At VivaTech 2026, one of the most revealing conversations wasn’t about AI models, autonomous agents, or the next generation of LLMs. It was about something much less visible: where artificial intelligence actually runs.
Behind the spectacular demonstrations, a more fundamental battle is being waged: who controls the AI infrastructure?
The End of the Global Cloud Illusion
For fifteen years, the cloud has been built on a simple promise: centralize to gain efficiency.
Companies moved their data to a handful of global hyperscalers, taking advantage of unprecedented economies of scale. The physical location didn’t matter as long as the services worked.
AI is changing this logic.
Now, the question is no longer just where the data is stored, but:
- Who oversees them?
- Under which jurisdiction?
- Who can access the templates?
- Who can suspend the service?
The example cited on stage was deliberately provocative: What would happen if Washington one day decided to cut off access to certain European cloud services?
It doesn’t matter whether this scenario is likely or not. The mere fact that governments, banks, and large companies are now asking themselves this question is, in itself, a major strategic shift.
Digital sovereignty has evolved from a political debate to an operational issue.
The Emergence of Sovereign AI
For a long time, sovereignty simply meant:
- a local data center,
- data hosted within the country,
- and national carriers.
This definition is no longer sufficient.
Companies are discovering that a data center located in Europe but operated by a company subject to U.S. law remains potentially exposed to extraterritorial constraints.
Sovereignty is now being redefined around three layers:
- Data Location
- The Identity of the Operators
- Applicable legal jurisdiction
This trend explains the emergence of new architectures:
- traditional public cloud,
- a sovereign European cloud,
- national cloud,
- cloud operated by local third parties,
- dedicated private infrastructure.
The future of AI therefore does not seem to be heading toward total centralization but toward controlled fragmentation.
The European Paradox
Europe finds itself in a unique position.
On the one hand, it wants to reduce its technological dependence.
On the other hand, the most advanced technologies still largely come from the United States.
The dilemma is clear:
How can Europe preserve its sovereignty without cutting itself off from global innovation?
The answer that seems to be emerging is not to replace American technologies but to embed them within local governance frameworks.
In other words:
The goal is not to recreate OpenAI, Oracle, or Anthropic in Europe, but to control the conditions under which their technologies are used.
This distinction is essential.
The real problem isn’t AI. It’s the data.
One of the most interesting observations from this session concerns the quiet failure of many AI projects.
Thousands of pilot projects are launched.
Very few make it to production.
The reason is generally not related to the models.
It’s related to the data.
In most large organizations:
- The information is scattered,
- systems do not communicate with one another,
- there are multiple data repositories,
- and historical data is fragmented.
Result:
Companies are investing in AI even though they haven’t yet laid the necessary groundwork to make the most of it.
So the issue is no longer:
“Which AI should I use?”
but rather:
“Can our data assets be utilized by AI?”
This distinction explains why certain highly regulated industries, particularly banking, are now among the most advanced.
Their advantage is not technological.
It is organizational.
The Return of the Local in a Global World
Another strong sign: the geography of AI is evolving.
The dominant model until now relied on a few massive data centers designed to train the models.
This approach is likely to continue.
But inference—that is, the day-to-day use of models—could gradually move closer to users.
A two-tier architecture is emerging:
Tier 1: Mega-training centers
A few giant facilities concentrate the computing power needed to train the models.
Level 2: Micro-inference centers
Much smaller facilities—sometimes deployed directly at customers’ sites or within their countries—run models locally that have been enriched with their private data.
This development is significant.
It means that AI could replicate a dynamic already observed in telecommunications:
- a global network at the core,
- distributed intelligence at the edge.
The Next Challenge: Energy
Behind sovereignty lies another battle.
Energy.
As AI becomes a critical national infrastructure, data centers are increasingly resembling reverse power plants.
Their location now depends less on proximity to major cities than on:
- access to electricity,
- availability of grids,
- access to renewable energy,
- cooling capacity.
The analogy drawn during the discussion is particularly apt:
Data centers are becoming public utilities, just like water, electricity, and transportation networks.
This shift implies a growing role for governments in planning AI infrastructure.
What this means for businesses
Most executives continue to view AI as merely a tool.
The reality is different.
AI is gradually becoming an architectural issue.
The organizations that will derive the most value from AI over the next five years will likely not be the ones that have tested the most models.
These will be the ones who have:
- clear data governance,
- a defined data sovereignty strategy,
- a scalable infrastructure,
- business applications already connected to their data repositories.
The VivaTech discussion ultimately reveals a profound shift.
The first decade of the cloud was all about centralization.
The next decade of AI could be one of controlled relocalization.
Artificial intelligence will remain global.
But the infrastructure that makes it possible will become increasingly local, regulated, and strategic.
And in this new balance, the decisive question may no longer be:
“What model do you use?”
But:
“Who owns the infrastructure he’s thinking about?”
Cette publication est également disponible en :
