The RTX 4090 is four years old and out of production. Yet in November 2026 it’s one of the most asked-about cards for local AI. With a new RTX 5090 selling for around $6,900, a used 4090 suddenly looks like a bargain.
But used 4090 prices have climbed too, and there are real risks when buying one second-hand. Here’s what it runs, what it really costs, how it compares, and exactly what to check before you pay.
- Specs: 24 GB GDDR6X, 1,008 GB/s, 16,384 CUDA cores, FP8 support, 450 W.
- Used price (US): eBay sold average about $2,340 (31 Oct 2026). Median asking price about $3,300 (6 Nov). Pay closer to the sold price.
- Runs: 27B-31B LLMs fully on the GPU (Qwen3.8 27B, Gemma 4 31B), most image models at full quality, video models with FP8 or GGUF.
- vs 5090: about 56% of the bandwidth and 75% of the VRAM, for roughly a third of the price.
- Verdict: worth it at $2,300-$2,600 if you need 24 GB and fast image/video generation. At $3,300+ it’s poor value.
RTX 4090 Specs for AI
| Spec | RTX 4090 | Why it matters |
|---|---|---|
| Architecture | Ada Lovelace (AD102), 2022 | Supports FP8 Tensor math; no FP4 |
| CUDA cores | 16,384 | Fast image/video generation |
| VRAM | 24 GB GDDR6X | Fits 27B-31B LLMs at Q4 |
| Memory bus / bandwidth | 384-bit / 1,008 GB/s | Fast token generation |
| Power | 450 W, 850 W PSU recommended | Needs a strong PSU and good airflow |
| Power connector | 16-pin 12VHPWR | Must be fully seated (see checklist) |
| Launch price | $1,599 (October 2022) | Now discontinued, so used only |
Note: FP8 support matters in 2026. Many image and video models ship FP8 files that run natively on the 4090, while the older RTX 3090 has to convert them.
Used RTX 4090 Prices in November 2026
Two trackers show very different numbers, and the difference is important.
| Source | What it measures | Price | Trend |
|---|---|---|---|
| GetPCParts (31 Oct) | eBay sold listings, weekly average | $2,342 | Down 10% in 30 days |
| DealHunter (6 Nov) | Asking prices, 104 listings | $3,300 median | Up from $2,499 on 20 Aug |
Sold prices are what people actually paid. Asking prices are what sellers hope for. So a fair used 4090 in the US right now is roughly $2,300-$2,600, and sellers asking $3,300 or more are betting on desperate buyers.
Tip: On eBay, filter by Sold Items to see real prices for the exact model before you make an offer.
For context on new cards, see why GPU prices are so high in 2026.
Which LLMs Fit in 24 GB
About 22-23 GB is usable after the driver and desktop. Sizes are Ollama download sizes; add 1-2 GB for context.
- Qwen3.8 27B (18 GB): fits fully with a 32K+ context. My first pick on 24 GB.
- Gemma 4 31B (19-20 GB): best for maths and reasoning.
- Muse Glimmer 30B (18 GB) for agents and Laguna XS 2.1 (20 GB).
- Flint 40B (23 GB, about 2.5B active): just fits with a short context, or with light offload.
- 70B dense models need about 42 GB at Q4, so you’d need two cards or heavy offload.
- 120B-class MoE (gpt-oss-120b, Nemotron 3.5 Super): run with experts in system RAM if you have 64-128 GB.
See my best Ollama models by VRAM for pull commands.
Expected LLM Speeds
Token generation is roughly limited to bandwidth divided by model size. Real speeds are usually 60-80% of that.
| Model (file size) | RTX 4090 ceiling | RTX 3090 ceiling | RTX 5090 ceiling |
|---|---|---|---|
| Gemma 4 12B (8 GB) | ~126 tok/s | ~117 tok/s | ~224 tok/s |
| Qwen3.8 27B (18 GB) | ~56 tok/s | ~52 tok/s | ~100 tok/s |
| Gemma 4 31B (20 GB) | ~50 tok/s | ~47 tok/s | ~90 tok/s |
Note: These are ceilings I calculated from bandwidth, not my own benchmarks. For chatting, the 4090 is only slightly faster than a 3090. Its big advantage is compute: prompt processing and image/video generation.
Image and Video Generation
This is where the 4090 earns its price. It has far more Tensor and CUDA performance than a 3090, plus native FP8.
- Image models: Qwen Image 2.1, Flux-class and SDXL run at full quality with room for LoRAs and ControlNets.
- Video models: Wan 2.2 and LTX 2.5 run in FP8 or GGUF with sensible clip lengths. My Wan 2.2 guide shows a fast setup.
- LoRA training: 24 GB handles image LoRAs comfortably and small video LoRAs with care.
- Speed-ups: SageAttention works well on Ada. See installing SageAttention on Windows.
4090 vs 5090 vs 3090 vs 5080
| Card | VRAM | Bandwidth | FP8 / FP4 | Power | Price (Nov 2026) |
|---|---|---|---|---|---|
| RTX 5090 | 32 GB | 1,792 GB/s | Yes / Yes | 575 W | ~$6,900 new |
| RTX 4090 (used) | 24 GB | 1,008 GB/s | Yes / No | 450 W | ~$2,300-$2,600 sold |
| RTX 3090 (used) | 24 GB | 936 GB/s | No / No | 350 W | ~$1,450 |
| RTX 5080 | 16 GB | 960 GB/s | Yes / Yes | 360 W | ~$2,000 new |
More detail: used RTX 3090 for AI, RTX 5090 vs 5080 for AI and the budget RTX 5060 Ti 16GB.
