A calculator for when a local LLM rig pays for itself
Sunk Cost takes a machine, a model and a daily token count, and works out how long the hardware takes to beat renting the same model by the token.
The advice circulating among founders is simple enough: buy a Mac, run models locally, and the hardware pays for itself. Sunk Cost, posted to Hacker News as a Show HN, is an attempt to check that arithmetic.
The builder says he kept hearing the claim and wanted to test it. The tool asks for three things: the machine, the model and how many tokens you get through in a day.
It then compares the cost of owning the box against paying an API provider for the same model, token by token, and returns a payback period.
The inputs are where the judgement sits. You set the memory in the machine, the wattage it draws under load and its memory bandwidth, then the context window you want and the ratio of input tokens to output tokens.
Three assumptions are left open to the reader. Electricity price in dollars per kilowatt-hour, the speed of the API you are comparing against in tokens per second, and either your own measured local speed or your current monthly bill.
There is a fourth, and it is the one most back-of-envelope calculations skip. The tool lets you assume API prices keep falling, and lets you set how fast.
That matters because a rig is paid for once and a token bill is paid every month. If the rental price halves during the payback window, the payback window moves.
Where local speed has not been measured, the tool estimates it from memory bandwidth divided by bytes read per token, and labels the figure as an estimate rather than a measurement.
The site also ranks models against Claude and GPT, and lists what it calls the quickest payback at each level of capability. Those capability ratings are described on the site as a judgement call informed by public benchmarks, and deliberately coarse.
No break-even figures are quoted on the page itself; the numbers depend entirely on what you put in. That is the honest answer, and it is also the reason to run it before buying rather than after.
The privacy terms are stated plainly. The machine, speed and bill you enter are sent to the site once you stop typing, with no cookies and no IP address, only the country the request came from. Browsers set not to be tracked send nothing.
For a founder, the practical use is narrow and worth an hour. Put in your real daily token volume rather than your hoped-for one, put in your actual electricity tariff, and assume API prices fall. If the rig still pays back inside a period you can plan for, the purchase argues for itself. If it only works at flat prices and heavy usage, it is a preference, not a saving.
More: Sunk Cost