Blog · Hardware Buying Guide · September 20, 2026 · 8 min read

Buying a Used GPU for AI: The 10-Point Checklist

Used cards are the best value in local AI — a secondhand 24 GB card often costs less than a new 16 GB one. But AI is a harsher owner than gaming: hours pinned at maximum memory and power. Used mining-era cards, for instance, ran 24/7 but usually at gentle memory clocks, while used gaming cards ran cool but hard. Either way, inspect before you pay.

The checklist

• 1. Test VRAM first — it is everything for AI. Ask to run a VRAM stress/scan tool for 15–30 minutes. VRAM errors show up as artifacts, crashes, or tool-reported failures. A card with faulty memory is worthless for inference even if games run fine.
• 2. Check memory size against your models, not vibes. Bring your target model list: our compatibility panel tells you the exact VRAM class you need before you negotiate.
• 3. Look for reflow/reball marks and warped boards. Discoloration around memory chips, bent PCBs, or missing screws suggest a hard life or a bad repair.
• 4. Inspect the fans. Spin each one by hand: grinding or free-spinning (no resistance) means worn bearings. Replacements exist but cost time and money.
• 5. Demand a live temperature test. Ten minutes of a heavy benchmark: you want core and memory temperatures that are stable and sensible, not throttling into the 90s °C.
• 6. Listen under load. Coil whine is common and cosmetic; rattling fans and clicking are not.
• 7. Match the serial number to the receipt/photos. Standard anti-swap practice for in-person deals.
• 8. Prefer sellers with return windows. Some VRAM faults only appear after an hour of load — a 24–48 hour return option is worth a slightly higher price.
• 9. Research the card’s history model. ex-mining cards can be fine (check memory and fans harder); ex-server compute cards often have monster durability but need adapters or lack video outputs. Know which you are buying.
• 10. Price it against new alternatives. If a used card costs nearly as much as a new one with the same VRAM and better efficiency, the used deal is not a deal. Compare bandwidth and VRAM on our Hardware panel side by side.

Red flags worth walking away from

• “Works fine, but I can’t demo it” — the single most common scam pattern.
• Prices dramatically below market “because I need it gone today.”
• Thermal pad replacement with clearly mismatched pads or a warped backplate — a symptom of a previous overheating failure.

Do it right and a used card is the cheapest ticket into 24 GB-class local AI. Do it wrong and you have bought a very expensive space heater. Check, then pay.