Home › Budget builds › Used GPUs
Used GPUs for local LLM — break-even, not hype
Sourced price anchors (not street prices)
- RTX 4090 launch MSRP $1,599 (2022-10-12) — Wikipedia GeForce RTX 40 series, retrieved 2026-10-20
- RTX 5090 launch MSRP $1,999 — Wikipedia GeForce RTX 50 series, retrieved 2026-10-20
- RTX 3090 launch MSRP $1,499 (2020-09-24) — Wikipedia GeForce RTX 30 series spec table, retrieved 2026-10-20. A 2020 launch price is a floor for the story, not today’s used price: the 3090 is the card the whole used-for-LLM trade runs on, and its MSRP is why a sub-$1,000 used card can pay back so fast.
Street prices: enter them into the calculator; check eBay searches for RTX 3090, RTX 4090. We intentionally pre-fill nothing.
Payback at cited rental rates (hours of use until buying matches renting)
| rental alternative | $/h | if card = $900 | if card = $1,599 |
|---|---|---|---|
| RunPod RTX 5090 32GB | $0.99 | 909 | 1615 |
| Lambda A100 40GB | $1.99 | 452 | 803 |
| RunPod A100 80GB | $1.59 | 566 | 1005 |
| RunPod H100 80GB | $3.49 | 257 | 458 |
| Lambda H100 80GB | $3.29 | 273 | 486 |
| RunPod H200 141GB | $4.59 | 196 | 348 |
modelvram.com public table (self-stated check 2026-09-29, retrieved 2026-10-20). Partial re-verify 2026-10-20: Lambda A100 40GB $1.99/h CONFIRMED on lambda.ai/pricing (fetched directly); Lambda H100 rows ambiguous on-page ($3.99 vs $4.09 configs) → not adopted; RunPod & Vast pricing pages are JS-rendered, no extractable static figures — verify rates visually before spending. Electricity offset: subtract your GPU's $/h power cost (enter in calculator).
FAQ
Which used GPU is the best value for local LLMs?
VRAM per dollar decides it — 24GB cards (3090/4090 class) fit most 20–32B q4 runs; the calculator's break-even answer beats any static 'best' list because street prices and rental rates both move.
How fast does a used card pay for itself vs renting?
break-even hours = street price ÷ (rental $/h − your GPU-hour electricity). With cited rates (A100 80GB $1.59/h, RTX 5090 $0.99/h; modelvram.com public table (self-stated check 2026-09-29, retrieved 2026-10-20). Partial re-verify 2026-10-20: Lambda A100 40GB $1.99/h CONFIRMED on lambda.ai/pricing (fetched directly); Lambda H100 rows ambiguous on-page ($3.99 vs $4.09 configs) → not adopted; RunPod & Vast pricing pages are JS-rendered, no extractable static figures — verify rates visually before spending) a $900 card vs 5090 rental pays back in ≈909 GPU-hours of usage. Enter your real numbers — we don't invent street prices.
Is buying used risky for this?
Card health/memory errors matter; factor a 6–12 month warranty or return window into the price you'll pay.