US curbs on AI spur European firms to spread the risk
European Companies Respond to US AI Restrictions
(Fixes name to Kuhn (previously Kahn) in paragraph 22)
By Leo Marchandon
PARIS, June 22 (Reuters) - Limits on access to some U.S. artificial intelligence services are pushing major European companies to speed up spreading risk across multiple providers and reinforcing the need for more domestic alternatives.
Recent restrictions include the U.S. government ordering San Francisco-based Anthropic, the company behind AI chatbot Claude, to suspend access to its Fable 5 and Mythos 5 models for foreign nationals, citing national security concerns.
Such limits highlight a vulnerability for companies reliant on proprietary AI services delivered remotely - as those services can be restricted by their providers and cannot be run independently on a company's own servers.
Adopting a Multi-Provider Approach
Executives from Siemens, Renault Group, Orange and ChapsVision told Reuters at last week's VivaTech conference in Paris that they already use a mix of U.S., Chinese and European models to avoid dependence on any one provider.
Siemens said it uses Chinese models DeepSeek and Alibaba's Qwen, as well as Nvidia's Nemotron alongside other U.S. and European models.
EU Efforts Toward Digital Sovereignty
EU officials have been trying to reduce dependence on U.S. technology, which they see as a threat to the region's economic future, and have compiled a sovereignty package of measures to strengthen the bloc's capacity in semiconductors, AI and its digital autonomy.
Yet, major companies say sovereignty is more about choice and diversity.
"You need flexibility," Cedrik Neike, chief executive of Digital Industries at Siemens, told Reuters. "Sovereignty often gets confused with autarky (economic self-sufficiency), and autarky is absolutely not the way to do it."
Limited European AI Providers
Europe's general-purpose AI providers, however, are few, headed by France's Mistral, while others like translation specialist DeepL have built strong positions in narrower fields.
The artificial intelligence market is broadly split into two camps: open-source or open-weight models, which companies can run on their own servers, and proprietary models, which are accessed remotely and remain under the developer's control.
"Today, in open source, when you look at European models, they're not impressive. At one point, the Americans were there, then they moved to closed source, and now there are only Chinese models in open source," OVHcloud Chief Executive Octave Klaba said.
Cost and Operational Pressures
Infrastructure and Control
Orange said its infrastructure could run all open-source models, including those developed in China, and described the risk in simple terms: running a Chinese model on European infrastructure is like buying a painting in China and bringing it home, because the model is self-operated and does not send data abroad when run on premises.
The Anthropic restrictions, Orange said, made "patently clear, if it wasn't before, how important it is for Europe to have access to an AI service that it can control, that will never be switched off on a whim."
Orange's Chief Executive Christel Heydemann, in a keynote address, called on Europe to build artificial intelligence it can access, govern and challenge on its own terms.
French AI and data analytics company ChapsVision, which has won government contracts in France and Germany to replace U.S. rival Palantir, said it uses models ranging from Mistral and Anthropic to OpenAI and Qwen. For ChapsVision, sovereignty means having a credible alternative ready if a key service is cut.
Software companies SAP and Sopra Steria agreed resilience comes through diversification, not isolation. IT group Capgemini said most AI providers were adapting their offerings beyond remote access to ease dependency concerns in Europe because the market was too large to lose, although it flagged it was a work in progress.
Rising Token Costs
Cost is also emerging as a pressure point.
Token costs - the fees charged per unit of information processed - are rising as companies move to systems in which software agents carry out tasks automatically.
Orange said executives would be "obsessed with cost per token" by the end of this year, citing Uber as an example of a company that had burned through its 2026 token budget in just four months.
Carmaker Renault Group works with Google, Microsoft, Mistral, DeepSeek and Dataiku, using both open-weight and proprietary models, although it says it is not using DeepSeek in a meaningful way yet.
"Renault Group already has an in-depth reflection on the cost of AI tokens, which have risen sharply and are pushing us to adapt," a spokesperson said.
Building Effective AI Infrastructure
Rudy Kuhn, a senior executive at German software firm Celonis, whose clients include BMW and Siemens, said companies need to first build the necessary infrastructure to give AI agents context on how their business works before deploying them.
"If you do not provide a context model, AI needs to extract every single fact from the data itself," he said. "This will just blow your token bill completely."
(Reporting by Leo Marchandon. Additional reporting by Dominique Patton and Gilles Guillaume; Editing by Matt Scuffham and Susan Fenton)
