Building a Secret GPU Cluster

This post is a draft. Last updated: September 2026.

• built an inference cluster across 30 idle university lab computers
• detected interactive logins and released the machine until the user left
• tunneled SSH traffic through the university VPN to reach lab machines isolated from the public internet
• bypassed per-user disk quotas by storing model weights in RAM instead
• used the cluster to run large-scale LLM evaluations without relying on external inference providers