“No country in the world right now controls every single layer,” cautioned Eva-Maria Hempe, executive director for Public Sector Emea at Nvidia, during a keynote on the five-layer AI sovereignty stack at Nexus Luxembourg 2026. Photo: Paperjam

“No country in the world right now controls every single layer,” cautioned Eva-Maria Hempe, executive director for Public Sector Emea at Nvidia, during a keynote on the five-layer AI sovereignty stack at Nexus Luxembourg 2026. Photo: Paperjam

Nvidia’s Eva-Maria Hempe argued that AI sovereignty is not about controlling everything, but about understanding dependencies across energy, chips, infrastructure, models, data and applications.

Eva-Maria Hempe, executive director for Public Sector Emea at Nvidia, warned that AI sovereignty should not be treated as a single concept, arguing that no country currently controls every layer of the AI stack. Speaking during a keynote at Nexus Luxembourg 2026, Hempe described AI sovereignty as “not a monolithic concept” but a “multi-layer concept”.

“What it does not mean is that you have to control every single layer, because I would argue no country in the world right now controls every single layer,” Hempe stated. She added that the issue was “really more about being aware of the different layers, how they interact, and then making very strategic and very informed choices about potential dependencies and how you manage them.”

Sovereignty as dependency management

Hempe described the global AI infrastructure build-out as “the largest infrastructure build out in human history”, with data centre projects “springing up everywhere”. She linked the investment to three factors: economic growth, security and sovereignty. Compute capacity, she stated, would be “one of the foundations for future economic growth”.

On security, Hempe cautioned that foreign AI infrastructure had become a strategic issue as geopolitical tensions increased.

“We are getting much more in the age of increasing geopolitical tensions,” she noted, adding that there was “a lot more nuance about what dependencies there are, and to which degree you can or cannot rely on foreign AI infrastructure as a category of strategic risk.” For sovereignty, she argued that the issue was control over technologies that shape public life.

“If you think about a simple LLM, an LLM carries values inside it,” Hempe stated. “The data it has been trained on, the way it has been trained, embeds certain values inside that model.”

She warned that this mattered when large language models were used in schools or by public servants to support decisions. “These values are becoming part of the decision-making process, and hence it is important for nations to ensure that their values are embedded in the technology they’re using,” she remarked.

The five-layer stack

Hempe presented AI sovereignty as a five-layer stack starting with energy, followed by chips, infrastructure, models and data, and applications. “It all starts with energy,” she stated, while adding that even countries with strong energy systems had dependencies. “Nobody is truly sovereign. France has great nuclear power stations, but France doesn’t have that much uranium to run them on.”

The same applied to chips, she observed. Nvidia’s chips are designed in Silicon Valley, produced in Taiwan by TSMC, use machines from Dutch company ASML and rely on mirrors made in Germany.

“We have a per se global supply chain also in this level of stack,” Hempe noted.

The infrastructure layer included “land, power delivery, cooling, networking, but basically data centres”, while the upper layers covered “the models and the data that feeds the models” and the applications built on top.

Europe’s compute gap

Hempe pointed to Europe’s limited share of AI computing power as a central sovereignty challenge. She stated that US hyperscalers accounted for 56% of global AI computing power, while other US players accounted for 40%. Europe, by contrast, had 4% of AI computing power in last year’s figures, rising to around 5%. “That’s not a lot,” Hempe remarked. “AI is going to be so essential for economic growth, for security, and for sovereignty.”

She added that more than 90% of frontier AI model capabilities originated from US companies, while Europe spent more than $12bn per year on US hyperscale infrastructure.

“In general, there is growing demand in Europe,” she observed. “There’s so much talent in Europe, there is money in Europe, so we really have an opportunity in our hands to have a sovereign AI future for Europe across the different layers of the stack.”

Every choice has a trade-off

Hempe cautioned that the infrastructure layer gave organisations several choices, but none represented full sovereignty. One option was to use hyperscalers and consume AI “off the tap”. Another was to use hyperscaler data centres physically located in Europe. A third option was to use European neo-cloud providers with “often enough higher sovereignty layer”. At the far end was the “fully air-gapped on-prem data centre”.

“All of these choices are perfectly legit,” Hempe stated. “You just have to make an informed decision.”

She argued that organisations had to decide “what are you trying to do, what is important to you, what are the trade-offs between cost, between performance, between risk, and what’s also your assessment of risk.”

Hempe said similar trade-offs applied to the model layer. Closed frontier models could be consumed through an API, while open-weight models could be downloaded and run locally. “Open weights is great, it gives you a lot of great options,” she acknowledged, adding that such models could be run in “your air-gapped super secure defence bunker” so that data “is never going to leave the premises”. But she cautioned that open weights did not remove all uncertainty.

“The thing about open weights, it’s still a black box,” Hempe warned. She said the alternative was to take openness further through open data.

“Now your black box becomes a transparent box, something you know what went into it,” Hempe explained. “It’s still not fully deterministic, but if you know what data was used to train your models, you have a lot more insight into how this model will operate when you’re actually using it.”

For schools, for example, she argued, transparency over training data would be especially important. “If I want to use a model in schools to use in lectures about history, I would very much like to know what version of history went into training this model,” Hempe remarked.

Building blocks, not full control

Hempe concluded that AI sovereignty required national compute infrastructure, open models with open weights and open training data for certain applications, and deployment and governance capabilities. “For certain applications, you have transparency requirements,” she stated. “You have to be able to disclose why decisions are made the way they are.”

But she cautioned that infrastructure alone is insufficient.

“It’s not enough to just put some GPUs on a field somewhere,” Hempe warned. “You need to know how to use them, you need to build the capabilities, need to build the knowledge, need to build knowledge also on the regulatory side.”

Her message was that AI sovereignty is not absolute control, but a set of informed choices across a stack where every country and organisation faces dependency. “You just need to know what you’re trying to do and what you’re willing to give,” Hempe concluded.