The hardest question first: data protection, the Cloud Act, your own infrastructure.
This is the question that holds most people back. That is why it stands here at the front and not in the small print. I show you both paths with the numbers — and you decide for yourself.
Everything in your own house.
Your own servers, your own models, nothing leaves the building. That is technically possible — and it has a price worth knowing before you say it out loud.
Hardware
Between CHF 100,000 and 600,000, depending on size and ambition.
Operation and power
From CHF 2,500 a month upwards — rack servers and electricity included. Open-ended at the top.
People and maintenance
Someone has to run it: set it up, keep it updated, fix faults. Either a position in-house or a contract with someone who can. This item keeps running long after the hardware is paid off.
Redundancy for outages
If the machine goes down, the work stops. Operational safety means a second system that takes over — and that duplicates part of the hardware.
When it pays off
When you are a large company and everyone works on that infrastructure. Then the effort is carried by many shoulders and the sum adds up.
Team plans that protect you.
On the team and business plans of the major providers, your content is not used for training and not stored in that sense. That is the difference from a free account, and it is the reason this path is an option for most companies at all.
Even so: everyone has to decide this for themselves — and depending on your industry the question belongs with your lawyer, not on a website. What I can give you is the experience from practice.
It may well be that you are allowed to use eighty per cent of your data and not the customer records. Then you work with eighty per cent. That is still eighty per cent more than today — and that is exactly what the day is about: a system that separates what may go in from what stays out →
And the open models?
If you want to run an open-source model yourself, the effort is immense. If you have someone else run it, exactly the problem you set out to solve stays: your sensitive data sits with a third party again.
Through the API the costs are lower. But the technical effort becomes enormously more complicated — and that is the item people forget when they do the sums.
The fact is: Anthropic, OpenAI and Google have models so strong that the open ones cannot really keep up in a direct comparison. The exception is a model trained very heavily on your own case — and that is a budget question again.
Working with Claude and GPT is like driving a Porsche. With open-source models it is a VW GTI. Both drive. Just not the same.
Calculated over five years.
The question everything comes down to in the end — and it can be put in figures.
Your own infrastructure
250,000 – 1,500,000CHF over five years
hardware, operation, power,
people and redundancy
Team plans
A few thousandCHF over five years
depending on the number of users
Both protect you if you set them up properly. The difference is the factor between those two figures — and the question of what else you could do with that money.
Simple wins.
My goal was to use AI so that the effect is maximal and the thing stays simple. So simple that it can be installed and maintained without hiring someone for it.
Complicated is always available. With AI, complicated is also very expensive, and then it is out of all proportion to what comes out at the other end.
Times have changed. That is an opportunity — but it takes the courage to think differently. Only you can decide that for yourself.
What this is not
This is not legal advice. What applies to your industry is something you settle with your lawyer and with your provider's terms. What I show you at the AI-Start is how to set up your system so that the line between «may go in» and «stays out» is clean from the start.