Sunk Cost tool tests the economics of running local AI models
The Hacker News project compares several local systems with paid APIs, but does not guarantee that buying hardware will pay off.
A new tool called Sunk Cost estimates how long local large language model hardware may take to pay for itself compared with using paid application programming interfaces.
Presented as a “Show HN” project on Hacker News, the tool is designed to help users assess the economics of running open models locally. It compares Mac mini, Mac Studio, DGX Spark and Strix Halo systems.
Sunk Cost indicates which open models fit each system, how they perform and how much usage would be needed to reach an estimated break-even point against paid APIs.
The supplied material does not include hardware prices, benchmark results, API rates, usage assumptions or specific break-even periods. It also does not establish whether costs such as electricity, maintenance or hardware depreciation are included.
The tool’s comparisons are estimates rather than independently verified financial conclusions. Its stated approach allows for the possibility that local AI hardware may not pay for itself.
