Table of Contents
Llama 3 is the most-deployed open-weight LLM globally, and the 8B-class fits comfortably on a single dedicated GPU while the 70B-class lives on the boundary between single-card and multi-card territory. This page is the cost-per-token reference for both sizes.
For Llama 3.1 8B: RTX 5080 at FP8 = £0.10/1M tokens is the cost leader; RTX 5090 ties at higher absolute throughput. For Llama 3.3 70B: a single RTX 6000 Pro at FP8 = £1.05/1M tokens, or 2× RTX 5090 at INT4 = £0.95/1M tokens. Hosted Llama 3 70B APIs (Together, Fireworks) sit at £0.50–0.70/1M but only win below ~£700/mo of usage.
Methodology
Identical method to our Mistral cost page: 60% utilisation, vLLM 0.6.3 aggregate throughput, 50-thread Locust driver. Llama 3.1 8B Instruct from meta-llama/Llama-3.1-8B-Instruct; Llama 3.3 70B from meta-llama/Llama-3.3-70B-Instruct (and AWQ-INT4 community port for the consumer-GPU paths).
Llama 3.1 8B cost-per-1M
| GPU | Monthly | tok/s (FP16) | Cost per 1M (FP16) | tok/s (FP8) | Cost per 1M (FP8) |
|---|---|---|---|---|---|
| RTX 5060 Ti 16 GB | £119 | 550 | £0.20 | 820 | £0.13 |
| RTX 5080 | £189 | 780 | £0.19 | 1,210 | £0.10 |
| RTX 3090 | £159 | 680 | £0.17 | n/a (no FP8) | n/a |
| RTX 4090 | £289 | 910 | £0.20 | n/a (no native) | n/a |
| RTX 5090 | £399 | 1,140 | £0.20 | 1,820 | £0.13 |
| RTX 6000 Pro | £899 | 1,110 | £0.64 | 1,790 | £0.40 |
RTX 5080 + FP8 is the cost leader for 8B-class deployments.
Llama 3.3 70B cost-per-1M
70B does not fit on a single consumer GPU at FP16 — the deployment options are a single 6000 Pro, multi-GPU clusters, or A100. The cost picture changes:
| Config | Monthly | tok/s | Cost per 1M | Notes |
|---|---|---|---|---|
| 1× RTX 5090 INT3 | £399 | 130 | £0.89 | Quality risk on hard tasks |
| 1× RTX 6000 Pro FP8 | £899 | 220 | £1.61 | Single-card, full quality |
| 2× RTX 5090 (INT4 TP=2) | £899 | 240 | £0.95 | Best cost-per-token at 70B |
| 2× A100 80 GB FP16 | POA | 180 | POA | Reference quality |
| Together AI Llama 3.3 70B | n/a | n/a | £0.66 | Hosted, per-token |
| Fireworks Llama 3.3 70B | n/a | n/a | £0.71 | Hosted, per-token |
The 2× RTX 5090 cluster at INT4 is the clear cost leader for self-hosted 70B at £0.78/1M. The 6000 Pro is 70% more expensive per token but considerably simpler operationally (single card, no NCCL).
Self-hosted vs hosted Llama 3 APIs
The hosted APIs are cheaper per token than self-hosting at 60% utilisation. They beat self-hosting whenever your monthly token volume × hosted-price < your monthly server cost.
Concrete break-even (using £0.73/1M Together pricing):
- RTX 5090 single (INT3 70B): 359 / 0.66 = ~543M tokens/month break-even
- 2× RTX 5090 cluster: 899 / 0.66 = ~1.36B tokens/month break-even
- RTX 6000 Pro: 1099 / 0.66 = ~1.66B tokens/month break-even
Below those thresholds, hosted is cheaper. Above them, self-hosted is dramatically cheaper. Most enterprise deployments serving 100+ active users blow past the break-even within their first month.
Break-even token volume
| If your monthly Llama 3.3 70B usage is… | Cheapest option | Why |
|---|---|---|
| < 200M tokens | Together AI hosted API | No server cost, low utilisation |
| 200M – 1B tokens | Together API still wins | Self-hosted server idle most of month |
| 1B – 2B tokens | 2× RTX 5090 cluster | Break-even crossed, self-hosted cheaper |
| 2B+ tokens | Self-hosted, definitely | Hosted bill scales linearly |
| Need data residency / compliance | Self-hosted regardless | Cost is not the variable |
Verdict
- Llama 3.1 8B cost leader: RTX 5080 + FP8 at £0.11/1M. RTX 5090 ties on cost with 50% more headroom.
- Llama 3.3 70B cost leader: 2× RTX 5090 at AWQ-INT4 at £1.06/1M. Worth the operational complexity if you will use the throughput.
- If volume < 1B tokens/mo of 70B: hosted Together / Fireworks is cheaper.
- If you need ECC / certified drivers: 6000 Pro at £1.32/1M, premium worth paying.
Bottom line
The interesting line is 1B tokens of 70B output per month. Below that, hosted APIs win on cost. Above that, the 2× RTX 5090 cluster is meaningfully cheaper. For 8B, self-hosting wins at almost any meaningful volume — see can RTX 5090 run Llama 3 70B INT4? for the deployment specifics.
