GPU comparison · live prices
H200 vs 5090: price per hour and cost per token
H100 compute with 76% more memory: fewer GPUs per model and longer contexts per GPU. Consumer card, cheap per hour; great for small models, slow when split across many.
Prices observed 12m ago · refreshed every 15 minutes
| NVIDIA H200 | NVIDIA RTX 5090 | |
|---|---|---|
| Memory | 141 GB HBM3e | 32 GB GDDR7 |
| Memory bandwidth | 4.8 TB/s | 1.8 TB/s |
| Cheapest on-demandper GPU-hour | $1.32Vast.ai | $0.46Vast.ai |
| Cheapest spotper GPU-hour | $0.79Nebius · availability not published | —None in stock |
| Providers in stock | 2 | 3 |
Per GPU-hour on demand, the 5090 costs 0.35× the H200. 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 5 of these models and 5090 for 1: it depends on model size.
| Model | H200 | 5090 |
|---|---|---|
| Llama 3.1 8BBF16 | $0.08 1× GPUs · ~7,200 tok/s | $0.08 1× GPUs · ~2,688 tok/s |
| Qwen 2.5 32BFP8 | $0.17 1× GPUs · ~3,545 tok/s | $0.27 2× GPUs · ~1,588 tok/s |
| Llama 3.3 70BFP8 | $0.37 1× GPUs · ~1,632 tok/s | $0.58 4× GPUs · ~1,462 tok/s |
| Mixtral 8x22BFP8 | $0.21 2× GPUs · ~5,908 tok/s | $0.32 8× GPUs · ~5,293 tok/s |
| Llama 3.1 405BFP8 | $2.15 4× GPUs · ~1,138 tok/s | — Doesn’t fit on 8 |
| DeepSeek V3 (671B MoE)FP8 | $0.20 8× GPUs · ~24,908 tok/s | — Doesn’t fit on 8 |
Estimates: tokens/s from memory bandwidth with batched serving, not a benchmark. Check each provider before you buy.