The RTX 3090 is still the best local AI GPU

Mostly, yes. Six years after launch, a used RTX 3090 is still the card I tell most people to start with.

Used cards sold for about $1,050 to $1,450 in October 2026. For that you get 24GB of GDDR6X at 936 GB/s, full CUDA, and a card the local-AI community has been building around for years. By our estimate, one 3090 runs Qwen3.8 27B at Q5_K_M at about 34 tok/s, and Qwen3.6 35B-A3B at about 159 tok/s. Nothing new at that price has 24GB.

The longer comparison, including the 5070 Ti, 4090, 5090, 7900 XTX, and the cards below, is in Is the RTX 3090 still the best GPU for local AI in 2026?. This is the short version, and the two places I would actually look instead.

Why it still wins

Token generation is mostly a memory-bandwidth problem. The 3090’s 936 GB/s is about 93% of a used RTX 4090, and the 4090 was about $2,500 in October 2026 for the same 24GB. You pay roughly 2.4 times as much for a small speed bump and faster prompt processing.

New cards at a similar price are a different trade. An RTX 5070 Ti is about $1,200, has a warranty, and supports FP4, but it is 16GB. That pushes Qwen3.8 27B down to a smaller quant. I would still take the extra 8GB for chat.

The 3090 also stacks. Two cards are about 47GB usable. Four are about 93GB, enough to hold gpt-oss-120b in VRAM, for roughly the price of one shortage-priced RTX 5090. The four-3090 IKEA build is the template I send people. You do not need NVLink for llama.cpp or Ollama. Layer split over PCIe is enough for one user.

Where I would not start with one

It is a 350W card with hot rear memory, no FP8 or FP4 hardware, and a used market that has been mined and gamed on for years. Power-limit each card to about 280 to 300W if you run more than one. Stick to Founders Edition, EVGA, and MSI if you care about fan control on Linux.

And 24GB on one card is the ceiling. If that is the problem you are solving, the next two posts are the interesting ones: a new Radeon AI PRO R9700 when you want 32GB you can put four of in a normal case, and an unlocked CMP 170HX when you want a used NVIDIA card to report 64GB. Neither replaces the 3090 as the default. Both beat it at the specific job they were bought for.