Nvidia says its AI accelerators can remain economically valuable longer than standard accounting schedules imply. Michael Burry sees a familiar argument in that claim: investors made a similar case for computer-leasing businesses in the 1960s, when demand was said to be nearly limitless.
What Nvidia’s slide argues
A presentation dated September 27 compared the residual value of Nvidia’s A100, H100 and B200 accelerators with a five-year accelerated depreciation schedule. Its argument was that companies may write down AI infrastructure faster than the equipment loses economic value.
Supporters took the slide as a response to criticism that GPUs become obsolete too quickly. But the accounting schedule and the market value of a used accelerator measure different things. An asset can still have value after it has been written down, while its resale or rental value can fall well before its accounting life ends.
Burry’s historical comparison
In a post titled Don’t Believe Your Lyin’ Eyes, GPU Depreciation & Useful Lives, Burry pointed to Adam Smith’s 1968 book The Money Game. The passage he cited describes investors in computer-leasing companies who expected demand for computing equipment to keep growing and profits to rise year after year.
Questions about how quickly the machines should be written off were treated as old-fashioned doubts about a new market. Burry sees a parallel in today’s arguments about AI infrastructure: confidence in sustained demand is being used to support confidence in the useful life of expensive equipment.
His comparison is not a claim that the two markets are identical. The point, as Burry frames it, is that market behavior can outlast the technology around it. A new generation of hardware may arrive quickly, while investors continue to make familiar arguments about durable demand and long-lived assets.
The rental-market question
The debate is not confined to accounting. In September, Nvidia CEO Jensen Huang described the company’s compute infrastructure as a “durable and well-rentable asset.” Investor Jim Chanos responded with a practical question: if GPUs are such attractive rental assets, why doesn’t Nvidia rent them to customers itself?
The exchange came amid rising rental rates for three-year-old H100 GPUs. That gives the discussion a useful complication. Older hardware can still command a rental price, but the existence of demand does not, by itself, show how long that demand will last or how its value will change as newer chips arrive.
What depreciation cannot settle
Nvidia’s presentation and Burry’s response focus on different risks. Nvidia points to value that remains after a standard depreciation period. Burry points to the possibility that newer GPU generations could reduce the value of older ones sooner than expected.
Neither an accounting schedule nor a single rental-market observation answers the central question: how long will existing AI infrastructure keep earning enough to justify its cost? That is a question about future demand and competition, not simply about when a spreadsheet says an asset has been written down.
The 1960s comparison is a warning about confidence, not proof that today’s AI market will follow the same path. GPUs may keep earning after their accounting lives end. The harder part is knowing when “still useful” stops meaning “still worth what investors paid.”