Where does your data live?

The first question is never about the model

I spent six years putting data and AI systems into regulated healthcare environments. One question came up in almost every serious conversation: where does our data live, and who can reach it?

Most leaders can't answer it yet, because vendors use "secure", "private", "EU-hosted" and "sovereign" as if they meant the same thing. They don't. The good news is that AI is software. It lives somewhere, runs somewhere and can be reached by someone, like every system you already govern. This two-part essay gives you the map.

Privacy and security are two different promises

Privacy is about who can see your information. Security is about how hard it is to break in.

You can have one without the other. A server under your desk is private, because nobody knows it is there, but you are the one who patches it and watches the door. A large cloud platform can be extremely secure and still not private from the company that runs it. When a vendor says "secure", ask which of the two they mean.

Every digital asset needs a home

Every file, database and application physically sits somewhere: on your own servers, on hardware you rent in a data centre, or on a cloud platform. Think of your data as a photo album. At any moment it is in one of three states:

• At rest: the album is on the shelf. The data is stored.
• In transit: you carry it to another room. The data moves across a network.
• In use: the album is open in your hands. The data is being processed, which is exactly what AI does.

Encryption at rest and in transit is standard today. Protecting data in use is newer, and it matters most for AI.

AI follows the same rules, with one extra question

A large language model is software too. It runs on GPUs in a data centre: your prompt travels there, gets processed, and the answer travels back.

AI adds one question that data alone does not: who owns the model, and whom are you trusting with your data?

• Closed models, such as Anthropic's Claude and OpenAI's GPT (both US), belong to the company that built them. You rent their intelligence and trust that company to keep to its contract. Today the most capable models are closed.
• Open-weight models, such as Mistral (France), Llama and Gemma (US), and DeepSeek and Kimi (China), can be downloaded. You hold your own copy and decide where it runs, so the trust sits with you. They have generally trailed the frontier, but the gap is narrowing fast.

So the practical question is: does your use case actually need frontier intelligence?

Sovereignty means one thing: who can reach your data

Residency is where your data sits. Sovereignty is whose law can reach it.

The US CLOUD Act lets US authorities compel a US provider to hand over data it holds, wherever in the world it is stored. A US company's data centre in Frankfurt is still a US company's data centre: the law follows the company, not the server. It is a legal process, rarely used against businesses, and it cannot force a provider to decrypt data it cannot read. Whether it matters depends on how sensitive your data is.

There are three ways to answer the question well:

1. Choose the jurisdiction: a provider owned and operated in the EU.
2. Remove the ability to read: encryption where you hold the keys, so there is nothing readable to hand over.
3. Hold it yourself: run the model and data on infrastructure you control.

Three phrases to test your vendor on