For years the advice for anyone building a local AI machine was simple: buy a used RTX 3090. It has 24 GB of VRAM, 936 GB/s of bandwidth and full CUDA support, and it used to sell for $600-$900. In late 2026 the 3090 is still one of the best AI cards around, but the price has more than doubled, and new 16 GB and 32 GB cards have changed the maths. This guide looks at what a used 3090 costs now, what it can run, how it compares with new alternatives, and exactly what to check before you buy one.
- Price: about $1,450 median on eBay US (September 2026), up 13.8% in 30 days. The old “$600-$900” advice is out of date.
- Still the best bandwidth per dollar: 936 GB/s for about $1.56 per GB/s, about twice as fast as an RTX 5060 Ti 16 GB on models that fit both.
- No longer the cheapest VRAM per dollar: the 32 GB Radeon AI PRO R9700 and the 16 GB RTX 5060 Ti cost less per gigabyte.
- Runs: 27B-32B LLMs at Q4, Flux-class and Qwen-Image models at fp8/GGUF, Wan 2.2 video with GGUF. Two of them run 70B at Q4.
- Before buying: stress-test VRAM with memtest_vulkan for 15-30 minutes, watch memory temperatures, and buy where you can return it.
Why the RTX 3090 Is Still an AI Card
Three specs keep the 2020-era RTX 3090 relevant:
- 24 GB of GDDR6X. Enough for 27B-32B language models at Q4 and nearly every image model. Only the RTX 4090, RTX 5090 and pro cards beat it in the GeForce-class world.
- 936 GB/s of bandwidth. Text generation speed depends mainly on bandwidth, and 936 GB/s is close to an RTX 5080 (960 GB/s) and more than twice an RTX 5060 Ti (448 GB/s).
- CUDA and Ampere tensor cores. Every AI tool supports it: Ollama, llama.cpp, vLLM, LM Studio, ComfyUI, kohya and Unsloth. It also supports BF16, which modern models use.
It is also the last GeForce card with an NVLink connector, which some multi-GPU training setups can still use.
Used RTX 3090 Prices in 2026
OpenClaw analysed 291 eBay US listings on 24 September 2026:
| Measure | Price |
|---|---|
| Average | $1,463 |
| Median | $1,450 |
| Fair asking range | $1,398-$1,502 |
| 90-day low | $1,244 |
| 30-day change | +13.8% |
That is roughly the 3090’s original $1,499 launch price, six years later. The reason is the 2026 GPU and memory shortage: new cards with 24 GB or more cost far more (the cheapest RTX 5090 was about $6,900 in early November), so demand for used 24 GB cards keeps rising. I explain the background in why GPU prices are so high in 2026.
RTX 3090 vs New Alternatives
| Card | Price | VRAM | Bandwidth | $ per GB | $ per GB/s |
|---|---|---|---|---|---|
| RTX 3090 (used) | about $1,463 | 24 GB | 936 GB/s | $61 | $1.56 |
| RTX 5060 Ti 16 GB (new) | about $770-$790 | 16 GB | 448 GB/s | $49 | $1.76 |
| Radeon AI PRO R9700 (new) | $1,299 MSRP; $1,700-$1,850 street | 32 GB | 640 GB/s | $41-$58 | $2.03-$2.89 |
| RTX 5090 (new) | about $6,900 | 32 GB | 1,792 GB/s | $216 | $3.85 |
Per-GB figures for the 3090, 5060 Ti and R9700 at MSRP are from OpenClaw’s analysis; I recalculated the R9700 at street price and the RTX 5090 at November prices. The picture: if you want the most speed per dollar, the 3090 still wins. If you want the most memory per dollar, the R9700 at a good price or a 5060 Ti wins. Read the best GPU for local LLMs for the full ranking.
