Cheapest hardware to run a local LLM
Ranked by measured cost per million tokens, which already includes electricity and hardware spread over five years. The cheapest to run is not always the cheapest to buy.
| # | Hardware | VRAM | Read tok/s | Power | Full rig | Cost/M tok | Payback |
|---|---|---|---|---|---|---|---|
| 1 | RTX 4090 | 24 GB | 4,000 | 550 W | $4,199 | $0.056 | 434 days |
| 2 | RTX 5080 | 16 GB | 3,713 | 510 W | $2,800 | $0.058 | 393 days |
| 3 | RTX 5070 Ti | 16 GB | 2,499 | 450 W | $2,380 | $0.060 | 403 days |
| 4 | RTX 3090 | 24 GB | 2,094 | 500 W | $2,700 | $0.060 | 410 days |
| 5 | RTX 4070 Ti SUPER 16GB | 16 GB | 1,700 | 435 W | $2,140 | $0.061 | 399 days |
| 6 | RTX 5090 | 32 GB | 5,803 | 700 W | $8,500 | $0.062 | 558 days |
What these lists do not cover
- Cheapest per token assumes the machine runs near full time. At a few hours a day, a cheaper rig with worse throughput can win on total spend.
- All figures are single-stream measurements on one workload. Multi-user serving, fine-tuning and server parts behave differently.