The question people frame as “self-host vs cloud” for AI coding is really a utilization bet, and most home rigs quietly lose it. Stephen Bochinski’s breakdown lays out three paths — buy the box and run open models locally, rent those same open weights from a provider at API rates, or min-max frontier subscriptions where roughly $400/month buys ~$2800 of list-price usage until you hit the metered ceiling.

The local option only pays off if you keep the machine saturated. That’s the variable everyone underestimates: capability isn’t the constraint, sustained utilization is. A 24GB card grinding a quantized model overnight is cheap per token only when there’s an overnight queue to grind. There usually isn’t, and the GPU you optimize around today looks like a bad bet a year later when the model and hardware churn underneath you. Renting open weights — a near one-line provider swap — is the boring right answer precisely because it keeps you liquid while the configurations are still in flux.

This maps cleanly onto enterprise inference capacity planning. You don’t provision GPUs for peak capability; you provision for sustained load, and you measure cost-per-successful-task, not cost-per-token. A coding agent that retries a “cheap” local model three times can cost more — in wall-clock and watts — than one clean frontier call that lands on the first attempt. The retry tax never shows up in the per-token sticker price.

The HN thread is full of home-rig war stories, and the honest ones all circle the same admission: the box sits idle more than expected. So here’s the real test before you buy: can you name the long-running workload that will keep it busy at 2am? If you can’t, you’re renting whether you own the hardware or not.