- Learn
- Pick your hardware
The best GPU for local AI under $300
The best graphics cards for running local LLMs on a $300 budget, what each one runs, estimated speeds, and which to pick for easy setup or raw value.
For most people, a used RTX 3060 12 GB. It costs about $180 to $230 used as of October 2026, it fits 14B models, and every tool supports it. For the most memory and speed per dollar, a Tesla P100 at under $100 each, if you're happy to add a cooling shroud and use Linux-friendly tools. Two P100s fit in the same budget and give you 32 GB.
The shortlist
| Card | Memory | Bandwidth | Qwen3 14B, estimated | Used price |
|---|---|---|---|---|
| RTX 3060 12 GB | 12 GB | 360 GB/s | ~21 tok/s | $180 to $230 |
| Tesla P100 | 16 GB | 732 GB/s | ~43 tok/s | We paid about $85 |
| Tesla P40 | 24 GB | 346 GB/s | ~20 tok/s | Check current prices |
| RTX 2080 Ti | 11 GB | 616 GB/s | ~36 tok/s | Check current prices |
| Arc B580 | 12 GB | 456 GB/s | ~27 tok/s | Check current prices |
| RX 7600 XT | 16 GB | 288 GB/s | ~17 tok/s | Check current prices |
Prices for the first two come from our builds page. Speeds are our calculator's estimates. Prices for the rest move a lot, so check before you buy and only consider ones that land under $300.
Pick 1: RTX 3060 12 GB, the easy one
It's slow on paper but has enough memory for Qwen3 14B, the point where local models start to feel smart. It works on Windows and Linux with Ollama, LM Studio, image generators and everything else. Avoid the 8 GB version.
Our "Starter" build puts it in a used desktop for $300 to $400 total. Full details: the RTX 3060 12 GB for AI.
Pick 2: Tesla P100, the value one
The P100 has 16 GB of HBM2 at 732 GB/s, about twice the 3060's bandwidth, so it generates text about twice as fast on anything that fits. It also fits gpt-oss 20B, at about 128 tok/s estimated.
Two of them, at $170 to $240 for the pair, give you 32 GB. On our rack, two P100s run Qwen3.6-35B-A3B at 53 tok/s, measured.
The catches: no fans (you need a shroud, about $15 to $25 per card), a CPU-style EPS power plug, no display output, slow prompt processing, and older software support. Read the Tesla P100 guide first.
The others
- Tesla P40: 24 GB in one card, but slow memory. Good for MoE models like Qwen3 30B-A3B, at about 61 tok/s estimated; slow on dense 32B models. Same cooling and power catches as the P100. See the Tesla P40 guide.
- RTX 2080 Ti: fast for its age, but 11 GB is tight for 14B models with long context.
- Arc B580: a new card with 12 GB, but narrower software support. See Intel Arc for local AI.
- RX 7600 XT: 16 GB is a lot for the money, but its memory is slow. See AMD GPUs for local AI.
What to avoid
- 8 GB cards if you can stretch to 12 GB. 8 GB stops at 8B models. See how much VRAM you need.
- Very old cards with 4 to 6 GB, like the GTX 1060. They'll run only the smallest models.
- Untested listings. Ask for proof it runs under load. See how to check a used GPU.
Which one?
- First card, want it to just work: RTX 3060 12 GB.
- Homelab, want the most for your money: two Tesla P100s.
- Can stretch your budget: see the best GPU under $1,000, where a used RTX 3090 changes the picture.
One email when there's a new guide, podcast episode, or result from the rack. No spam.