Buying a machine to run a model in‑house, renting a GPU, or paying for an API: at what volume does each option become the cheapest? This tool compares all three based on your actual usage.
How to use it
- Enter your monthly input and output tokens, in millions, and select the API model for comparison.
- Describe your machine: price, depreciation period, power, electricity cost, generation throughput, and any standby consumption.
- Provide the hourly rate for GPU rental.
- Read the three monthly costs, the machine’s occupancy rate, and the verdict statement.
Principles behind the calculation
- The API cost is the monthly volume multiplied by the chosen model’s prices, converted to euros.
- Compute time comes from throughput: output ÷ generation throughput, plus input ÷ (throughput × prompt reading multiple).
- The on‑premises cost adds depreciation (price ÷ months) and electricity consumed during that compute time, plus any standby; rental multiplies compute hours by the hourly rate.
- The break‑even point is the hardware price divided by the monthly savings compared to the API, after deducting electricity. Occupancy compares compute hours to the 720 hours in a month.
Limitations to be aware of
A local model is rarely equivalent to a cutting‑edge API model: compare it to a model of the same class. The calculation ignores your time, maintenance, quality of service, load spikes, caching, and long context.
