Would an $1,800 rig ever pay for itself against just paying per token?

Hybrid, and I have not gone back. Local model does the bulk, private and repetitive work where being merely good is fine, and anything hard or customer-facing goes to an API. My API bill dropped by roughly two thirds rather than to zero, and the local box earns its keep on the documents I genuinely cannot send anywhere. Trying to make one of the two do everything is what makes people unhappy with both.

Do the sum out loud, it is not close on cost alone. $1,800 over three years is $50 a month before you plug it in. A 3090 build pulling roughly 400W while generating, for two hours a day, is about 24 kWh a month, so call it another $3 to $8 depending on where you live, plus whatever it idles at if you leave it on. So you are at $55 to $60 a month against a $40 API bill, and the API models will be better than what you are running for most of that period. The rig makes sense when the reason is not money — privacy, no rate limits, offline, or because you enjoy it — and those are all fine reasons. It just is not a savings decision at two hours a day.

I bought the rig, used it constantly for six weeks, and then used it about twice a month for a year. Be honest with yourself about whether the appeal is the capability or the project of assembling it, because for me it was substantially the second one. The saving grace is that a 3090 build is also just a good computer, so it was not wasted, but I would not call it a $1,800 inference decision in hindsight.

Rent a 3090 or 4090 by the hour from a compute provider for a weekend first. Twenty dollars tells you whether the models you can actually run at that size are good enough for your work. Most of the regret posts in here would have been prevented by that weekend.

The comparison people make is usually unfair in the other direction too, because the workload changes once inference is free at the margin. My API usage was disciplined — I thought about each call. Locally I run things I would never have paid for: re-summarising the same document five ways, batch tagging a few thousand files overnight, throwing an entire repository at a long context because why not. My effective usage went up something like tenfold. Whether that is value or waste depends on the person, but do not assume your two hours a day stays two hours a day.

Politely disagreeing with the cost-focused answers. The reason I run local is that my document set is not mine to upload, and once that is true the price comparison stops existing. If your summarising work involves anything under a confidentiality obligation, work out what the alternative actually is before comparing dollars, because for some of us there is no cloud column in the table.