Models · Meta · live prices
Llama 3.1 405B: GPUs needed and cost to self-host
405B parameters. In FP8 that is about 486 GB with room for the KV cache.
Memory, FP8
486 GB
weights + ~20%
Cheapest per 1M output tokens
$1.29
8× A100 · on-demand · 60% utilization
Hosted API, sample list price
$3.50
per 1M output tokens
Cost per 1M output tokens, FP8
Prices observed 12m ago. Cheapest on-demand offer per GPU, whole instances, 60% average utilization.
- A10080 GB 8× on Vast.ai · on-demand · $2.68/hr · ~967 tok/s $1.29per 1M out
- H200141 GB 4× on Vast.ai · on-demand · $5.27/hr · ~1,138 tok/s $2.15per 1M out
- H10080 GB 8× on RunPod · community · $15.92/hr · ~1,588 tok/s $4.64per 1M out
- B300288 GB 2× on RunPod · community · $13.88/hr · ~948 tok/s $6.78per 1M out
- RTX PRO 600096 GB 8× on Vast.ai · on-demand · $8.02/hr · ~510 tok/s $7.28per 1M out
- 409024 GB Doesn’t fit on 8 GPUs
- 509032 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 | BF16972 GB | FP8486 GB | INT4243 GB |
|---|---|---|---|
| H10080 GB | — | 8× | 4× |
| H200141 GB | 8× | 4× | 2× |
| B300288 GB | 4× | 2× | 1× |
| A10080 GB | — | 8× | 4× |
| RTX PRO 600096 GB | — | 8× | 4× |
| 509032 GB | — | — | — |
| 409024 GB | — | — | — |