AI sovereignty: for companies, the power to choose is what matters

For a company, AI sovereignty means knowing where its data lives and being able to switch language models without rebuilding the project. Owning every piece of the technology is not realistic. It isn’t for most countries, as the World Economic Forum writes in its white paper “Rethinking AI Sovereignty”, published in January 2026 with Bain & Company. Even less so for a single business.

Why are even countries giving up on self-sufficiency?

The United States and China attract more than 60% of global AI investment and are the only ones present along the entire value chain, from chips to applications. Everyone else has to choose.

The WEF groups them into five archetypes. Italy isn’t named, but the profile that fits it best is that of the “selective players”, alongside Germany and France: industrial economies that invest in data and applications for manufacturing and rely on foreign partners for hardware and compute. The recommended path is to focus where you have an advantage and to treat partnerships as part of sovereignty, not as a fallback.

Where is AI value shifting?

To applications. According to the WEF, by 2030 annual investment in AI applications and services will reach around $1.5 trillion, more than infrastructure and models combined. With PCs, value moved from hardware to software; with smartphones, from devices to apps. For a manufacturing company, the game is played in how AI enters its processes.

What changes when this reasoning reaches the company?

Data and processes stay within the company perimeter. The language model, on the other hand, is a choice, and it will change, because new ones come out every few months with different prices and performance. The risk is getting tied to a single vendor with no way out, with data leaving the company and costs rising with usage.

“Which data, specifically?”

Riccardo Celli, CEO of Aidia, hears it in client requests: “The main concern is intellectual property: the code that goes in, patents, technical specifications. And then customers: the fear of losing them, of someone taking them away or seeing how much they spend.”

The WEF lists this among its five priorities: avoiding vendor lock-in requires interoperability and portability, with room for both proprietary and open-source models. Operational control matters more than ownership.

Why does every choice weigh more in Europe?

In an analysis published on 6 October, ECB economists observed that positive AI news boosts investment in both the United States and the euro area, but more so across the Atlantic. Their suggested explanation: in the US, venture capital and equity markets are deeper and the market is unified. In Europe, capital for AI is harder to find. A project that has to be rebuilt after two years, because it was designed around the wrong vendor, costs more here than elsewhere.

And the cost isn’t only financial. “The most expensive choice is getting used to an improper use of AI today,” says Celli. “Companies seriously risk ending up in a stranglehold with current providers. Changing workflows is costly, but above all it takes far too much time.”

Three questions to ask before committing

Where is data processed, and who can access it?
How much does it cost to switch to another model?
Do compute costs stay predictable as usage grows?

AVA, Aidia’s platform that connects enterprise systems to any LLM, proprietary or open source, is built to answer these three questions. It runs on-premise, lets you switch models without redesigning the system and keeps compute costs fixed. Data never leaves the company.

If you’re evaluating an AI project and want to understand which parts to keep under your control, let’s talk.