Useful-life assumptions for GPU fleets used to be an accounting detail. Now they’re a financing question.
Lenders are underwriting against residual value directly, and NVIDIA has moved to guarantee up to 25% of it to unlock roughly $500 billion in AI infrastructure financing.
Most of the debate so far has run on modeled depreciation curves and asking prices. Our new report, The Productive Life of a Data Center GPU, looks at the observed record instead. It draws on fleet retirement notices, published rental rates, and settled sales from Sprout’s own ledger.
A few things stand out.
- Microsoft ran several GPU generations for 6.9 to 8.8 years before retiring them, and its V100 fleet ran 8.4 years against a six-year book life.
- AWS says it has never retired an A100.
- Rental rates for prior-generation GPUs rose through 2026 instead of falling.
- Six-year-old A100 systems are selling at roughly 38% to 43% of their launch price, above the 25% residual now used in lending.
There’s also real liquidity behind those prices: nine NVIDIA datacenter generations, from 2014 Kepler to 2024 Blackwell, are in our inventory at once, and every one of them has buyers.
The report covers each of these findings with sources, plus a look at how the largest operators have revised their depreciation schedules and what the data doesn't tell you.
