What does it cost to self-host an LLM?

Our workstation cost around CHF 8'450, plus power costs of about CHF 3.45 per full-load day. Calculated over two years, that comes to CHF 0.55 per thousand requests at full utilisation, and CHF 4.64 at a realistic 10 percent utilisation. Staff costs for operating it come on top.

Is a self-hosted LLM cheaper than an API?

Only from high utilisation onward. Against Infomaniak's price for a 30-billion-parameter model (90 centimes per thousand requests with 10,000 input tokens and 2,000 output), our machine pays off over two years from 57 percent utilisation, over three years from 39 percent. That equals 14 and 9 hours of full load per day, respectively, every day.

How much power does a local LLM use?

Measured, our two RTX 4090s together drew 406 watts under full load; with the processor and board, we estimate around 520. Per million tokens generated, that is 0.23 kilowatt-hours, or just over 6 centimes at the Zurich peak tariff. Power is therefore a minor item; the purchase cost dominates the calculation.

Why self-host if it costs more?

For control: the data never leaves the building, no contract needs checking, no provider switch needs fearing. At realistic utilisation, this control costs around CHF 310 per month in our calculation over two years, against the cheapest Swiss provider price.

Self-hosting an LLM: the real cost and the break-even

Whether a machine of your own pays off against Swiss token prices comes down to one figure: utilisation. The full calculation to check for yourself, from the parts list to the break-even, with measured power instead of estimates.

What does a self-hosted LLM cost?

Markedly more than an API, unless the machine runs at close to full load around the clock. Our CHF 8’450 workstation beats the Swiss provider price over two years only from 57 percent utilisation onward, meaning 14 hours of full load per day. At realistic utilisation, the machine of your own costs around CHF 310 more per month over two years. That is the price of control.

In brief

Compare against the Swiss provider, not against American frontier prices: our reference request (10,000 input tokens, 2,000 output) costs 90 centimes per thousand at Infomaniak on a 30-billion-parameter model, Swiss data hosting included.
Cross power off your worry list. Measured, it costs just over 6 centimes per million tokens; the purchase cost dominates the calculation.
Think in utilisation, because that is what decides everything. The same machine costs anywhere from 0.6 times to 5 times the provider price, depending on utilisation.
Lift utilisation with night work: overnight batch processing is the most effective lever toward the break-even.
Budget for the person as well. Operation, updates and getting things running again belong on no parts list, and this calculation deliberately does not include them.

This is the final part of our series on local inference, and it answers the question all the previous parts left open: does the machine from the first part pay off against buying per token? Here is the calculation, complete and ready to check, with measured power and the Zurich tariff.

The price you have to beat

The benchmark for this calculation is the cheapest sovereign provider. Infomaniak hosts a model of the same size class and design as ours (30 billion parameters, MoE) in Geneva and Zurich, without the US CLOUD Act, for 5 centimes per million input tokens and 20 centimes per million output tokens. Whether it matches ours in quality, we did not measure; for the price comparison, it is the closest match on the market.

For a typical request from our measurement series (10,000 input tokens, 2,000 output), that means: 90 centimes for a thousand requests. That is the bar to clear. Anyone self-hosting is competing against a world where a single request costs a tenth of a centime and Swiss data hosting is already included.

What the machine of your own costs

You know the purchase cost from the first part: around CHF 8’450 for two RTX 4090s, a processor, 128 GB of system memory and the rest of the parts list.

We measured the power draw. Under full load, both cards together drew 405.7 watts, 28 at idle. With the estimated surcharge for the processor and board, we calculate 520 watts under load and 130 at idle. At the Zurich peak tariff of 27.65 centimes per kilowatt-hour (excl. VAT), that comes to just over 6 centimes per million tokens generated.

That settles the biggest misconception in the debate: power is a minor item. A full-load day costs CHF 3.45 at the socket, while the purchase cost, spread over two years, comes to CHF 11.58 per day. Amortisation dominates, and it runs whether or not anyone is asking anything.

