The value pick
A used RTX 3090: 24GB of VRAM for $1,499. The most memory per dollar in local AI.
The path the used-market already knows: one RTX 3090, 24GB of VRAM at 936GB/s for $1,499. It runs the 27B class at 4-bit in VRAM alone, carries the biggest agentic workloads with the newest split-serve recipes, and does image generation the Macs cannot touch. Add 96GB of RAM and an NVMe drive and the same build becomes a $3,000 machine that runs Qwen3.8-Flash-Next - Artificial Analysis scores it level with GPT-6 Sol (medium). It costs about $770 less than the Mac above it, and it is loud.
You are comfortable buying used hardware: find a card with a return policy, stress-test the memory the day it arrives, and you save real money for one afternoon of setup. If you already own a desktop, the 3090 drops in beside your parts and the build lands under budget.
last check: Sep 29, 2026
The parts
What it runs
The 300B+ MoE class: 24GB of VRAM holds 27B models at 4-bit, not 300B models at any quant. Used cards carry no warranty, which is part of the price.
About 350W under AI load, and the 3090 is loud: the main trade against the Mac. Plan an 850W power supply and a case with real airflow. No dedicated circuit needed, but the room knows the card is working.
This is the tier where local AI becomes boring, which is the highest compliment. The box is silent, the answers are fast, and the models are big enough that you stop asking what a model can do and start asking what you want done. Your writing assistant, your code reviewer, your document librarian, your form-filler: all local, all day, on hardware that uses less power than the monitor next to it.
See the $5k tier →