What a 3090 Can Run
| Workload | On one RTX 3090 |
|---|---|
| 8B-14B LLMs | Q8 with long context, fast |
| 27B-32B LLMs (Qwen3.8 27B, Gemma 4 31B) | Q4_K_M with 8K-32K context |
| 30-35B MoE (Qwen3.5-35B-A3B) | Q4 fully on GPU, very fast |
| gpt-oss-120b and 100B+ MoE | With 64 GB+ system RAM and expert offload |
| FLUX 3 Klein, Qwen-Image 2.1, MOSS-Image | fp8 or GGUF Q5-Q8 |
| Wan 2.2 / LTX-2.5 video | GGUF at 720p; slower than RTX 40/50 cards |
| LoRA training | SDXL and most image model LoRAs; QLoRA on 8B-14B LLMs |
As a rough ceiling, a 20 GB model file on 936 GB/s cannot generate faster than about 45 tokens/s; real speeds land below that. Use my VRAM requirements guide to check any model before you download it, and the best Ollama models by VRAM for the 24 GB tier.
Limitations to Know
- No FP8 or FP4 hardware. Ampere tensor cores do FP16/BF16 and INT8. FP8 model files still load (they are upcast), but you do not get the speed boost that RTX 40 and 50 cards get. GGUF files work normally.
- Slower compute for images and video. Diffusion models are compute-heavy, so an RTX 5070 Ti can be faster than a 3090 on image generation even with less VRAM. My gpu-unlock BF16 guide covers a useful speed trick for RTX 30 cards in ComfyUI.
- Power and heat. 350 W board power, with spikes above that. NVIDIA specifies a 750 W system power supply for one card.
- Size. Most 3090s are 3 slots thick and over 30 cm long. Check your case and motherboard slot spacing, especially for two cards.
- Age and no warranty. Every used 3090 is at least four years old. Treat it as having no warranty.
Checklist Before You Buy
- Price check. Compare the listing with the current fair range (about $1,400-$1,500 in late 2026). Very cheap listings are often scams or broken cards.
- Buy with returns. eBay with buyer protection, a local seller who lets you test, or a retailer’s refurbished program.
- Ask about history. Mining is not an automatic no (steady mining temperatures can be gentler than gaming cycles), but ask whether pads or paste were replaced.
- Look at photos. Check for bent fins, missing screws, broken fan blades, burnt or melted power connectors and signs that the cooler was opened.
- Check the model. Founders Edition uses a 12-pin connector (adapter needed); partner cards usually use two or three 8-pin connectors. Plan your power supply cables.
How to Test a Used 3090
Do this as soon as the card arrives, well inside the return window:
- Install the latest NVIDIA driver and check that
nvidia-smishows 24576 MiB of memory. - Run memtest_vulkan for 15-30 minutes. Any memory error means return the card. VRAM tends to fail before the GPU core on these cards.
- Watch memory junction temperature (GPU-Z or HWiNFO on Windows). The GDDR6X chips on the back of the board run hot; sustained readings near 100-110 C under load mean the thermal pads need replacing.
- Run a real AI load: load a 20 GB model in Ollama or llama.cpp and generate for 10 minutes, then run a long ComfyUI batch. Look for crashes, artifacts or driver resets.
- Listen to the fans across their speed range for rattles or grinding.
Two RTX 3090s for 70B Models
Two used 3090s (about $2,900) give you 48 GB, enough for Llama 3.3 70B at Q4_K_M (42.5 GB) or other 70B-class models with a modest context. llama.cpp, Ollama and LM Studio split layers across both cards automatically. In older llama.cpp benchmarks, two 3090s ran a 70B Q4_K_M model at about 16 tokens/s. You need a 1,000-1,200 W power supply, a board with two well-spaced x16-length slots, and strong case airflow (or a riser and open frame). For comparison, the same money buys a Mac Studio M5 Max with up to 128 GB of unified memory; see Mac vs PC for local AI.
