Buying a used RTX 3090 for AI: what to check and what to pay

The RTX 3090 came out in 2020 and it's still the card most people buy for local AI. The reason is simple: 24 GB of fast memory for the price of a mid-range new card. Memory decides which models you can run (see how much VRAM you need), and nothing else this cheap has 24 GB at 936 GB/s.

What it runs

RTX 3090, 24 GB · estimated with our calculator · full page

Qwen3 4BQuick everyday chat and summaries
~210 tok/s
Runs greatqwen3:4b
Qwen3 8BA solid all-rounder for chat and writing
~99 tok/s
Runs greatqwen3:8b
Gemma 3 12BReads images as well as text
~64 tok/s
Runs greatgemma3:12b
Qwen3 14BNoticeably smarter at reasoning and code
~55 tok/s
Runs greatqwen3:14b
gpt-oss 20BOpenAI's open model. Fast, strong reasoning
~160 tok/s
Runs greatgpt-oss:20b
Gemma 3 27BGoogle's best open model for its size
~30 tok/s
Runs greatgemma3:27b
Qwen3 30B-A3B30B smarts at small-model speed
~160 tok/s
Runs greatqwen3:30b
Qwen3 32BStrong coding and long answers
~26 tok/s
Usableqwen3:32b
Llama 3.3 70BClose to the big cloud models
needs 45 GB
Too bigllama3.3:70b
gpt-oss 120BOpenAI's largest open model
needs 67 GB
Too biggpt-oss:120b
Qwen3 235B-A22BAlibaba's flagship open model, for multi-GPU rigs
needs 144 GB
Too bigqwen3:235b
DeepSeek R1 671BThe full DeepSeek R1, for the biggest rigs
needs 406 GB
Too bigdeepseek-r1:671b

The highlights: Qwen3 32B at about 26 tok/s and Qwen3 30B-A3B at around 160 tok/s. Two of them run Llama 3.3 70B.

What to pay

As of October 2026, a used 3090 goes for roughly $650 to $850. The low end is usually ex-mining or a blower model; the high end is a clean card from a gaming PC, sometimes with a box.

The 3090 Ti is a little faster and runs cooler memory, but it usually costs more for little gain on this job. A 4090 costs far more and generates text only about 8% faster; it's much faster at prompt processing and image generation.

What to check before you buy

Run a real load. Ask the seller to show it under a stress test, or test it yourself: 15 minutes of a GPU benchmark, or a model generating non-stop. Watch for crashes, artifacts, or the fans going wild.
Memory temperature. The 3090 has memory chips on the back of the board, and they run hot. Tools like HWiNFO (Windows) show "memory junction temperature". Under load it should stay below about 100°C. Above 105°C, plan to replace the thermal pads.
Fans. Spin each one up. Grinding or clicking means a fan replacement soon (cheap, but annoying).
Ex-mining isn't automatically bad. Mining ran cards at steady, lowered power, which is gentle on the chip. The risks are dried-out thermal pads and worn fans, both fixable. Price it accordingly.
The power connectors. Look for scorching or melted plastic around the 8-pin plugs.

What to pair it with

Power supply: 850 W of good quality. The 3090 can spike well above its 350 W rating for milliseconds, and cheap supplies shut down. For two cards, 1,200 W or more.
Case clearance. Most 3090s are over 30 cm long and take three slots. Measure.
System RAM: 32 GB or more, so loading models is quick and you have headroom if one spills over.
Any CPU from the last several years. The GPU does the work.

Our full parts list with prices is the "Popular" build on the builds page.

Want to save power?

You can cap the card's power with almost no loss in text generation speed, since it's limited by memory, not compute. On Linux:

sudo nvidia-smi -pl 280

That caps it at 280 W. On Windows, MSI Afterburner's power limit slider does the same.

Alternatives

Less money: a Tesla P100 gives you 16 GB of very fast memory for under $100, with some homework.
Quieter: a Mac with lots of memory is silent and fits bigger models, but runs them slower.
More memory: two 3090s give 48 GB for 70B models.