more necessary as total depreciation will continue to grow along with soaring capex. For the tech giants, the worst-case scenario is slower growth or a decline in profits. For the neo-cloud players, however, the stakes are much higher: they may no longer be able to repay the enormous debt taken on to finance their expansion.