We will talk about sovereignty and AI sovereignty. I must confess that I’m not an AI person; I’m not a machine learning or data science guy. I do a lot of things about graphs and agentic memory — that’s my favorite topic.
Before that, I was a general architect, focusing for the last five years on self-sovereign identity, cryptography, decentralized systems, and all these interesting things that give people sovereignty. So, what is sovereignty, why does it matter, and what does it mean in terms of AI?
Defining Sovereignty
We will talk about this from the human angle, but also from the angle of the agent. This raises interesting philosophical questions: agents and AI systems are no longer just microservices. If they mimic our identity and represent us, should they have traits of our identity and some kind of rights?
Sovereignty as a term appeared in politics for countries: you control your own territory and make independent decisions without the influence of others in internal and external politics.
Digital Sovereignty in the Web 2.0 Era
What does sovereignty mean for the digital world? We usually talk about Web 2.0 and Web 3.0. Big platforms took away permission from people. In the Web 2.0 revolution, we gave too much power to platforms.
Currently, platforms are the representatives of humans on the internet. If you want to do something, you have to go through a platform. We want to change this setup and return to the state of the internet before the big platforms and the Web 2.0 revolution.
If you want to talk to somebody, you shouldn’t need a man-in-the-middle or a representative. You could do everything on behalf of yourself; you control your own data and your own compute. To represent yourself in a network, we need some form of identity — a crypto identity — to represent yourself in a secure but anonymous way.
The Three Pillars of Digital Sovereignty
Digital sovereignty is mainly about ownership: you own your identity, you control it, and you own your data. There are three big pieces to this:
Data Ownership: You own your data and control access to it.
Software Control: You own or control the software that runs on top of your data to make it useful. In terms of AI, this means owning the code or the software around your agent. You should be able to audit it, change it, and adapt it to your needs.
Compute Sovereignty: You need compute to run this software. This is the most challenging part because running powerful models requires heavy computational power, GPUs, and powerful servers.
The GPU and Infrastructure Challenge
Some people compare access to GPUs to access to power stations. They say you don’t need to run your own power station to consume electricity. However, for electricity, we have laws that protect consumers, multiple suppliers, and better regulation.
With LLMs (Large Language Models), we don’t have this. If a player goes out of the market, you are left with “garbage” that doesn’t work anymore. If a provider like OpenAI decides you are “not so good for them” and cuts your access, you lose everything: your conversations and your workflow.
The Path to Sovereign AI
We need to own the computational power and the models just as we own the data. This opens the question of how to own expensive infrastructure or create a “commodity” for it that a group of people or states could manage.
For example, there are initiatives in Switzerland where they are making their own LLMs aligned with the state, with security guardrails acceptable to Switzerland rather than the US, and accessible to all citizens.
Hosting the model yourself offers the highest level of privacy because you have a “zero sharing” policy. You ensure your data is not contributing to training models that might discriminate against groups you belong to.
The Future: Local and Edge Agents
The goal is to own your data, your compute, and the infrastructure. While AI infrastructure is not easily achievable for everyone, “local edge agents” are.
We can use smaller LLMs that run on our own machines.
We can afford servers with a few GPUs to run middle-sized models.
Some agents will run on LLMs, some on distilled SLMs (Small Language Models) for specialized tasks, and some on classical machine learning or neuro-symbolic approaches.
These can run not just on powerful machines, but even on a user’s phone. For example, at King, we built an agent that made inferences directly on the user’s device (iPhone or Android). This is a target for private and sovereign AI. If you own your phone, you own the hardware where the agent runs, and you have a full data snapshot of all your conversations.
Conclusion
Sovereignty matters because nobody can “cut you off”. You own, change, and do anything you want with your system. Furthermore, if you decide to contribute to training a model, you can do so in a fairer way by receiving benefits for providing your data, creating a more equal data economy.
Ultimately, sovereignty is about protecting our future from surveillance capitalism and the control and domination of big companies or countries.