Power Limit and Cooling Tips
- Power limit it. LLM inference is limited by memory, not power, so you can cap the card at around 250-280 W with
sudo nvidia-smi -pl 280(Linux) or MSI Afterburner (Windows) and lose very little speed while cutting heat and noise. - Replace thermal pads if memory temperatures are high. It is a common, cheap fix, but it voids any remaining warranty and needs the right pad thickness for your model.
- Space your cards. Two 3-slot cards next to each other will choke the top card. Leave a slot gap or use a riser.
Setting Up a Used 3090 for AI (Step by Step)
- Power: use a quality 850 W or bigger PSU for one card (1,200 W+ for two) and separate PCIe 8-pin cables for each connector, not one daisy-chained cable. The 3090 has short power spikes well above its 350 W rating.
- Clean driver install: on Windows, remove old drivers with DDU in Safe Mode, then install the latest NVIDIA Studio or Game Ready driver. On Linux, install the distribution’s recommended NVIDIA driver and check
nvidia-smi. - Install your AI tools: Ollama or LM Studio for chat models, ComfyUI (portable build on Windows) for image and video. Both detect the card automatically.
- First model: run
ollama run qwen3.8:27bor another model of about 15 to 20 GB to fill most of the 24 GB, then watchnvidia-smito confirm the GPU, not the CPU, is doing the work. - Set the power limit (see the tuning tips above) and save it as a startup task so it survives reboots.
Used 3090 Problems and Fixes
| Problem | Likely cause | Fix |
|---|---|---|
| PC shuts off under AI load | PSU cannot handle power spikes | Bigger PSU, separate PCIe cables, or a lower power limit |
| Artifacts, driver resets or “CUDA error: an illegal memory access” | Overheating or failing VRAM | Check memory junction temperature; run memtest_vulkan; repad or return the card |
| Ollama runs on CPU and is very slow | Driver not detected or model too large | Update the driver, restart Ollama, and use a smaller model or quant that fits in 24 GB |
| CUDA out of memory in ComfyUI | Model plus image size exceed 24 GB | Use FP8 or GGUF model versions, lower resolution, or enable low-VRAM options |
| Loud coil whine | Normal on some cards at high frame or token rates | Lower the power limit; it is harmless |
| Second 3090 not detected | Slot, riser or BIOS setting | Reseat the card, test the riser, and set the slot to PCIe Gen 4 or Gen 3 in the BIOS |
Frequently Asked Questions
Is the RTX 3090 still good for AI in 2026?
Yes. 24 GB of VRAM and 936 GB/s of bandwidth still make it one of the best cards for local LLMs. It is slower than RTX 40 and 50 cards for image and video generation and lacks FP8 hardware.
How much is a used RTX 3090 in 2026?
About $1,450 median on eBay US in September 2026, with a fair range of roughly $1,400-$1,500 and a 90-day low of about $1,244.
RTX 3090 or RTX 5060 Ti 16 GB for AI?
The 3090 has 50% more VRAM and about twice the bandwidth, so it runs bigger models faster. The 5060 Ti is new, uses much less power and has FP8/FP4 support, and costs about half as much.
Is it safe to buy an ex-mining RTX 3090?
It can be. Ask about the card’s history, buy with a return option, and run memtest_vulkan and a long AI workload before the return window closes.
Can two RTX 3090s run a 70B model?
Yes. Together they have 48 GB, which fits a 70B model at Q4_K_M (about 42.5 GB) with a modest context. llama.cpp and Ollama split the model across both cards automatically.
What power supply do I need for an RTX 3090?
A good 850 W unit for a single card with a normal CPU, and 1,200 W or more for two cards. Use separate PCIe power cables for each 8-pin socket.
Final Thoughts
The used RTX 3090 is no longer a bargain, but it is still a smart buy if you want fast 24 GB CUDA performance without paying RTX 5090 prices. Test it properly, power-limit it, and it will run most of today’s open models comfortably. If you only need it for occasional big jobs, renting one at about $0.50 an hour may be smarter; compare in the cheapest cloud GPUs for AI.
