Local-First 52-Agent Recursive Swarm Infrastructure Bottleneck and Hardware Recommendations

Hey Community, Looking for some input on a unique infrastructure bottleneck we've hit. I’m currently deploying a 'local-first, 52-agent recursive swarm' architecture for a client, moving entirely off the legacy cloud stack. Early tests are showing significant results—a 450% efficiency recovery and near-zero latency compared to traditional cloud setups. The Bottleneck: The swarm is incredibly compute-heavy. My current local hardware is hitting severe thermal throttling just trying to keep up with the recursive logic. It's clear the workload demands dedicated, top-tier hardware—specifically looking at an M5 Max (128GB RAM) spec node to sustain it. My Question: Has anyone else here managed 'sovereign' (off-cloud) AI workloads of this density? Is the immediate jump to the M5 Max the right move, or are there other alternative hardware configurations that can handle this level of recursive compute without melting down? Appreciate any insights.