Used RTX 3090 hits $1000 for local LLMs

For local AI, VRAM capacity is the only metric that truly matters, making a used 3090 a better investment than a new mid-range card. Its 24GB VRAM allows for 30B+ parameter models at 4-bit quantization that 16GB cards cannot accommodate, while its 384-bit memory bus offers nearly triple the bandwidth for faster token generation. Although hardware age and high power requirements carry some risk, the 3090 remains future-proof for independent developers as the VRAM floor for useful local models continues to rise.