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

What does a self-hosted LLM cost?

Your own machine for AI costs more than a Swiss provider, unless it runs close to full load around the clock. Our CHF 8’450 workstation undercuts the price of Infomaniak, a Swiss host, for a 30-billion-parameter model (12 August 2026) only from 57 percent utilisation over two years. At 10 percent, it costs CHF 310 more per month. In return, the data stays in your own operation.

In brief

• Compare against a Swiss provider. Infomaniak, a Swiss host with data held in Geneva and Zurich, charges 90 centimes per thousand requests for our reference request on a model with 30 billion parameters?ParametersThe learned numerical values of a language model. Their number describes its size: "8B" means eight billion parameters. Larger models can do more but need more GPU memory.. The request comprises 10,000 input tokens?TokenThe smallest billing unit of a language model; a word is roughly one to three tokens. Providers charge per million tokens, separately for input and output. and 2,000 output tokens, and a token is a text fragment of a few letters.
• Power barely registers. It costs just over 6 centimes per million tokens generated (graphics cards measured, Zurich peak tariff). The purchase of the hardware makes up the main part of the cost.
• Think in utilisation. It is the share of requests that the machine actually processes, measured against what it could manage around the clock. This figure alone decides: over two years, the same machine costs 0.6 times the provider price per request at 100 percent utilisation and 5 times at 10 percent.
• Raise utilisation with overnight work. Batch processing overnight, meaning many jobs one after another without intervention, is the most effective way to reach the break-even, the point from which your own machine is cheaper than the provider.
• Budget for the person who runs the machine. Operation, updates and getting things running again after failures appear on no parts list, and this calculation deliberately does not include them.

This article is the fifth and final part of our series on local inference?InferenceRunning a trained model in production: a request goes in, an answer comes out. Inference is what you buy from a host per token., meaning the operation of AI models on hardware of your own. The first part described a machine with two graphics cards. This part answers whether it pays off against buying per token. The calculation is set out in full and ready to check, with measured power and the Zurich tariff.

The price you have to beat

The benchmark?BenchmarkA standardised test with fixed tasks used to compare models. A good benchmark result says little by itself about your own task. for this calculation is the cheapest sovereign provider↗Sovereign AI · Part 3Sovereign AI infrastructure: why the model is the easy partThe language model is the easiest part of your AI to swap. What ties you to one provider are the models that read your documents and make them searchable (OCR and embeddings). Which Swiss hosters offer every component.Read article. Infomaniak hosts a model of the same size class and design as ours (30 billion parameters, Mixture of Experts?MoEMixture of Experts: a model that stores many sub-models ("experts") but computes only a few of them per token. All parameters are stored; only a fraction are used for any given calculation. design, or MoE for short: a model that uses only a small part of its parameters per token) in Geneva and Zurich, without the US CLOUD Act?US CLOUD ActA 2018 US law: American providers must hand over data under their control regardless of where in the world the servers stand. It also reaches the European data centres of US corporations., for 5 centimes per million input tokens and 20 centimes per million output tokens. We did not measure whether its quality matches ours; 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 comparison price. Anyone self-hosting is up against an offer in which a single request costs a tenth of a centime and Swiss data hosting is already included.

What your own machine 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 watts at idle. With the estimated surcharge for the processor and board, we calculate 520 watts under load and 130 watts at idle. At the Zurich peak tariff of 27.65 centimes per kilowatt-hour (excl. VAT), power costs just over 6 centimes per million tokens generated.

Power is therefore a minor item. A full-load day costs CHF 3.45 in power, while the purchase cost, spread over two years, comes to CHF 11.58 per day. The largest share of the cost is depreciation, and it runs whether or not anyone makes a request.

The calculation everything hinges on

The table uses the MoE model, because it reaches the highest throughput in the measurement series: at full utilisation, the machine delivers 633 tokens per second with it, which is just over 27,000 requests per day. It reaches index 32.1 (between GPT-5 mini and GPT-5). The model at GPT-5.1 level from the first part manages only 240 tokens per second, and with it there is no break-even at all: even continuous full load 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. The table shows how strongly utilisation changes the price per request:

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. In hours: the cards would have to run at full load for 14 and 9 hours a day respectively, every day, weekends included.

What utilisation really means in an office

What this means for an office can be worked out. 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 round down to 10 percent utilisation. At night and over the weekend, the depreciation costs keep running, without any requests coming in.

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

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.

Overnight work raises utilisation. The machine’s most expensive property is its unused hours. Anyone who has jobs worked through in overnight batches, such as checking contracts, indexing files or preparing reports, puts idle time to use for 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 reaches a higher utilisation and therefore sits closer to the break-even. Since our calculation applies the peak tariff throughout, night-time electricity also costs less than calculated.

A longer depreciation horizon lowers the daily 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 will still be enough the day after tomorrow.

When your own machine is still the right call

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

• Some data must not leave your own operation, whatever a contract promises.
• With professional secrecy?Professional and official secrecyThe criminally protected duty of confidentiality of authorities (official secrecy, Art. 320 of the Swiss Criminal Code) and a list of professions named in the law (Art. 321), such as lawyers, doctors and pharmacists. It applies independently of data protection law and reaches further. Banking secrecy sits separately in the Banking Act., even reviewing provider contracts already means effort and residual risk.
• Your own machine makes you independent of price increases, model retirements and contract changes, as the provider series documents.
• The machine will still do tomorrow exactly what it does today.

Anyone who buys for these reasons is making the right purchase, 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

With these four steps, you set up the calculation for your business:

1. 01

   Count your requests

   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 instead of estimating it.

2. 02

   Derive the utilisation from it

   Divide the daily requests by the daily capacity of the machine you want (requests per day at full load). If the result sits below a third, the purchase does not pay off at today's Swiss provider prices.

3. 03

   Look for night-time load before you buy

   Is there batch work that goes undone today? Case archives, records, recurring reports? It raises utilisation the most.

4. 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 provide the foundations of this calculation: the machine, the memory sizing, the quantisation?QuantisationCompressing a model by lowering the numerical precision of its weights (e.g. FP8 or 4-bit). Operating it becomes markedly cheaper; past a certain point, answer quality drops measurably. and the engines, each with all its measured values.

Ever thought about buying your own machine for AI? All data in your own operation, nobody reading along, and money saved on top?

We worked out what that means financially, against a sovereign Swiss provider, not a US cloud. At Infomaniak, a Swiss host, a thousand requests to a 30-billion-parameter model (10,000 input tokens, 2,000 output) cost 90 centimes, with data held in Geneva and Zurich.

Your own machine, by contrast: CHF 8'450 in hardware, 520 watts under full load (the graphics cards measured, the rest estimated), Zurich peak tariff of the utility ewz. From 57 percent utilisation (the share of possible requests that the machine processes), meaning 14 hours of full load per day, your own machine becomes cheaper than the provider over two years (break-even). A normal office runs at around 10 percent and therefore pays about CHF 310 more per month. Power barely registers here (just over 6 centimes per million tokens); the purchase cost dominates.

So nothing is saved. Your own machine is a quantifiable premium for keeping every byte inside your own operation, and for most businesses it comes out smaller than expected.

The article carries the full calculation to check for yourself: both depreciation horizons, the utilisation curves, and what is deliberately left out, namely the person who runs the operation. Link in the comments.

#AI #DataSovereignty #Costs #Switzerland