The calculation everything hinges on

At full utilisation, the machine delivers 633 tokens per second, which is just over 27,000 requests per day. That is the best figure in the measurement series, achieved with the MoE model (index 32.1, between GPT-5 mini and GPT-5); the model at GPT-5.1 level from the first part manages 240 tokens per second, and with it there is no break-even at all, even running at full load continuously would remain more expensive than the provider. The daily cost (the purchase cost’s daily share plus power) is spread across however many requests are actually made, and this is exactly where everything is decided:

Utilisation	CHF per 1,000 requests (2 years)	(3 years)	vs. Infomaniak (0.90)
10 %	4.64	3.23	5.2x and 3.6x more expensive
30 %	1.61	1.14	1.8x and 1.3x more expensive
60 %	0.85	0.62	cheaper
100 %	0.55	0.41	clearly cheaper

The break-even sits at 57 percent utilisation over two years, at 39 over three. Translated into hours: the cards would have to compute at full load 14 and 9 hours per day, respectively, every day, including weekends.

What utilisation really means in an office

Now for the uncomfortable translation, to check for yourself. A business with twenty to thirty employees uses the machine during 40 of the 168 hours in a week, and during those hours perhaps half its capacity is occupied on average. That works out to 12 percent; we calculate with around 10 percent utilisation. At night and over the weekend, the purchase cost keeps counting, just without any requests.

At 10 percent, the machine of your own costs five times what the provider does. In francs: around CHF 310 more per month than the same requests would cost at Infomaniak.

Important for context: this calculation deliberately leaves out the largest item. The person who operates the machine, keeps it updated and gets it running again on a Friday evening appears on no parts list. Every hour they need shifts the comparison further in the provider’s favour.

How the break-even can be shifted

The calculation is not a law of nature; two levers can move it.

Night work raises utilisation. The machine’s most expensive property is its dead hours. Anyone running batches overnight, checking contracts, indexing files, preparing reports, turns idle time into requests. The extreme case from part one fits exactly here: the 284-billion model is too slow for the screen and ideal for the night. A machine that answers the team during the day and works through the case archive at night comes seriously close to the break-even, and since our calculation applies the peak tariff throughout, the cheaper night-time electricity works in your favour on top.

A longer horizon stretches the purchase cost. Over three years instead of two, the break-even falls from 57 to 39 percent. The counter-argument: three years is a long time in this technology. The RTX 4090 on our list was replaced after just over two years by a successor with a third more memory, and model quality within each size class keeps rising. Anyone depreciating over a longer period is betting that today’s machine still suffices the day after tomorrow.

When the machine of your own is still the right call

After all the figures, reasons remain that no cost calculation captures, and in practice, buying happens because of them:

Data that must not leave the building, whatever a contract promises. Professional secrecy, where even reviewing provider contracts already means effort and residual risk. Independence from price increases, model retirements and contract changes, as the provider series documents. And the certainty that the machine will still do tomorrow exactly what it does today.

Anyone buying for these reasons is buying correctly, and now knows what it costs. Anyone who just wants to save money is better served by the Swiss provider.

How to calculate your own case

01

Count your requests, do not estimate them

How many AI requests does your business make per day today, and how long are they? Anyone not yet using AI starts with a month of API usage and reads the figure off the bill.
02

Derive the utilisation from it

Daily requests divided by the capacity of the machine you want. If the result sits below a third, the purchase does not pay off on price in our pricing world.
03

Look for night-time load before you buy

Is there batch work that goes undone today? Case archives, records, recurring reports? That is the lever that turns the calculation around.
04

Budget the operation in full

Who keeps the machine running, and what does that time cost? Only with this line item is the comparison against the provider price complete.

This closes the series. Parts one through four carry the foundations of this calculation: the machine, the memory sizing, the quantisation and the engines, each with all its measured values.