US export controls on AI chips have been tightened to prevent Chinese companies from remotely accessing advanced processors located outside China. The move addresses a loophole highlighted by the Nvidia-Megaspeed saga, but it does not resolve a wider contradiction in Washington’s strategy for securing American dominance in artificial intelligence.
The United States is encouraging the global spread of data centres equipped with American-designed chips, including in Latin America, the Middle East and south-east Asia. Yet the same facilities can also provide the computing power needed to deliver Chinese artificial intelligence models to users around the world.
US infrastructure is powering Chinese models
US-owned data centres using chips from companies such as Nvidia continue to rent capacity to inference providers. These firms run models on behalf of customers, meaning Chinese large language models can be served globally using American hardware and cloud infrastructure.
That arrangement reflects the strength of the US technology sector. Nvidia, Google, AMD and Amazon design nearly all deployed data-centre AI chips, while Amazon, Microsoft and Google together control more than half of the cloud-services market. A growing group of US “neoclouds” has further expanded America’s position.
China is increasingly using domestic chips within its own borders, but it cannot produce enough advanced processors to meet the needs of its companies, let alone overseas demand. Chinese model developers therefore remain able to benefit from infrastructure controlled by American businesses.
Nvidia provides resources to help companies deploy open models, including several developed in China. US-based providers such as Together AI, Baseten and Fireworks also serve Chinese models on systems powered by US chips.
DeepSeek’s latest model, for example, is available through at least 16 inference providers headquartered outside China, including Amazon Bedrock. Wider access can help Chinese models gain users, generate revenue for their developers and strengthen China’s influence, particularly in developing markets.
Moonshot AI has reportedly sought revenue-sharing agreements with companies serving its Kimi K3 model. Under the reported terms, businesses generating more than 20 million dollars in annual sales would have to negotiate a licence, with Microsoft, Amazon and Google allegedly asked for 30 per cent of their Kimi K3 revenue.
The open-weight model challenge
The impact of American computing infrastructure depends partly on the type of model it is running. Capacity used for proprietary systems such as Anthropic’s Claude or OpenAI’s GPT largely benefits US developers, which currently produce the strongest models in that category.
Open-weight models present a different challenge. They can be adapted for specialist uses and operated locally, and Chinese developers are currently highly competitive in that market. Infrastructure used to distribute those models can therefore help their creators expand internationally.
The Trump administration has been seeking US leadership across the technology chain, from chips and cloud services to AI models. But exporting American infrastructure will not automatically result in the global adoption of American-developed models.
A blanket ban on US cloud providers allowing non-Chinese companies to serve Chinese models would carry significant costs, potentially restricting businesses in the United States and allied countries while bringing limited national-security benefits. It could also prompt retaliation from Beijing.
Instead, the US should seek to outperform its competitors, including by supporting American open-weight models. Meta and Thinking Machines are among the companies developing such systems, and greater efficiency could encourage more of the world’s computing capacity to be used for US models.
Federal agencies could identify tasks suited to open-weight systems and direct procurement towards American models. Any future AI regulation would also need to avoid making it harder for US developers to build and distribute them.
The new export rule closes an important access route to advanced chips, but the broader competition will also be decided by which models people and businesses choose to use. American control of the hardware will not by itself prevent Chinese systems from gaining a global foothold.
