A few months back my team lead asked why I run a local model at all when "the cloud stuff is basically free." It's a fair question until you actually add it up, and I'd never done the math. So I did.
Start with the subscription. The chat model we were paying for came to about $20 a month per seat. Four people actively used it. That's $80 a month, roughly a grand a year, and it only grows with your team. It's a rent check you keep paying for as long as you use the thing.
Now the home lab. I already had an old workstation with a used GPU I'd picked up for a few hundred dollars. LM Studio and Open WebUI are free. A few quantized models that fit in its VRAM are free. Docker keeps the whole thing in one folder I can back up or nuke without ceremony. The only real ongoing cost was power, and honestly a mid-range GPU idling at a spooled-down quant draws less than people assume. I measured it at the wall: about the same as leaving a small space heater on, and I don't run it 24/7.
Here's the embarrassing part. The break-even landed somewhere around the seven-month mark. Seven months of rent vs one lump sum I've already spent, and after that the gear is mine. I can run as many experiments as I want without watching a metered bill tick up. My drafts, my notes, the occasional bit of config I'd rather not paste into someone's cloud feed — all of it stays on my own disk.
I'm not going to pretend it's a drop-in replacement. The small models I run are a cut below the big frontier ones on hard reasoning, and my context window means I can't throw a whole repo at it. I've learned to live with that trade. For the steady day-to-day stuff — drafting, summarizing, explaining a flag I haven't touched in two years — a good quant is more than enough, and the privacy is a genuine side win, not a slogan.
The thing that surprised me most wasn't the savings. It was how quickly the local setup stopped feeling like a project and started feeling like just another tool in the stack. You get past the setup once, and then it's just there, doing the work, costing nothing per click.
If you want a sensible place to start without reinventing it, I put together a self-hosted AI home lab bundle covering the LM Studio and Open WebUI wiring, quantized models, and the Docker layout that actually works: https://symshah.gumroad.com/l/selfhosted-ai-homelab. No cloud account required.
Top comments (0)