LISBON, Portugal — Tech giants are increasingly investing in the development of so-called “sovereign” artificial intelligence models as they seek to strengthen their competitiveness by focusing more on local infrastructure.
Data sovereignty refers to the idea that people’s data should be stored on the infrastructure of the country or continent in which they reside.
“Sovereign AI is a relatively new term that has emerged over the last year,” Chris Gow, head of European public policy at Brussels-based computer networking giant Cisco, told CNBC.
Currently, many of the largest language models (LLMs), like OpenAI’s ChatGPT and Anthropic’s Claude, use U.S.-based data centers to store data and process requests via the cloud.
This has sparked concern among European politicians and regulators, who see reliance on US technology as detrimental to the continent’s competitiveness – and, more worryingly, technological resilience.
Where does “AI sovereignty” come from?
The notion of data and technological sovereignty is an issue that was previously on the European agenda. This is partly due to companies’ reaction to new regulations.
The European Union’s General Data Protection Regulation, for example, requires companies to process user data securely and compliantly, while respecting their right to privacy. High-profile cases in the EU have also raised doubts about whether European citizens’ data can be securely transferred across borders.
In 2020, the European Court of Justice invalidated a data sharing framework between the EU and the United States, on the grounds that the agreement did not offer the same level of protection as that guaranteed within the EU by the General Data Protection Regulation (GDPR). Last year, the EU-US Data Privacy Framework was created to ensure that data can flow securely between the EU and the United States
These policy developments ultimately resulted in a push toward localization of cloud infrastructure, where data is stored and processed for many online services.
Filippo Sanesi, global head of startup marketing and operations at OVHCloud, said the French cloud computing company is seeing strong demand for its infrastructure located in Europe because it “understands the value of having its data in Europe, which are subject to European legislation. “
“As this concept of data sovereignty becomes more mature and people understand what it means, we are seeing more and more companies understand the importance of having your data locally and under jurisdiction and governance specific,” Sanesi told CNBC. “We host a lot of data,” he added. “This data is sovereign in specific countries, within specific regulations.”
“Now, with this data, you can actually create products and services for AI, and those services should then be sovereign, controlled, deployed and developed locally by local talent for the local population or businesses.”
The push toward AI sovereignty hasn’t been pushed forward by regulators — at least not yet, according to Cisco’s Gow. Instead, it’s coming from private companies, which are opening more data centers — facilities containing large amounts of computing equipment to enable cloud-based AI tools — in Europe, he said.
Sovereign AI is “more driven by the industry that nominates it as well as policymakers,” Gow said. “You don’t see the terminology of ‘AI sovereignty’ used on the regulator side yet.”
Countries are championing the idea of AI sovereignty because they recognize that AI is “the future” and an “extremely strategic technology,” Gow said.
Governments are focused on strengthening their domestic technology businesses and ecosystems, as well as the all-important back-end infrastructure that enables AI services.
“The AI workload uses 20 times the bandwidth of a traditional workload,” Gow said. According to Gow, it’s also about empowering the workforce because businesses need skilled workers to succeed.
But the most important thing is the data. “What you see is a lot of attempts on that side to think about training LLMs on localized data, in language,” Gow said.
“Reflecting values”
In Italy, the first LLM trained specifically on Italian language datacalled Italia 9B, launched this summer.
The goal of the Italia project is to store results in a given jurisdiction and draw on data from citizens of that region so that the results produced by AI systems are more anchored in languages, culture and l local history.
“Sovereign AI is about reflecting the values of an organization or, also, the country you are in, as well as the values and the language,” David Hogan, head of EMEA enterprise sales for the tech giant chip manufacturing. Nvidiatold CNBC.
“The main challenge is that most of today’s frontier models have been trained primarily on Western data in general,” Hogan added.
In Denmark, for example, where Nvidia has a large presence, officials are concerned that vital services such as healthcare and telecommunications are provided by AI systems that do not “reflect” culture and values local Danish companies, according to Hogan.
On Wednesday, Denmark released a landmark white paper outlining how businesses can use AI under the new EU AI law, the world’s first major law on AI. The document is intended to serve as a model for other EU countries to follow and adopt.
“If you’re in a European country that’s not one of the major internationally spoken linguistic countries, probably less than 2 percent of the data is about your language, let alone your culture,” Hogan said.
How regulation fueled a mindset shift
That’s not to say that regulation hasn’t proven to be an important factor in getting tech giants to think more about building localized AI infrastructure in Europe.
OVHCloud’s Sanesi said regulations such as the EU GDPR have catalyzed much of the interest in offshoring data processing within a given region.
The concept of AI sovereignty is also attracting support from local European technology companies.
Earlier this week, Berlin-based search engine Ecosia and its Paris-based counterpart Qwant announced a joint venture to develop a European search index from scratchaimed at serving better results in French and German.
Meanwhile, the French telecom operator Orange said it is in discussions with a number of AI model founding companies about creating a smartphone-based “sovereign AI” model for its customers that more accurately reflects their own language and own culture.
“It wouldn’t make sense to create our own LLMs. So there’s a lot of discussion at the moment about how do we collaborate with existing providers to make it more local and safer?” Bruno Zerbib, Orange’s chief technology officer, told CNBC.
“There are many use cases where (AI data) can be processed locally (on a phone) instead of being processed on the cloud,” Zerbib added. Orange has not yet selected a partner for these sovereign AI model ambitions.