Sold prices / best / local AI
Best used GPU for local AI
The cheapest VRAM in a used desktop card with 12GB or more right now is the Quadro P6000 24GB: 24GB for a median $410 on US eBay, or $17 per GB. The cheapest 24GB card, enough for 32B models at 4-bit, is the Quadro P6000 24GB at $410. Prices as of 2026-11-12.
RTX 2060 Super: flat over the past week.
Cheapest card at each VRAM size
VRAM decides how big a model fits. Speed matters less than fitting at all.
| VRAM | Card | Sold median | Runs at 4-bit | Buy |
|---|---|---|---|---|
| 8GB+ | RTX 2060 SuperOr Tesla P4 8GB at $113, a fanless server card | $190 | 8B models | Find on eBay |
| 12GB+ | RTX 2060 12GBOr Tesla P100 16GB at $159, a fanless server card | $243 | 14B models | Find on eBay |
| 16GB+ | RTX 4060 Ti 16GBOr Tesla P100 16GB at $159, a fanless server card | $661 | 14B models | Buy $650 |
| 24GB+ | Quadro P6000 24GBOr Tesla P40 24GB at $293, a fanless server card | $410 | 32B models | Find on eBay |
| 32GB+ | RTX 5000 AdaOr Tesla V100 32GB at $858, a fanless server card | $3,430 | 32B models | Find on eBay |
| 48GB+ | Quadro RTX 8000 | $2,000 | 70B models | Find on eBay |
| 80GB+ | RTX Pro 6000 Blackwell 96GB | $15,515 | 70B models | Find on eBay |
All NVIDIA cards by price per GB of VRAM
| Card | VRAM | Sold median | Per GB |
|---|---|---|---|
| Tesla P100 16GB44 sales · server card, no fan | 16GB | $159 | $10 |
| Tesla P40 24GB16 sales · server card, no fan | 24GB | $293 | $12 |
| Tesla P4 8GB22 sales · server card, no fan | 8GB | $113 | $14 |
| Quadro P6000 24GB13 sales | 24GB | $410 | $17 |
| RTX 2060 12GB13 sales | 12GB | $243 | $20 |
| Tesla V100 16GB14 sales · server card, no fan | 16GB | $327 | $20 |
| RTX 2060 Super48 sales | 8GB | $190 | $24 |
| RTX 207062 sales | 8GB | $190 | $24 |
| RTX 3050 8GB44 sales | 8GB | $195 | $24 |
| Tesla V100 32GB14 sales · server card, no fan | 32GB | $858 | $27 |
| RTX 2070 Super72 sales | 8GB | $215 | $27 |
| RTX 208039 sales | 8GB | $220 | $28 |
| RTX 2080 Ti89 sales | 11GB | $306 | $28 |
| Tesla T4 16GB17 sales · server card, no fan | 16GB | $450 | $28 |
| RTX 3060 12GB254 sales | 12GB | $340 | $28 |
| RTX 2080 Super44 sales | 8GB | $244 | $31 |
| RTX 3060 8GB16 sales | 8GB | $260 | $33 |
| RTX 3060 Ti163 sales | 8GB | $281 | $35 |
| RTX 3070200 sales | 8GB | $301 | $38 |
| Titan RTX10 sales | 24GB | $908 | $38 |
| RTX 505011 sales | 8GB | $309 | $39 |
| Quadro RTX 600014 sales | 24GB | $972 | $41 |
| RTX 4060 Ti 16GB30 sales | 16GB | $661 | $41 |
| RTX A2000 12GB24 sales | 12GB | $500 | $42 |
| Quadro RTX 800014 sales | 48GB | $2,000 | $42 |
NVIDIA only, because CUDA is what most local AI tools support best. Older Tesla and Pascal cards are cheap per GB but slow and need extra cooling. Model sizes assume 4-bit weights with room for context.
Complete local AI builds with every part priced
Prices are medians of completed US eBay sales over the last 14 days (30 when sales are thin), shipping included, with broken, parts-only, and bundle listings removed. Cards with too few sales or a wide spread are left out. How we price