Local LLM workstation, $12k budget.

Independent ML researcher / consultant.

Needed to run 70B-class quantized models locally without offloading to cloud for every test.

• $12k all-in budget
• Standard 15A residential circuit
• Home office — noise mattered
• Linux preferred

Two-GPU configuration sized to VRAM ceiling, high-airflow case, A/B compared against a single-GPU alternative.

Stable inference performance on target model sizes, quiet enough to share a room, and a documented upgrade path to a third GPU.

VRAM was the real constraint. CPU and RAM choices were comfortably second-order.