Online model or local LLM
For , buying instead of paying for pays for itself in 1 year and 5 months, electricity included, counting 8% a year on money invested elsewhere.
Example: 20 questions per person per working day, 3,000 tokens read and 500 written each time.
The return can be changed: drag it, use the arrow keys, or tap it to type a value.
- Online
- Local
Move along the curve with the pointer or the arrow keys to read each month.
- Machine price
- €6,599
- Largest open model on this machine
- Mistral Medium 3.5 128B
- Tokens a month
- 74 million
- Million tokens on the API
- €5.62
- API, a month
- €413
- Electricity, a month
- €13
- Lost returns over 5 years
- €3,097
- Saved by local over 5 years
- €19,500
A simulation, not investment advice. Prices read on 4 November 2026, at 1.1179 dollars to the euro (ECB, 3 November 2026). Electricity at €0.20 per kWh. Machine counted over 5 years, without resale.
Method
- 01
Online
The token volume of the usage at the public API price: reading, writing and cache, each at its rate.
- 02
Local
The price of the machine, then its electricity: its power times its working hours.
- 03
Lost returns
Each euro spent could have earned a return: both curves grow at the chosen rate.
Questions
Is a local model as good as an online one?
Not necessarily: the best online models do not run at home. But a local model keeps your data with you, which no price replaces when it must not leave.
What if I have a subscription?
For one person, a subscription often costs far less than the API, at a fixed price. Its limits are not published in tokens, so the tool compares prices per token.
Will the machine handle this volume?
Not always. The tool checks memory, not speed. Several people at once or long documents may need more than one machine: measure on your own usage before buying.
Where do the prices come from?
The APIs, from the official pages of Anthropic, OpenAI, Google and Mistral. The machines, from the press, since Apple and NVIDIA do not show them on their product pages: they vary by reseller.
Which memory for which model?
The details give the largest open model that fits the machine at 4 bits. Our LLM memory tool does the calculation for each model, context and number of users.
Updated 4 November 2026