Models · Meta · live prices
Llama 3.3 70B: GPUs needed and cost to self-host
70.6B parameters. In FP8 that is about 85 GB with room for the KV cache.
Memory, FP8
85 GB
weights + ~20%
Cheapest per 1M output tokens
$0.22
2× A100 · on-demand · 60% utilization
Hosted API, sample list price
$0.88
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 2× on Vast.ai · on-demand · $0.67/hr · ~1,386 tok/s $0.22per 1M out
- H200141 GB 1× on Vast.ai · on-demand · $1.32/hr · ~1,632 tok/s $0.37per 1M out
- 509032 GB 4× on Vast.ai · on-demand · $1.83/hr · ~1,462 tok/s $0.58per 1M out
- 409024 GB 4× on Vast.ai · on-demand · $1.24/hr · ~822 tok/s $0.70per 1M out
- RTX PRO 600096 GB 1× on Vast.ai · on-demand · $1.00/hr · ~609 tok/s $0.76per 1M out
- H10080 GB 2× on RunPod · community · $3.98/hr · ~2,278 tok/s $0.81per 1M out
- B300288 GB 1× on RunPod · community · $6.94/hr · ~2,720 tok/s $1.18per 1M out
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 | BF16169 GB | FP885 GB | INT442 GB |
|---|---|---|---|
| H10080 GB | 4× | 2× | 1× |
| H200141 GB | 2× | 1× | 1× |
| B300288 GB | 1× | 1× | 1× |
| A10080 GB | 4× | 2× | 1× |
| RTX PRO 600096 GB | 2× | 1× | 1× |
| 509032 GB | 8× | 4× | 2× |
| 409024 GB | 8× | 4× | 2× |