GPU comparison · live prices
H100 vs H200: price per hour and cost per token
The default for serving 70B-class models: one node of 8 fits Llama 3.1 405B in FP8. H100 compute with 76% more memory: fewer GPUs per model and longer contexts per GPU.
Prices observed 11m ago · refreshed every 15 minutes
| NVIDIA H100 | NVIDIA H200 | |
|---|---|---|
| Memory | 80 GB HBM3 | 141 GB HBM3e |
| Memory bandwidth | 3.35 TB/s | 4.8 TB/s |
| Cheapest on-demandper GPU-hour | $2.20Vast.ai | $2.63Vast.ai |
| Cheapest spotper GPU-hour | $1.00Nebius · availability not published | $0.79Nebius · availability not published |
| Providers in stock | 4 | 1 |
Per GPU-hour on demand, the H200 costs 1.20× the H100. Per token, that depends on the model:
Cost per 1M output tokens
Cheapest on-demand offer, GPUs per replica to fit the model, 60% average utilization.
H200 is cheaper per token for every model below that runs on both at today’s prices.
| Model | H100 | H200 |
|---|---|---|
| Llama 3.1 8BBF16 | $0.18 1× GPUs · ~5,025 tok/s | $0.17 1× GPUs · ~7,200 tok/s |
| Qwen 2.5 32BFP8 | $0.37 1× GPUs · ~2,474 tok/s | $0.34 1× GPUs · ~3,545 tok/s |
| Llama 3.3 70BFP8 | $0.81 2× GPUs · ~2,278 tok/s | $0.75 1× GPUs · ~1,632 tok/s |
| Mixtral 8x22BFP8 | $0.45 4× GPUs · ~8,246 tok/s | $0.41 2× GPUs · ~5,908 tok/s |
| Llama 3.1 405BFP8 | $4.64 8× GPUs · ~1,588 tok/s | $4.29 4× GPUs · ~1,138 tok/s |
| DeepSeek V3 (671B MoE)FP8 | — Doesn’t fit on 8 | $0.39 8× GPUs · ~24,908 tok/s |
Estimates: tokens/s from memory bandwidth with batched serving, not a benchmark. Check each provider before you buy.