The used GPU market is where most local-AI budgets actually get decided. A used RTX 3090 delivers 24GB of VRAM for $600-$900, while new NVIDIA cards with 24GB start around $2,000 (RTX 5090) and a used RTX 4090 runs $1,400-1,600. For the VRAM-driven decision that defines local AI, that gap is the whole thesis of this page. One 2026 wrinkle: the DRAM shortage has pushed memory prices up across the board, and used cards have risen with them. All prices below are bands, and they move weekly.
Why 24GB used is the VRAM-per-dollar king
If you’ve read how much VRAM you need, you know the throughline: 24GB is the sweet spot — the tier where 27B-30B models run at good quantization with real context, and you stop fighting memory pressure. The used market is the cheapest way into that tier.
Concretely, a 24GB card comfortably fits:
- 30B-class MoE models like Qwen3-Coder-30B-A3B at Q4_K_M — we measured 18.1GB VRAM used at 8K context on our Arc B60 Pro in bm-001, and 22.6GB at Q6_K in our quantization sweep. Both fit.
- 27B dense models like Qwen3.6-27B at Q4_K_M — 21.24GB measured in bm-013.
- 70B models at Q3 as a stretch, with partial CPU offload — usable, not ideal.
The cheapest new 24GB card is Intel’s Arc Pro B60 at ~$450-500 — but it’s Vulkan-only, no CUDA. If CUDA matters, the used 3090 undercuts every other 24GB option: the RX 7900 XTX at ~$900 (AMD) and the used RTX 4090 at $1,400-1,600. That’s why it remains the default budget-24GB recommendation, and why we wrote a full used RTX 3090 deep-dive with the complete buying checklist.
The RTX 3090 in 2026: what $600-$900 buys
At $600-$900 (moves weekly), a used 3090 buys the 24GB tier, CUDA compatibility with every local-AI tool, and a 350W card that runs hot and needs a serious PSU.
We don’t run 3090s on our benchmark fleet. Our four lab machines span Intel Arc (B60 Pro), AMD (Radeon AI PRO R9700, RX 7900 XT), and NVIDIA Blackwell (RTX 5070 laptop) — see the fleet. What we can speak to from direct experience is the 24GB tier itself: we’ve benchmarked it on Arc B60 Pro and know exactly what fits. The VRAM math transfers; the 3090’s specific tok/s figures are community-reported, not ours.
Mining-card caveats are real. Used 3090s are overwhelmingly ex-mining cards. That’s not automatically bad — mining operators often repadded the VRAM, which is actually better for sustained AI loads — but it means you verify before buying. Short version: check VRAM thermal pads, fan health, seller reputation, and confirm the card computes under load inside the return window. Full checklist in the 3090 deep-dive.
The 12GB sibling warning. The RTX 3080 looks like a cheaper sibling, but its 10GB VRAM is the binding constraint for local AI — and even the 3080 Ti’s 12GB falls short of the 24GB tier. If your workload fits in 12GB, a 12GB starter is the cheaper path; if you need 24GB, the 3090 is the used-market answer. Check Model Fit before assuming you need 24GB.
Other used candidates worth a look
| Card | Price band (moves weekly) | VRAM | Verdict |
|---|---|---|---|
| Used RTX 3090 | $600-$900 | 24GB | The value king. CUDA, ex-mining caveats. |
| Used RTX 4090 | $1,400-$1,600 | 24GB | Faster, but pricier per GB. |
| Used RX 7900 XTX | ~$900 | 24GB | AMD path to 24GB; ROCm/Vulkan setup cost. |
| Used RTX 3080 | varies | 10GB | VRAM-limited; only if 10GB fits your models. |
| Used RTX A4000 | varies | 16GB | Blower pro card; quiet, long-duty-cycle design. |
| BC250 crypto APU | ~$150 | shared | Community-reported FLUX.2-klein runs; unverified by us. |
Used RTX 4090 ($1,400-$1,600). Faster than a 3090, but roughly double the price for the same 24GB — worse per-GB math. You’re buying speed, not capacity. If your workload is VRAM-bound, the 3090 wins on value; if you need speed for image generation or long-context prompt processing, the 4090 is the used-market speed pick.
