European firms spread AI bets after US access curbs

TL;DR:

• Major European firms are mixing US, Chinese and European AI models to avoid depending on any single provider.
• The US order forcing Anthropic to suspend access for foreign nationals exposed the fragility of remote proprietary services.
• Rising “cost per token” is emerging as a second pressure point as companies deploy autonomous agents.

Limits on access to some US artificial intelligence services are pushing major European companies to spread risk across multiple providers and to push harder for domestic alternatives. The trigger is familiar to anyone following the story: Washington’s order for Anthropic to suspend access to its Fable 5 and Mythos 5 models for foreign nationals on national-security grounds. That move laid bare a structural weakness — proprietary services delivered remotely can be switched off by their providers and cannot be run on a company’s own servers.

Diversity, not autarky

Executives from Siemens, Renault, Orange and ChapsVision told Reuters at VivaTech in Paris that they already blend US, Chinese and European models. Siemens uses China’s DeepSeek and Alibaba’s Qwen alongside Nvidia’s Nemotron and other systems. “Sovereignty often gets confused with autarky, and autarky is absolutely not the way to do it,” said Siemens digital industries chief Cedrik Neike. Orange put the Anthropic episode bluntly: it made “patently clear” how important it is for Europe to have AI it can control. The picture echoes recent Resultsense coverage of Europe confronting its US AI dependence and France dropping Palantir for domestic options.

A second theme is cost. Token charges are rising as firms shift to agentic systems, with Orange predicting executives will be “obsessed with cost per token” by year-end and citing Uber burning through its 2026 token budget in four months. Celonis warned that deploying agents without a “context model” of how a business works will “blow your token bill completely”.

Looking forward

For UK organisations, the lesson is twofold. First, resilience now means multi-model architectures rather than betting on one frontier provider — a posture that demands engineering discipline most firms have not yet built. Second, the economics of agentic AI are tightening just as adoption accelerates, making cost governance as important as model choice. Britain’s own sovereign-AI debate runs along the same fault line: whether open-weight models, run on local infrastructure, offer a credible hedge against services that can be cut off “on a whim”.