[Today’s Signal] The U.S. Blocked GPUs to China. Now It Says AI Knowledge Is Moving Through APIs

The U.S. government’s focus on restricting physical GPU exports to China has been sidestepped by API-based access to American frontier AI models. This development significantly alters the landscape of technology control, as the challenge shifts from managing hardware supply chains to protecting intellectual property embedded in remotely accessible software. The alleged knowledge distillation by Chinese firms like Alibaba and Moonshot AI, involving billions of tokens, points to a sophisticated and distributed approach to acquiring AI capabilities without needing advanced computing infrastructure on their own soil. This poses a direct challenge to the effectiveness of current U.S. export controls. For Asian AI developers and cloud providers, this situation underscores the growing scrutiny on how AI models are accessed and utilized across borders. Companies operating in the region that offer API access to advanced models or provide cloud infrastructure could face increased regulatory pressure from the U.S. The incident with Alibaba’s alleged activity against Anthropic’s Claude model, involving over 151 million exchanges, highlights the scale of the problem and suggests that model providers will need to implement more robust access controls and monitoring mechanisms to prevent unauthorized knowledge transfer. The thing to watch is how U.S. policy adapts to control software-based access, which is inherently more complex to regulate than physical goods.