Models · DeepSeek · live prices
DeepSeek V3 (671B MoE): GPUs needed and cost to self-host
671B parameters, 37B active per token (mixture of experts: every expert sits in memory, but only the active ones are read per token, so it decodes like a 37B model). In FP8 that is about 805 GB with room for the KV cache.
Memory, FP8
805 GB
weights + ~20%
Cheapest per 1M output tokens
$0.20
8× H200 · on-demand · 60% utilization
Hosted API, sample list price
$1.10
per 1M output tokens
Cost per 1M output tokens, FP8
Prices observed 11m ago. Cheapest on-demand offer per GPU, whole instances, 60% average utilization.
- H200141 GB 8× on Vast.ai · on-demand · $10.55/hr · ~24,908 tok/s $0.20per 1M out
- B300288 GB 4× on RunPod · community · $27.76/hr · ~20,757 tok/s $0.62per 1M out
- 409024 GB Doesn’t fit on 8 GPUs
- 509032 GB Doesn’t fit on 8 GPUs
- H10080 GB Doesn’t fit on 8 GPUs
- A10080 GB Doesn’t fit on 8 GPUs
- RTX PRO 600096 GB Doesn’t fit on 8 GPUs
The tick marks the sample hosted-API price. Estimates: tokens/s from memory bandwidth with batched serving, not a benchmark. Idle hours still bill, so lower utilization means pricier tokens.
GPUs per replica by precision
One node, split 1, 2, 4 or 8 ways; 90% of VRAM usable.
| GPU | BF161610 GB | FP8805 GB | INT4403 GB |
|---|---|---|---|
| H10080 GB | — | — | 8× |
| H200141 GB | — | 8× | 4× |
| B300288 GB | 8× | 4× | 2× |
| A10080 GB | — | — | 8× |
| RTX PRO 600096 GB | — | — | 8× |
| 509032 GB | — | — | — |
| 409024 GB | — | — | — |