GPU comparison · live prices
B300 vs 4090: price per hour and cost per token
Blackwell Ultra: twice the H200’s memory, so one node of 8 holds the largest open models with room for long contexts. The cheapest way to serve 7B–14B models, mostly on marketplaces and spot.
Prices observed 12m ago · refreshed every 15 minutes
| NVIDIA B300 | NVIDIA RTX 4090 | |
|---|---|---|
| Memory | 288 GB HBM3e | 24 GB GDDR6X |
| Memory bandwidth | 8.0 TB/s | 1.0 TB/s |
| Cheapest on-demandper GPU-hour | $8.99RunPod | $0.31Vast.ai |
| Cheapest spotper GPU-hour | $9.30Nebius · availability not published | —None in stock |
| Providers in stock | 2 | 3 |
Per GPU-hour on demand, the 4090 costs 0.03× the B300. 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.
B300 is cheaper per token for 2 of these models and 4090 for 4: it depends on model size.
| Model | B300 | 4090 |
|---|---|---|
| Llama 3.1 8BBF16 | $0.27 1× GPUs · ~12,000 tok/s | $0.10 1× GPUs · ~1,512 tok/s |
| Qwen 2.5 32BFP8 | $0.54 1× GPUs · ~5,908 tok/s | $0.32 2× GPUs · ~893 tok/s |
| Llama 3.3 70BFP8 | $1.18 1× GPUs · ~2,720 tok/s | $0.70 4× GPUs · ~822 tok/s |
| Mixtral 8x22BFP8 | $0.65 1× GPUs · ~4,923 tok/s | $0.39 8× GPUs · ~2,977 tok/s |
| Llama 3.1 405BFP8 | $6.78 2× GPUs · ~948 tok/s | — Doesn’t fit on 8 |
| DeepSeek V3 (671B MoE)FP8 | $0.62 4× GPUs · ~20,757 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.