Used RTX 3080 (10GB). VRAM-limited. Only if 10GB genuinely fits your models — check Model Fit first. For most 2026 local-AI workloads, 10GB is tight.
Used RTX A4000 (16GB). A blower-style pro card designed for long duty cycles and quiet operation. Spec-sheet analysis — we have not measured this hardware. It sits between the 12GB starter tier and 24GB; worth considering if 16GB fits your models.
BC250 crypto APU (~$150). Community reports of FLUX.2-klein image generation on these repurposed mining APUs circulate regularly. Community-reported — we have not verified this in our lab. A curiosity-tier experiment, not a recommendation.
The AMD/Intel used angle
- Used RX 7900 XTX (~$900). 24GB on the AMD side. We run an RX 7900 XT (20GB) in our lab and measured 25.21 tok/s on Qwen3.6-27B at Q4_K_M in bm-013 — a serious inference host; the XTX adds 4GB. The tradeoff is the ROCm/Vulkan software tax versus CUDA’s frictionless path.
- Arc B580 new at used prices ($300-$330). Intel’s Battlemage card is new, warrantied, and priced like a used card. Its 12GB is the starter tier — for reference, we measured the 24GB Arc Pro B60 at 38.6 tok/s on Qwen3-Coder-30B-A3B Q4_K_M in bm-001. Full cross-vendor picture: best GPUs for local AI.
The risks checklist
- VRAM thermal pads. Mining runs memory hot, and the 3090’s GDDR6X modules run very hot on stock pads. Ask whether the card was repadded; a repadded card is actually better for sustained AI loads.
- No warranty. Most used cards are out of warranty. Buy from sellers with returns accepted and tested-listing language; eBay’s money-back guarantee and dedicated GPU resellers are safer than no-returns private listings.
- PSU requirements. The 3090’s 350W draw needs 850W minimum for a single card, more for two — budget for a quality 850W PSU alongside the card; it’s not optional.
- Verify it actually computes. On arrival: run a sustained llama.cpp generation, watch temps and VRAM with
nvidia-smi -l 1, and confirm 24GB is addressable with no artifacting or throttling. Do this inside the return window.
When NEW beats USED in 2026
The DRAM shortage (“rampocalypse”) has pushed memory prices up across the board — and used GPU prices have risen with them. The 3090’s $600-$900 band is higher than a year ago, and the gap to new cards has narrowed. The honest answer to buy-now-or-wait: the used premium has grown, so the used discount has shrunk.
Buy new when:
- You want warranty and known history. No detective work, no return-window stress test. If your time matters more than the savings, buy new.
- The used discount has shrunk too far. If a used 3090 is within ~$200 of a new 24GB card, the risk stops justifying the savings. The Arc Pro B60 at ~$450-500 is the new-card 24GB path worth pricing against it.
- Power or heat is a constraint. The 3090’s 350W draw is a real cost — in electricity (see our cost analysis) and in the cooling and PSU hardware around it. Newer cards do more with less power.
So is the 3090 still the value king? Yes — on raw VRAM-per-dollar, it remains the used-market default for the 24GB tier. But the margin is thinner than it was. Price both paths the week you buy.
Where this data comes from
- bm-001 — Qwen3-Coder-30B-A3B on Arc B60 Pro vs. dual Radeon AI PRO R9700 (18.1GB VRAM at Q4_K_M, 38.6 tok/s)
- bm-003 — Q4_K_M vs. Q6_K vs. Q8_0 quantization sweep on Arc B60 Pro (22.6GB at Q6_K)
- bm-013 — Qwen3.6-27B cross-machine baseline (21.24GB, 25.21 tok/s on RX 7900 XT)
- Used RTX 3090 deep-dive — the full buying checklist for ex-mining cards
Pricing reflects the US used market as of October 2026. Prices are bands and move weekly; treat ranges as directional, not a quote. Community-reported claims are labeled as such and have not been verified in our lab.