Abhay Chakra Sadineni

Research paper · November 2026

The Economic Life of an Accelerator

Abhay Chakra Sadineni

GPU hardwareAI infrastructureDepreciationEnergy

Abstract

The largest buyers of data-centre GPUs depreciate them over five to six years. A common critique holds that a new chip every year makes three years more realistic. Both sides argue from the release calendar. This paper argues from prices. It sets public rental indices for the NVIDIA A100, H100, H200 and B200 against each chip's arithmetic throughput, memory bandwidth, memory capacity and system power, and asks when an owner should stop running a chip. Three results follow. First, older chips rent well above arithmetic parity with the newest one. On 26 October 2026 the six-year-old A100 rented for 70 percent more than the B200's price per unit of dense BF16 throughput would imply. Second, once launch scarcity passed, rents fell slowly, by about 13 percent a year for the A100 since April 2023 and 9 percent a year for the H100 since October 2024. Third, the decision to retire a chip turns on power. Where power and space can be bought at market prices, an A100 earns about five times its operating cost before idle hours, and operating cost alone would not retire it before age eight even if rents fell 40 percent a year. Where power is the binding constraint, the answer turns on how fast the replacement's own rent falls. If B200 rents fall 9.4 percent a year, as H100 rents have, the A100 sits at the replacement threshold now, at six and a half years. If they hold flat, both the A100 and the H100 should already be replaced. If they fall 13 percent a year or more, both remain economic. The six-year schedule is therefore plausible for power-limited sites and conservative elsewhere, and its accuracy depends on a rent path no one yet observes. The paper closes with what this means for lenders who take GPUs as collateral.

Cite

Abhay Chakra Sadineni (2026). The Economic Life of an Accelerator. Research paper. https://doi.org/10.5281/zenodo.24210935 https://www.abhaychakra.com/papers/economic-life-of-an-accelerator/