Beijing Pushes Open-Weight AI Models to Challenge US Dominance
China is betting that the global adoption of its adaptable artificial intelligence models will reinforce its industrial dominance and extend its international reach. US President Donald Trump stated in September that whoever wins AI wins the global race, declaring that the US leads in frontier models. Meanwhile, Beijing is distributing open-weight models that users can download and modify, offering a cheaper and more flexible alternative to proprietary Western systems.
The strategy relies on driving international governance and multilateral cooperation. At July’s UN Global Dialogue on AI Governance, participants stressed that governance should reflect the priorities of all nations, not just the most technologically advanced. Two weeks later at the World AI Conference & High-Level Meeting on Global AI Governance in Shanghai, Xi Jinping urged governments to adhere to openness and prevent historical injustices in technology development.
To institutionalize this approach, China launched the World AI Cooperation Organization in July. Twenty-nine countries signed the founding charter of the Shanghai-based organization, including Brazil, Indonesia, South Africa, and Russia. The body aims to promote a non-discriminatory environment for AI development.
Washington Responds With Closed Models and Trusted Networks
While Chinese competitors distribute open-weight models with published parameters, leading US AI companies keep their closed models under wraps. Users can access US systems only through restricted interfaces. US policymakers view closed models as essential for maintaining control over frontier technology and securing potential military and industrial advantages.
To protect supply chains and counter Chinese dissemination, the US launched the Pax Silica initiative in late 2025 to reduce coercive dependencies and protect critical materials. The initiative includes the EU and India among more than 20 trusted partners. The American AI Exports Program seeks to ensure worldwide adoption of US technology standards and governance models.
Despite US export controls, Chinese developers have narrowed the capability gap. Recent near-frontier open-weight releases like Xiaomi’s -V2.6 and Moonshot’s Kimi K3 demonstrate that domestic developers are only months behind their US peers.
Security Risks and State Control in Chinese Open-Weight Systems
China’s ready access to capable models without US provider dependence presents an attractive option for governments and industries. However, openness does not equal freedom from state control. Wendy Chang and Olaf Schulz warn that all Chinese AI services require approval from the Cyberspace Administration of China, posing risks of propaganda and manipulation.
Unlike open-source software, open-weight models do not expose their underlying code and training data, leaving limited room for external security scrutiny. Customizing these models creates deep integration into public administration and logistics IT systems, making switching providers costly and difficult.
Analysts note that prepositioned Chinese AI models embedded in critical infrastructure could create new exploitation opportunities. August revelations concerning an AI attack on the developer platform Hugging Face challenged Silicon Valley’s orthodoxy on closed-source safety, though questions remain regarding whether Chinese firms can freely disclose incidents involving domestic agents.
Frequently Asked Questions About US-China AI Competition
Which countries signed the World AI Cooperation Organization charter?
Twenty-nine countries signed the founding charter in Shanghai in July, including Brazil, Indonesia, South Africa, and Russia.
What is the primary difference between US and Chinese AI deployment strategies?
US companies use closed models accessible only through controlled interfaces to maintain proprietary advantage, while Chinese competitors distribute open-weight models that users can download and modify.
What security concerns do open-weight models present?
Open-weight models do not reveal their underlying code and training data, limiting scrutiny while risking deep integration into critical IT systems that are difficult to dismantle.
Related reading