Run DeepSeek's 671B Model on 8 Mac Minis for $20K

EXO Labs demonstrated that DeepSeek-V3's full 671B-parameter model runs on a cluster of 8 M4 Pro Mac Minis (64GB RAM each), totalling 512GB of unified memory across the cluster. Cost: roughly $20,000 one-time.

The secret is DeepSeek's Mixture-of-Experts (MoE) architecture — only ~37B parameters activate per inference pass, so the model distributes cleanly across Apple Silicon. The whole cluster draws around 1,120W at peak.

Comparable cloud capacity on AWS runs $20,000–$40,000 per month. The hardware pays itself off before your first invoice clears. Data stays local, latency is yours to control, and there's no vendor throttling during peak hours.

Why it matters: the gap between 'rent AI forever' and 'own your inference stack' just became a one-time hardware purchase decision.