Used RTX 4090 Buying Checklist
- Check the 16-pin connector. Look for melted, browned or bent pins on the card and cable. 12VHPWR melting was a known 4090 problem. Walk away if you see damage.
- Ask for a stress test. Ask the seller for a video of a 15-minute stress test (FurMark, OCCT or a ComfyUI run) with
nvidia-smishowing 24 GB, temperature and power. - Verify the model.
nvidia-smiand GPU-Z should report “NVIDIA GeForce RTX 4090” with 24,564 MiB. Check the serial number on the box matches the card. - Weigh it and look inside. There have been scams where the GPU chip was removed and replaced. A card that is suspiciously light or won’t post is a red flag.
- Check hotspot and memory temps. Under load, memory junction temps far above 100 °C mean worn thermal pads.
- Ask about history. Cards from mining or AI rental farms ran 24/7. That isn’t always bad, but expect to replace pads and paste.
- Prefer buyer protection. Use eBay or PayPal Goods and Services, and test on arrival within the return window.
Warning: Never pay by bank transfer, gift cards or crypto for a used GPU. A “too good to be true” 4090 under $1,500 is almost always a scam.
4090D and 48 GB Modded Cards
- RTX 4090D: a China-market version with fewer CUDA cores (14,592) and slightly lower performance. Fine for AI, but should cost less than a full 4090.
- 48 GB modded 4090s: Chinese workshops sell 4090 and 4090D cards with the memory upgraded to 48 GB. They’re popular for LLMs but have no NVIDIA warranty, can need special drivers or firmware, and are hard to resell. Buy only from a seller with a track record.
Setup and Power Tips
- PSU: a quality 850 W or bigger ATX 3.0/3.1 unit with a native 12V-2×6 or 12VHPWR cable.
- Cable: push the connector in until it clicks, and avoid sharp bends near the plug.
- Power limit: for AI, capping at 70-80% (
nvidia-smi -pl 350) usually costs only a few percent of speed and runs much cooler. - Drivers: use a current Studio or Game Ready driver. Ollama, llama.cpp, LM Studio and ComfyUI all support Ada out of the box.
- Monitoring: see my GPU and VRAM monitoring guide.
Should You Buy One?
| If you… | My pick |
|---|---|
| Mostly chat with 27B-32B LLMs | Used RTX 3090 (cheaper, nearly as fast) |
| Do a lot of image and video generation | Used RTX 4090 at $2,300-$2,600 |
| Need 32 GB for LLMs on a budget | Radeon AI PRO R9700 (~$1,700-$1,850) |
| Want the fastest single card regardless of price | RTX 5090 |
| Need a big GPU only occasionally | Rent one (see my cloud GPU guide) |
Troubleshooting
The PC shuts down under AI load
Usually the PSU. The 4090 has short power spikes above 450 W. Try a power limit with nvidia-smi -pl 350, check the cable, or use a bigger PSU.
Out of memory with a 27B or 31B model
Lower the context or use KV cache quantisation. My CUDA out of memory fix covers Ollama, llama.cpp and ComfyUI.
Memory temperature is very high
Improve case airflow and lower the power limit. On an old card, new thermal pads often drop memory temps a lot, but that voids any remaining warranty.
ComfyUI is slower than expected
Check that you’re using FP8 or full-precision files rather than heavy GGUF quants you don’t need, and install SageAttention. Make sure the model isn’t spilling into system RAM.
Frequently Asked Questions
Is the RTX 4090 still good for AI in 2026?
Yes. 24 GB and 1,008 GB/s still run 27B-31B LLMs fully on the GPU, and its compute and FP8 support make it fast for image and video models.
How much is a used RTX 4090 in November 2026?
eBay sold listings averaged about $2,340 at the end of October, while the median asking price was about $3,300 in early November. Aim for the sold price.
RTX 4090 or RTX 5090 for AI?
The 5090 has 32 GB and about 1.8x the bandwidth, but costs around $6,900 new. A used 4090 at $2,300-$2,600 is much better value unless you need 32 GB on one card.
RTX 4090 or used RTX 3090 for local LLMs?
For chat speed they’re close, because bandwidth is similar. The 3090 is about $900 cheaper. Pick the 4090 if you also generate images and video.
Can an RTX 4090 run a 70B model?
Not fully on the GPU. A 70B model needs about 42 GB at Q4, so you need two cards, a 48 GB modded card, or heavy CPU offload.
What should I check when buying a used RTX 4090?
Inspect the 16-pin connector for melting, ask for a stress-test video with nvidia-smi, verify the model and serial number, and buy with buyer protection.
Is a 48 GB modded RTX 4090 worth it?
It can be for large LLMs, but it has no NVIDIA warranty, may need special drivers and is hard to resell. Buy only from a seller with a proven record.
Final Thoughts
The RTX 4090 is still a superb AI card, and with new prices this high, it’s often the most sensible high-end choice. The key is paying a sold-listing price, not an asking price, and checking the card carefully.
If you mainly run LLMs, a used 3090 saves real money. If you also make images and video, the 4090 is the one I’d buy. Compare everything in my best GPU for local LLMs guide.