Knowledge · Local AI

Cloud or local AI: when does your own hardware pay off?

The cloud costs per request, your own hardware costs once. Somewhere the lines cross.

For your volume, the calculator shows what the cloud API and what your own hardware cost per month, and from how many requests the balance tips. All assumptions are stated openly on this page, the calculation runs in your browser and nothing is transmitted. And where the honest answer is «cloud», the calculator says so.

To the calculator Which hardware? To the hardware calculator

The calculator

Your volume, both options, one figure

Four inputs are enough. The classes behind them (API prices, hardware tiers) are declared orders of magnitude, not quotes; what lies behind them is explained in the section «How the calculator works» below.

Last updated on 17 September 2026 · maintained quarterly together with the hardware calculator. API prices change often; check them on the linked price pages before deciding.

Transparency

How the calculator works

No magic, one formula from four figures, so that you can check each of them and replace it with your own:

The cloud side

Requests times size gives the token volume per month; times the blended price of your model class gives the API bill. The classes (~CHF 0.50 / 1.50 / 3.50 / 8.00 per million tokens, as of September 2026) are averaged from the official price lists of OpenAI, Anthropic and Google: three parts input to one part output, dollars to francs at 0.82, introductory prices counted as the current price. The newest top models (GPT-6 Astra, Claude Fable 5.1) are at around CHF 16.50 per million tokens. Prices change often, which is why the price pages are linked below.

The local side

The purchase depreciated over 36 months, plus a base load of around CHF 8 a month for continuous operation and electricity of roughly CHF 0.30 per million tokens. The surprising truth of this line: electricity is almost never the cost driver, depreciation is. The hardware tiers match our Ollama hardware matrix; which tier your task needs is answered by the hardware calculator.

For reference

Three example calculations

The same formula, three typical cases. Calculated with the mid-range API (CHF 1.50 per million tokens) and the hardware tiers from the calculator.

Scenario	Volume	Cloud API	Own hardware	Result
Small business emails, drafts, occasional summaries	1'000 medium requests (5 million tokens)	CHF 8/month	CHF 79/month (mid-range, CHF 2'500)	Cloud, clearly. At small volumes the API costs less than a lunch; local only pays off here for data protection reasons.
Document flow analysing receipts, contracts, reports	5'000 document requests (100 million tokens)	CHF 150/month	CHF 107/month (mid-range, CHF 2'500)	Local ahead. The break-even is at around 3'200 documents a month; below it the cloud stays cheaper, above it the balance tips.
Continuous operation agents and analyses around the clock	20'000 document requests (400 million tokens)	CHF 600/month	CHF 253/month (workstation, CHF 4'500)	Local is cheaper, and the gap grows with every additional job. From here the calculation is a real argument.

Calculated against a top model (CHF 8.00 per million tokens), local would look brilliant everywhere, but that would be dishonest: a local mid-range model does not replace a top cloud model. Compare options that both genuinely solve your task; whether they do is clarified by a test with real tasks.

Honestly assessed

What the calculation deliberately leaves out

Four items that are in no cost formula and still decide the right way.

Maintenance and safeguards

Your own AI server needs updates, backups and monitoring. If nobody in-house can take this on, it is bought in as a service or you stay in the cloud, where it is included in the price. This item regularly tips close calculations back towards the cloud.

The quality difference

The calculator compares cost, not capability. The top cloud models remain ahead of local models on demanding tasks. The comparison is only fair between options that demonstrably both solve your task, and that is checked with real documents, not with data sheets.

Data protection is the real reason

Most businesses that go local do not do so because of cost: customer files, personnel data and documents under professional secrecy should not leave the building. This reason applies regardless of what the calculation says, and it is legitimate even when local is more expensive.

Subscriptions are a different comparison

ChatGPT or Claude subscriptions for individual employees cost fixed amounts per person and month, regardless of volume. The calculator compares the API, that is the machine volume behind automated processes. For the chat workplace of individual people, a subscription usually remains the simplest option.

Frequently asked questions

FAQ: what AI costs in day-to-day business

Is local AI cheaper than ChatGPT?

At small volumes almost never: a few thousand medium-sized requests a month cost ten to thirty francs via the API, while your own hardware costs many times that in depreciation alone. Local only becomes cheaper at a permanently high volume, typically from over ten thousand medium requests or a few thousand analysed documents a month. The more common reason for local AI is not cost but data protection and independence.

At what point does your own AI hardware pay off financially?

As an order of magnitude: a suitable mid-range computer (around CHF 2'500) costs about CHF 77 a month including base load, depreciated over three years. Against a mid-range API, the break-even is roughly 65 million tokens a month, which is around 3'200 analysed documents or around 13'000 medium requests. The calculator above works it out for your figures.

How much electricity does local AI use?

Less than most people expect: depending on load and model, roughly CHF 0.10 to 1.00 per million tokens, plus a base load of a few francs a month in continuous operation. The cost driver for your own hardware is almost never electricity but the depreciation of the purchase.

What does the ChatGPT or Claude API cost?

Billing is per million tokens, and prices differ by more than a factor of ten between model classes: small models cost roughly CHF 0.50 per million tokens as a blended price, the mid-range around CHF 1.50, the large all-rounders around CHF 3.50, top models CHF 8.00 and more (as of September 2026). Prices change often; rely on the providers' linked price pages, not on remembered figures.

Buy AI hardware or rent a GPU server?

Renting is the right way to start: a GPU server is available by the hour or by the month, so quality and volume can be tested on a real case before money is tied up. Buying pays off for permanently plannable operation: compared with a continuously running GPU server at around CHF 300 a month, a purchase of CHF 2'500 to 4'500 has paid for itself after about 9 to 15 months (as of October 2026, electricity and maintenance not included). The calculator above has both options built in.

Why is the calculation in tokens?

Tokens are the billing unit of language models: a part of a word, roughly three quarters of a word in German. A short request including the answer is around a thousand tokens, a medium one five thousand, analysing a longer document around twenty thousand. Requests times size gives the monthly volume on which both cost calculations are based.

Sources to check

OpenAI: API pricing, the official token prices per model
Anthropic: API pricing, the same for the Claude models
Google: Gemini API pricing, the same for the Gemini models, including time-limited introductory prices
Our price history (CSV), the raw prices per model and reference date from which the classes are calculated; the derived classes per date as JSON
ElCom electricity price overview, the official Swiss electricity tariffs behind our electricity assumption (in German, French and Italian)
Our hardware calculator, which hardware tier your task needs at all

Which use case needs which hardware: local AI use cases. More detail in German: the German page and the comparison of AI operating options over 12, 24 and 36 months.

Next step

The calculation is the beginning, not the answer

Whether local AI pays off for your business depends on your task, your volume and your data protection obligations. In a free initial consultation, in German or English, we work through your case together, and if the honest answer is «stay in the cloud», we say so. Measure first, then invest. If you want help with the implementation, see what we can take on for you.