This week I had the honor of speaking at the UN during the Open Source Week, and I discussed about the importance of “true open source AI”, and sovereignty. It was amazing to be able to share insights and conversations with amazing individuals. Here are my remarks published (as seen in other media)
If you are interested in seeing it live, here is the recording (Right after Yan Lecunn): https://webtv.un.org/en/asset/k14/k14ej1ucqu
I was starting to write my remarks for today when the news came: “Anthropic banned from offering Fable and Mythos models to non-US citizens due to security concerns” And suddenly writing this became much much easier.
Imagine a future where every individual doing any kind of work can only come from three companies and three companies only. And these three companies can dictate who these individuals are, their skills and their salaries. And they dictate that for every company and country in the world. Well, that is the future we are creating.
The Global Digital Compact gives us a clear direction: an open, safe and secure digital future, with artificial intelligence developed for the benefit of humanity. Wherever the humans are.
We often speak as though AI begins with a model. It does not. It begins with data, infrastructure, institutions and people.
If those foundations are concentrated, opaque and inaccessible, the AI built on top will reproduce those conditions—only faster, and at far greater scale. In other words: Let’s not start with the weights. But let’s weigh the infrastructure.
I would like to make three points today:
Interoperability is a condition for participation.
Sovereignty is a condition for continuity.
And private AI is a condition for operational responsibility.
1. Interoperability is freedom
Interoperability is often treated as a technical detail. It is not. It is a condition for freedom.
A hospital should be able to use one storage system, another analytics engine and a locally adapted AI model—without copying sensitive patient data into a proprietary island.
A government should be able to replace a technology provider without having to rebuild an essential public service.
A researcher anywhere should be able to use open formats and interfaces and adapt them to local languages, local data and local needs. Seamlessly.
Therefore Open-source AI cannot simply mean publishing model weights.
An open model running on proprietary data formats, proprietary orchestration, proprietary cloud interfaces and proprietary governance is still a locked system—with an open door painted on it. Especially when any party can unilaterally condition access or price.
We need openness across the full, complete architecture:
Private companies and institutions must see this in very practical terms. Open table formats like Iceberg, open engines and catalogs (like Polaris), open source compute engines and APIs, and common governance policies allow different teams and technologies to work on the same governed data—without forcing unnecessary copies or dependence on one vendor. Or forcing everyone to move your data to a single provider in order for it to work.
Interoperability allows an institution to replace one component without chaging everything. Or without forcing all your data to go into one single place or vendor.
And it prevents today’s procurement decision from becoming tomorrow’s permanent dependency. This is fundamental for the future, PRESENT of the agentic economy we are building.
2. Sovereignty is no longer optional
For years, digital sovereignty was treated as a policy preference—something desirable for governments, perhaps, but too complicated or expensive for most organizations.
That position is no longer credible.
AI is becoming a fundamental building block of public administration, healthcare, education, defence, finance and critical infrastructure. And when that happens, control over that technology becomes a matter of continuity. The technology itself becomes the critical infrastructure.
A private provider can change the price of a token. It can change a rate limit, retire a model, modify its licence, or the quality of the output. It can alter the terms of service or decide that a particular capability will no longer be available in a particular market.
A government can impose an export restriction. A geopolitical dispute can affect access overnight.
And an institution that believed it had bought a technological capability suddenly discovers that it had only rented permission to use it.
We have just seen a powerful example last week.
Organizations and entire countries are building processes around capabilities that they do not ultimately control. Whether the decision originates in a boardroom, a regulator’s office or a geopolitical confrontation, the operational result can be the same: access disappears.
Imagine that dependency inside a hospital.
Inside a tax authority.
Inside an electricity grid.
Inside an education system.
Your national AI strategy cannot depend on somebody else’s terms of service.
Your AI sovereign strategy must answer the following questions:
1. Where does our data reside?
2. Who can access it and how?
3. Which models can use it?
4. Can we move the workloads?
5. Can we replace the models instantly?
6. Can we audit and inspect the system?
7. And most importantly: Can we continue operating if a provider changes its commercial or political position?
Real sovereignty is not isolation. It is the ability to participate in a global ecosystem without surrendering control of essential capabilities. What is the solution for that?
Well, Open source is central to the answer because it converts dependence on a single supplier into participation in a shared ecosystem. But open source alone does not create sovereignty.
Open source removes the toll. You still need to build the road.
We must invest beyond pilots, but in permanent capability. Not only in innovation, but in what happens after, in maintenance. Not only in model access, but in institutional independence. Day two operations. And that leads me to my third point
3. Private AI makes responsibility operational
My third point is private AI.
Private AI does not mean Private models. It means Open in its foundations. Private in its operation. Accountable in its outcomes.
Institutions cannot simply send all their sensitive data to an external black box. They need to bring AI to the data, rather than continually moving data to the AI.
They need the ability to run models in a public cloud, a sovereign cloud, a private data centre or at the edge—while applying consistent security, governance and audit controls.
This is why interoperability, sovereignty and private AI are inseparable.
Interoperability makes replacement possible.
Sovereignty makes the decision possible.
Private AI makes controlled operation possible.
Open models can be deployed privately, adapted to local knowledge and used within clearly defined boundaries. Investment on true open source AI is a key element to realise fair AI for humanity.
Responsible deployment requires identity and access controls, data lineage, evaluation, bias and security testing, read-teaming tools, continuous monitoring, transparent incident reporting, audit logs, human oversight, appeal mechanisms and the ability to shut a system down.
All that should become digital public goods.
And this shared safety layer is one of the most important missing pieces in today’s AI ecosystem.
4. From impressive demonstrations to dependable systems
Finally, we must move from innovation to deployment.
Our industry loves demonstrations. Wow factor. But humanity does not benefit from demonstrations. It benefits from systems that work on Monday morning, serve real people and survive an audit on Friday afternoon. From HIPAA to GDPR compliance.
That requires deployment discipline:
A clearly defined public benefit.
Legitimate and representative data.
Measurable performance.
Security testing.
Continuous monitoring.
A route for human appeal and redress.
And a plan to retire the system when it no longer performs as intended.
Governments should protect rights and use procurement to require open standards, interoperability, portability, auditability and credible exit plans.
Researchers and civil society should test claims, expose failures, localize systems and ensure that affected communities have a meaningful voice.
And the United Nations can help create common ground: coordinating standards, sharing safety resources, supporting capacity and ensuring that countries currently at the edge of AI development become active participants in its future.
Closing
If we get this right, open source will not merely make AI cheaper. It will make AI more contestable, adaptable, locally relevant and worthy of trust.
AI for humanity means that a teacher can adapt a system to a local curriculum. That a public-health agency can use sensitive information without surrendering control. That a small country can participate in the AI economy without renting its future to a company valued at 10 times its Gross Domestic Product.
And that a citizen can know when AI is being used, understand the basis of an important decision and challenge it when necessary.
Open source gives us the possibility of building a commons.
Interoperability keeps that commons connected.
Private AI allows institutions to operate responsibly.
And sovereignty ensures that humanity—not a contract, a pricing page or a distant political decision—remains in control.
Sovereignty is not technological nationalism. It is not isolation. It is the foundation of resilient, democratic and human-centered AI. And it is no longer optional.
Thank you.