Announcements have been piling up for some time now: free artificial intelligence solutions offered by certain Swiss players, services made available by public administrations for a few Swiss francs per employee per month, even promises to democratize AI at almost negligible cost…
These announcements deserve our attention. But above all, they should prompt us to ask a few questions, because there is a fundamental difference between an AI that is free for its user and an AI that costs nothing to build and run.
I am not writing these lines as a neutral observer: at Deeplink, we develop, host and continuously improve an AI platform in Switzerland. But it is precisely this position that allows me to talk about real costs rather than assumed ones. The figures that follow come from that experience.
What does the cost of AI really cover?
An artificial intelligence platform is much more than a chat window and a few calls to an algorithm. Behind the service lies an entire value chain:
- research and development of the models;
- development of the platform and its features;
- maintenance, security and operations;
- computing infrastructure, servers, storage and energy;
- engineers and specialized skills;
- inference costs, meaning the actual cost of every interaction with AI models;
- the constant upgrades needed to keep pace with a sector that is advancing at exceptional speed.
A CHF 2 per month offer: what are you really paying for?
Let’s take a concrete example. In a recent call for tenders, a Swiss canton estimated an employee’s monthly usage at around 1 million input tokens, 150,000 output tokens and 18 minutes of audio transcription.
With a mid-sized open-source model hosted in Switzerland, computing alone (inference) comes to around one and a half Swiss francs per employee per month for this profile. But this profile describes standard conversational use: one question, one answer.
With truly agentic assistants, able to chain actions, use tools, search for information and carry out complex tasks, the equation changes scale. An agent doesn’t call the model once, but several times: to choose the right tool, look up information, triage the request, pre-fill a form, check the quality of the answer, filter out abuse. And with every call, it resends its instructions, the description of its tools and the conversation history. In other words, you burn through tokens. A standard question consumes fewer than 10,000 tokens. An agentic conversation of three exchanges consumes nearly 100,000, ten times as much. And a complete procedure, such as filing a complaint, can take 8 to 12 exchanges.
So it only takes part of the usage becoming agentic for inference to exceed two francs per employee per month. In other words, an AI advertised at two francs a month covers, at best, the cost of computing, and nothing more. Not the teams, the platform, security, operations, maintenance or upgrades: everything on the list above still has to be paid for, and it is by far the largest share.
Presenting the cost of inference as if it were the price of AI is a bit like pricing a bus line by counting only the fuel. And with large proprietary models, even that fuel costs considerably more.
Admittedly, some architectures are more efficient than others, and some models require less computing power. But that reduces costs; it does not make them disappear.
Why is an incomplete price misleading?
When it is a public administration announcing these few francs per employee per month, the figure seems simple, reassuring, almost trivial, but it is harmful to the market. If platform development, engineers, infrastructure, operations and maintenance are handled in-house, those costs do not appear in the announced price. They have not disappeared, though: they have simply been moved elsewhere in the public budget.
The figure communicated then says nothing about the real cost of AI. Worse, it is misleading. A publicly announced price, especially one from an institution, quickly becomes a benchmark. It settles into the collective mind, among citizens, decision-makers and businesses: “AI costs a few francs a month.” It is a completely false reference point, yet everything else will now be measured against it: market offers, future budgets, the very value of these technologies.
The problem then goes beyond communication: it also distorts competition. A private company cannot shift its costs into another budget. It has to charge what its service really costs: research, development, infrastructure, teams, security and upgrades. Next to an artificially low public price, its offer immediately looks expensive, even unjustified, when it is simply honest. This risks weakening precisely the local players who invest in building high-performing, sovereign solutions, the very players a Swiss AI ecosystem needs.
Never forget: nobody does anything for free
When a service is offered for free, you simply need to understand where the trade-off lies:
- either you fund it indirectly, notably through your taxes, as we have just seen;
- or you pay for it another way, for example with your data or your usage: in that case, you are partly the product;
- or the free offer serves a broader strategy or broader interests for whoever provides it: locking users in, or demonstrating a market’s appeal to potential investors, who will eventually have to foot the bill. It is a well-known pattern, widely used by US players. But make no mistake: that money will have to be recouped one day, with a profit.
Free vs. real cost: the key takeaways
So don’t confuse free of charge with free of cost. Don’t confuse the displayed price with the real cost. And above all, never lose sight of the economic interests hiding behind an offer.
Every initiative that aims to democratize access to artificial intelligence is welcome, but it is time to move past a certain pricing illusion. AI can be free for those who use it. It never is for those who develop it, train it, host it, maintain it and keep it evolving.
If we want solutions that are truly high-performing, sustainable, sovereign and able to meet users’ expectations, we must start by looking at their real cost and owning it, whether it is paid by the user, the taxpayer or an investor.
Free can be a distribution model, but it will never be a business model in itself.
