Renting a data-center GPU got more expensive over the past year. The H100's median on-demand price rose —% to $— an hour, while purchase price-performance keeps doubling about every — years.
H100 on-demand, last 4 weeks
$—
—% since late Oct 2025, per GPU-hour
Median price change, last 12 months
—%
Across — GPUs; —% got cheaper
Rental index, Oct 2026 (Nov 2025 = 100)
—
Computers CPI: — (Sep 2026)
Spot discount vs on-demand
—%
Median, same GPU and week, last 12 weeks
Weekly median across providers, USD per GPU-hour, log scale. Hover for values.
Median of the last 4 weekly prices vs the first 4, on-demand. Data-center GPUs rose; the RTX 4090 fell.
Chain-linked index over GPUs priced every week since Oct 27, 2025, monthly average, Nov 2025 = 100. CPI is the BLS index for computers, peripherals and smart home assistants.
Discount vs the on-demand median for the same GPU and week, last 12 weeks. Bar = median, line = middle 50%. Spot beats every reservation shorter than 3 years.
FP32 GFLOP/s per launch dollar, inflation-adjusted to Sep 2026, log scale. Dashed line: log-linear trend across 54 GPUs.
Hours of use before the launch price beats renting at today's median, after power at $0.15/kWh. Data-center prices exclude the host server.
| GPU | Launch price | Rent $/hr | Break-even hours | Months 24/7 | Months, 8 h weekdays |
|---|
Sources: GetDeploying GPU rental price history and GPU models (CC BY 4.0); Epoch AI, Data on Machine Learning Hardware (CC BY 4.0); U.S. Bureau of Labor Statistics CPI via FRED. Prices through the week of Oct 26, 2026. Part of a data analysis portfolio: Python, DuckDB SQL and statsmodels.