Souverana.

Insights

AI, explained soberly.

Swiss-specific guides to sovereign AI, data protection and architecture, written for decision-makers.

Law & Regulation

6 parts

  1. Part 01 Guide

    AI regulation in Switzerland: what already applies

    Waiting for a Swiss AI law is not worth it. Almost every AI project is already measured against law that has been in force for years.

    · Updated

  2. Part 02 Guide

    revFADP and AI: what the law actually requires

    The fine of up to CHF 250’000 falls on the person who approved the tool, and only exceptionally on the company. That makes the approval a matter for management.

    · Updated

  3. Part 03 Guide

    Professional secrecy and AI: what may go to the cloud

    'That does not work for us' is often the answer to AI in law firms and medical practices. The criminal-law article on professional secrecy (Art. 321 of the Swiss Criminal Code) is drawn more narrowly, and most of daily work falls outside it entirely.

    · Updated

  4. Part 04 Template

    AI governance: one page is enough to start

    An AI policy, the internal rule for how to use AI tools, that takes three months to finish protects no one today. Shadow AI, the use of unapproved tools, grows for exactly as long as the document sits in draft.

    · Updated

  5. Part 05 Guide

    EU AI Act Switzerland: role first, then duties

    Most Swiss companies owe the EU AI Act almost nothing. The provider role gets expensive, that is, whoever ships an AI tool under its own name. Whoever lands in it usually notices too late.

    · Updated

  6. Part 06 Guide

    Shadow AI in the workplace: what reduces it

    A customer letter in a translation service, a note-taking bot in a video call: many companies have not yet decided how to handle AI. Their employees have.

Buying inference

3 parts

  1. Part 01 Analysis

    What LLMs really cost: the price per token is misleading

    Comparing price lists is time wasted: token consumption, cache hit rate and quantisation decide what you actually pay.

    · Updated

  2. Part 02 Analysis

    Prompt caching: why agents cost more to run in Switzerland

    With every message to a language model, the whole conversation so far is sent again. For long conversations and AI agents, that makes everything slow and expensive. Some providers bill the already-processed history up to 50 times cheaper; the Swiss hosts we checked do not. Here is how it works.

  3. Part 03 Technical

    Prompt caching in operation: what a host must set up

    Whether prompt caching works at a host depends on its technology: where the cache lives, how long it keeps entries and whether the host measures hits. This article shows the building blocks so that you can ask a provider the right questions, and explains why the Swiss model Apertus 70B needs about 30 times as much cache memory than DeepSeek V4.

Provider Data-Protection Check

5 parts

  1. Part 01 Analysis

    Is ChatGPT GDPR-compliant? What OpenAI's contracts cover

    A Zurich law firm reviewed the contracts of OpenAI, the provider of ChatGPT: none of them holds up for lawyers, doctors and other holders of professional secrecy. What that means for companies, and the routes around it that the review does not name.

    · Updated

  2. Part 02 Analysis

    Anthropic Claude and Swiss data protection: what to check

    Claude, Anthropic's AI assistant, works for business data, not for professional secrecy, according to the law firm VISCHER. The Zero Data Retention promise, meaning no storage of inputs and outputs, has not held for two models since June 2026.

    · Updated

  3. Part 03 Analysis

    Google Gemini & data protection: what SMEs need to know

    Google is the only large cloud provider with an AI web search usable with professional-secrecy data, but makes access cumbersome. The professional-secrecy addendum alone is not enough while the standard web search stays active.

    · Updated

  4. Part 04 Analysis

    Using Microsoft Copilot safely: what data protection allows

    Microsoft settles the contracts almost completely, except for web search. What Swiss companies need to check on Copilot and Azure OpenAI.

    · Updated

  5. Part 05 Analysis

    Proton Lumo: what privacy-friendly AI looks like

    Proton's AI assistant Lumo 2.0 relies on open models, its own servers and encrypted chats. Two catches remain: web search and the missing API.

    · Updated

Sovereign AI

5 parts

  1. Part 01 Guide

    What sovereign AI actually means for Swiss companies

    If you sell discretion, you have a problem the moment a client asks where your data goes when you use AI. Sovereign AI answers that question, and the degree of control can be set like a dial.

    · Updated

  2. Part 02 Strategy

    AI strategy for SMEs: the questions to answer first

    Most AI strategies are tool lists. Nine questions for senior management set how much control a Swiss SME keeps over its AI. No ready-made framework.

    · Updated

  3. Part 03 Guide

    Sovereign AI infrastructure: why the model is the easy part

    The 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.

    · Updated

  4. Part 04 Platform

    Sovereign AI platform: why it binds more than the model

    The model is easy to swap. The platform, the software your team uses for AI every day, holds your chats, files and knowledge, and that is what binds you longest. How to choose it sovereignly.

    · Updated

  5. Part 05 Integration

    Sovereign AI integration: when AI starts to act

    The moment an AI agent writes in your systems, a mistake becomes an incident. Where agents are connected to your software, your IP (intellectual property) is built too. How to keep both under control.

    · Updated

Local Inference

5 parts

  1. Part 01 Analysis

    Self-hosting an LLM: GPT-5 performance for CHF 8’500

    A machine with two graphics cards (RTX 4090), around CHF 8’500, three weeks of measurements. Which tasks a machine of your own handles, which it does not, and where the limit sits.

    · Updated

  2. Part 02 Technical

    VRAM for LLMs: how to size it correctly before you buy

    How to calculate how much graphics memory you need from parameter count, quantisation, context length and concurrent requests. With a calculator and the measurements from two RTX 4090s.

    · Updated

  3. Part 03 Technique

    LLM quantisation: how it works and how to use it well

    Quantisation stores a model's numbers with fewer bits, so that a large model fits onto affordable graphics cards and answers faster. The article shows how this works, how to quantise a model yourself and how to find the level at which quality still holds.

    · Updated

  4. Part 04 Technical

    Ollama, llama.cpp or vLLM: choosing the right engine

    Ollama, llama.cpp or vLLM: the choice depends on two questions. How many people work on the model at the same time, and how much graphics memory do you have relative to the model you want? Three tools, measured on the same hardware.

    · Updated

  5. Part 05 Analysis

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

    Whether your own machine pays off against a Swiss provider's prices comes down to one figure: utilisation, meaning the share of possible requests that the machine processes. The full calculation to check for yourself, from the parts list to the break-even (the point from which the purchase pays off), with measured power.

    · Updated

Start the AI Sovereignty Check