Independent researcher's $48K home GPU rig beat cloud rental by $17K

An ex-FAANG researcher built ‘grumbl,’ a six-GPU RTX 6000 Ada server, for $48,000 after quitting to do independent work. The 6000 Ada won out over A100 and H100 on price-to-throughput for the inference-heavy reinforcement learning workloads planned, and apartment power limits forced an unusual two-PSU build wired into separate circuits — a setup risky enough to justify hiring a professional builder. Ironically, the rig later moved to a parents’ basement where proper circuits could be installed.

Usage logging showed 76% lifetime utilization (85% since January 2025), below the author’s 95%+ expectation but enough to break even. Tallying daily on-demand cloud equivalents against actual GPU-hours used, plus roughly $3,000 in electricity, the home rig has saved about $17,000 as of March 2026 and now nets $90–$105/day in avoided rental costs. The compromise was a motherboard with slow GPU interconnect — fine for parallel small experiments, bad for model-parallel training.

The author’s stronger argument is psychological: owning flips the cost calculus so idle GPUs feel wasteful rather than each experiment feeling expensive, which encouraged more risk-taking. They’d still do it again, but next time would buy a standard datacenter box and rent colo space rather than hand-building. Renter’s insurance refused to cover the hardware; business insurance was required.