Few prices in technology moved as much as GPU rental rates over the last three years. Understanding the cycle helps you decide when to lock in a rate and when to stay flexible.
2023–2024: scarcity
After ChatGPT, demand for H100s outran supply. Hyperscaler on-demand rates sat around $7–8 per GPU-hour and lead times for clusters ran to months.
2024–2025: the glut
New datacenters came online, and many buyers who had reserved capacity found they needed less of it. Marketplace rates fell steadily, and one-year contract prices reached a low of about $1.70 per GPU-hour in October 2025, according to SemiAnalysis.
2026: tight again
Inference and media-generation demand absorbed the surplus faster than expected. By March 2026 roughly half of providers had no on-demand H100 stock, and capacity coming online through the summer was booked in advance. The Silicon Data H100 rental index stood at $3.06 in September 2026.
What it means for buyers
- Lock rates for steady workloads. When the market is tight, a reservation buys availability as much as a discount.
- Keep bursts flexible. On-demand capacity for experiments avoids committing at the top of a cycle.
- Don't buy hardware to escape the cycle. Owners of GPUs bought in 2023 watched their rental value fall by about three quarters before it recovered.
Sources: SemiAnalysis, Silicon Data, public provider pricing.
