| GPU model | Average pay-as-you-go price | Average spot price | Price difference |
|---|---|---|---|
| RTX 4090 | ¥1.75/hour | ¥1.41/hour | 20% lower |
| A100 80GB | ¥10.79/hour | ¥7.61/hour | 29% lower |
| H100 80GB | ¥23.12/hour | ¥14.17/hour | 39% lower |
| H200 141GB | ¥31.27/hour | ¥21.39/hour | 32% lower |
For online inference, stateful services, or training tasks with tight deadlines, avoid spot/preemptible instances—interruption costs often far exceed savings.
Always enable checkpointing (save model every N steps) in your training script and use a scheduler that supports automatic restarts; offload one-time tasks like data preprocessing to pay-as-you-go instances.
rentgpu.link（GPU算力雷达）《What’s the price difference between spot and pay-as-you-go GPU instances?》，Data更新 2026-10-28，https://rentgpu.link/answers/spot-vs-ondemand（CC BY 4.0）
rentgpu.link
A neutral GPU compute rental price comparison tool: Public prices from 26 global platforms are collected every 45 minutes. We own no GPUs and sell no compute.