China vs G7: AI at the Heart of a New Global Battle

China vs G7: AI at the Heart of a New Global Battle

The G7 concluded in Évian by limiting access to American models to allies only. Meanwhile, Wang Yi announced from Beijing the creation of a global cooperation organization in AI open to all. The governance fracture between these two blocs has never been more evident.

Since June 12, when the United States ordered Anthropic to suspend global access to its two most advanced models for national security reasons, the G7 summit in Évian-les-Bains has taken place in an atmosphere of geopolitical tension regarding AI. On the sidelines of this meeting, Dario Amodei (Anthropic), Demis Hassabis (Google DeepMind), and Sam Altman (OpenAI) advocated for a Washington-led coalition, restricting access to cutting-edge models to a circle of "trusted partners." On Wednesday, as leaders debated without reaching an agreement, Wang Yi spoke up to announce an opposing approach.

Wang Yi Plays the Openness Card

The Chinese Foreign Minister stated that Beijing is "accelerating the establishment of a global cooperation organization on AI" and invites all countries to participate. This announcement was accompanied by the release of a white paper on global governance, which directly targeted Washington without naming it. Zhao Haibing, vice president of China's main economic agency, explicitly criticized "closed, exclusive, and monopolistic approaches to technological development." To support this initiative, Beijing indicated it is relying on the BRICS and the Shanghai Cooperation Organization, multilateral structures where the United States has no voice.

This maneuver fits into a carefully crafted strategy since July 2025, when Li Qiang announced this global organization in Shanghai, just days after Washington released its own action plan for AI. Since then, Beijing has built its commercial argument: its open-source models (DeepSeek, Qwen) account for 17.1% of global downloads on Hugging Face (the most used platform for distributing open models), surpassing American models (15.86%) for the first time since January 2025, according to a MIT-Hugging Face study. These models are available for free, run on modest hardware, and require no subscription or data localization conditions. In Africa, Southeast Asia, and Latin America, they fill a gap that American export restrictions have widened over the past two years, supported by access to computing power independent of NVIDIA and other American industry giants.

Europe Caught Between Two Incompatible Architectures

The European Union was not merely an observer in Évian. Emmanuel Macron warned his counterparts that while Washington could "flip the switch overnight," American AI companies would also suffer. Narendra Modi expressed similar concerns. No agreement was reached. The status of "trusted partner" would remain defined by Washington, according to Reuters, and could be unilaterally revoked by the Department of Commerce.

Mistral AI illustrates this ambivalence on the Old Continent. The French unicorn had built its identity on open efficiency before its most ambitious business model went through Microsoft's Azure. On Hugging Face, Qwen3-72B recorded 640,000 downloads in the first week of April, compared to 380,000 for Codestral. Between the American ecosystem it commercially depends on and the Chinese open-source dynamic encroaching on its territory, Mistral maintains an increasingly precarious balance.

Other European players are pursuing their own paths. Multiverse Computing, a Basque startup founded in Donostia, aims for a valuation of over 1.5 billion euros by compressing high-performing open-source models using quantum-inspired mathematics to deploy them on consumer hardware, with Iberdrola, Bosch, and the Bank of Canada among its declared clients. The approach is coherent, and the niche is real. But as Arthur Mensch, CEO of Mistral, pointed out at the Évian summit: the paramount question for non-American countries is control over their AI supply chain.

The good news for Europe: two options are available, one free and the other paid. The less good news: neither one awaits its opinion to exist.