Sunk Cost, a new web application, provides estimates for the time required for a local large language model (LLM) hardware setup to pay for itself.
Sunk Cost Tool Estimates Payback Period for Local LLM Hardware
The tool calculates the payback period by factoring in the initial purchase price of the hardware, electricity consumption under load, and the estimated savings compared to using cloud-based APIs. Users can input specific hardware configurations, such as a Mac Studio with a specific chip and memory capacity, and define usage patterns, including tokens per day and context window size. For example, with a Mac Studio (64GB) running Qwen3.8 27B for agentic coding tasks, the tool estimates a payback period of approximately 43.6 years based on current API costs and electricity usage.
To provide a comparison of speed, Sunk Cost estimates local generation rates using a formula involving memory bandwidth and bytes read per token, though actual measured speeds are used where available. The tool also includes capability ratings that compare various models against frontier models like Claude and GPT based on public benchmarks. Regarding privacy, the service states that it does not store cookies or IP addresses, collecting only the request's country to improve estimates.
Sources
- Show HN: Sunk Cost – How long until a local LLM rig pays for itself? （Hacker News Frontpage, 2026-09-15）
- ローカルでAIを動かして元を取るまで何年かかるかがわかる「Sunk Cost」、例えばメモリ64GBのMac Studioではどれだけの時間が必要なのか？ （GIGAZINE, 2026-09-15）