the $549 new GPU or the $1,000 used one? for local AI, the old card wins — and it's not close

The short answer

For raw local-AI capability per dollar, a used RTX 3090 (24GB, ~$1,000) beats a new RTX 5060 Ti 16GB (~$549): the 3090's 24GB runs 32B-class models the 16GB card can't hold at all, and it's faster…

VRAM is the whole game. 24GB runs Qwen3-32B at Q4 (~30 t/s measured); 16GB tops out around 14B-class models. That capacity gap can't be bought back with speed.

On models both can run, the 3090 is faster: 87.5 vs 69.2 t/s on an 8B (Q4, 4k context) in hardware-corner's measurements.

The 5060 Ti's real wins: a warranty, ~180W vs ~350W power (cheaper to run, easier to cool/power), and no mining-history gamble.

Price-per-token/sec actually favors the 5060 Ti ($16.68 vs $19.18 on a 14B) — but only because it's cheaper, not because it does more. It can't touch the 3090's ceiling.

Buy the 5060 Ti if your models are ≤14B and you value warranty/efficiency; buy the 3090 if you want to run 32B-class models at home, which 16GB simply can't.