AI hardware · break-even
When does local hardware beat paying per token?
Hardware cost, resale value, electricity and your monthly token volume against an API price you type in; months to break even.
Inputs
Starting values are illustrative, not current prices or specifications. Replace each with your own.
Result
Answer
Break-even is the hardware cost you have to recover, minus resale, divided by the monthly saving. The monthly saving is the API bill you avoid minus the local running cost, which here is electricity plus anything else you add. If local running cost is not below the API price, there is no break-even. The API price is whatever you type in; this page does not look up prices.
Worked example
Hardware 2,000 with 800 resale leaves 1,200 to recover. At 350 W and 40 tokens per second, electricity is 0.3646 per million tokens at 0.15 per kWh. At 20 million tokens a month and an example API price of 5 per million, the API costs 100.00 a month and local electricity 7.29, a saving of 92.71. Break-even is 1,200 ÷ 92.71 = 12.9 months. At 40 tokens per second the machine can generate at most 103.7 million tokens in 30 days.
Formulas and notes
- Local energy per million tokens = watts ÷ 1,000 × (1,000,000 ÷ tokens per second ÷ 3,600) × price per kWh.
- Monthly saving = API price × volume − (energy per million × volume + other monthly cost).
- Months to break even = (hardware cost − resale) ÷ monthly saving.
- Capacity check: tokens per second × 3,600 × 24 × 30 ÷ 1,000,000 million tokens a month, assuming it runs nonstop.
- Not included: model quality differences, depreciation beyond the resale you enter, interest, downtime and the value of your time. A local open model and an API model are not equivalent products. Planning estimate, not financial advice.
Planning estimate only, not a benchmark: real memory use and speed depend on the model file, the runtime, the driver and the hardware.
FAQ
How do I compare a local LLM with an API on cost?
Work out your electricity cost per million tokens, subtract it from the API price per million tokens to get the saving per million, and divide the hardware cost you must recover by the saving over your monthly volume.
What API price should I enter?
Use your own blended price per million tokens from the provider's current price page, weighted by your mix of input and output tokens. The 5 in the example is a placeholder, not a quote.
Why does it say there is no break-even?
Because the local running cost per month is equal to or above the API cost at your volume. At low volumes the hardware may never pay back.
What does the capacity warning mean?
Your monthly volume is more than the machine could generate in 30 days of nonstop running at the speed you entered, so the numbers cannot be achieved with one machine.
Does this include the quality difference between models?
No. It compares cost only. A local open model may not match a hosted model on your task, which is a separate question.