GPU rental rates for Nvidia's A100, H100, and H200 accelerators continue climbing, with B200 pricing stabilizing near $6 per hour.

GPU depreciation has long been contentious in the AI buildout, with some arguing for longer equipment shelf lives while others maintain that a three-year useful life estimate is most appropriate. The latest hourly rental rate data for NVIDIA GPUs, however, supports the case for extended useful lives—a development favorable to hyperscaler profitability.

Contrary to claims of AI compute overbuild, GPU rental prices are not collapsing as new capacity comes online. Instead, they are rising across nearly every generation. According to Silicon Data, NVIDIA H100 GPUs commanded hourly rental rates just under $2 in January 2026; they now exceed that threshold and approach $3. NVIDIA's B200 GPUs, meanwhile, have climbed from just below $5 per hour in January to between $5.50 and $5.80.

One explanation for this counterintuitive price appreciation lies in rapid model efficiency gains. As AI models become more efficient to run, the same compute hardware can serve more tokens, increasing revenue and economic value per node. This dynamic mirrors the Jevons paradox: as barriers to a technology ease, its uptake increases—and in this case, older NVIDIA GPUs better retain their value in the face of that heightened demand.

The implications are significant. Microsoft CFO Amy Hood noted that "effective at the start of FY 27, we are extending the estimated useful life of our data centers and office buildings from 15 to 25 years." This same logic applies to hyperscalers' GPU inventories. By increasing useful life estimates for NVIDIA hardware, companies can reduce the annual depreciation expense on their income statements, directly boosting net income without a change in actual economic performance.