DCGM

YC-backed Expanse claims to cut wasted GPU compute in HPC/Kubernetes/SLURM clusters by predicting exact resource needs from live telemetry, offering paid pilots with measurable efficiency data and LLM-benchmarking context.

AI-driven k10s experiment demonstrates why architecture matters; author shifts to hand-designed architecture and Rust, with concrete guardrails to prevent future AI-induced bloat.