Glossary term
GPU Depreciation
The accounting write-down of AI accelerators over their assumed useful life — typically five to six years at the hyperscalers — and the debate over whether that life is too long. A shorter life raises reported costs, lowers earnings and cuts the collateral value behind GPU-backed debt.
AI-generated — produced automatically by Closelook’s systems under this site’s editorial policy.
What it means
A data-centre GPU is a fixed asset. Its cost is spread over its useful life as depreciation, a non-cash expense that reduces reported earnings. The hyperscalers lengthened the assumed life of their servers to five or six years between 2020 and 2023, which lowered annual depreciation and raised earnings. Critics argue that AI accelerators lose economic value faster: each new generation is several times faster per dollar, and rental prices for older chips fall as the new ones ship.
The question is not academic. Depreciation policy decides hyperscaler earnings, the equity value of neoclouds whose assets are almost entirely GPUs, and the loan-to-value of the project debt secured on them.
Why it matters for the AI trade
Every rung of the funding ladder below the top holds GPUs as collateral. If a chip that cost $30,000 is worth half that after two years rather than after three, the lenders to a neocloud are under-secured before the contracts have run. The spot rental market is the public test: if older-generation rental prices fall faster than the accounting life implies, the bears are right about the numbers even if they are wrong about demand.
How Closelook uses it
The compute-cost boards track spot GPU rental prices by chip generation as the market’s own depreciation curve; a planned second version compares actual rental prices with an expected-depreciation curve by launch date — the difference is the signal. The inference economics and valuation gap reads carry the argument; the interest coverage entry shows why depreciation choices change the credit picture.
Common questions
- How long do hyperscalers depreciate GPUs?
- Most assume five to six years for servers including accelerators, after extending the life from three or four years earlier in the decade. Neoclouds tend to use similar or slightly shorter lives.
- Why would a shorter life hurt stock prices?
- It raises the annual depreciation charge, which lowers reported earnings without changing cash, and it lowers the book value of the assets that back debt. For a neocloud it can also mean the contracts end before the hardware is paid off.
- What evidence settles the debate?
- Rental prices. If an older chip still rents for most of its original rate three years on, the long life is justified; if its price has collapsed, the books are optimistic. The spot market publishes that number every day.