The US-China rivalry in artificial intelligence is intensifying, as Chinese companies demonstrate a rapidly advancing AI ecosystem. Moonshot AI's Kimi K3 model, released as open-source, has notably surprised Wall Street and Silicon Valley by performing at or near the level of leading US models on several benchmarks. This development, along with advancements from Alibaba's Qwen family, Tencent's Hunyuan, and Zhipu AI, suggests China has cultivated an AI environment capable of repeatedly producing world-class capabilities across multiple firms, a strategic reality that the US often evaluates on a company-by-company basis.
Despite Washington's efforts to restrict China's access to advanced Nvidia chips, Chinese firms are finding ways to train powerful AI models. For instance, Moonshot reportedly secured a computing agreement with Alibaba for approximately 20,000 Nvidia Hopper-generation chips, specifically H200s, to power its Kimi K3 model. While the US has barred the sale of its most powerful chips to China and now requires licenses even for less capable models like the H200, some Chinese cloud providers are exploiting a workaround by renting computing power from overseas data centers that house restricted processors. This access challenges the effectiveness of US export controls, as evidenced by Kimi K3's strong performance.
The competitive landscape is further complicated by accusations of intellectual property theft and distillation. Alibaba's Qwen lab has been accused by Anthropic of running an extensive model distillation campaign against Claude, involving 28.8 million unauthorized exchanges through 25,000 fake accounts. Moonshot, a company backed by Alibaba and utilizing its infrastructure, now faces similar accusations regarding the distillation of Anthropic's Fable 5 to develop Kimi K3. While the timeline for such distillation for Kimi K3 has been disputed by researchers, the US Treasury Secretary has warned of potential sanctions and Entity List designations for PRC firms engaging in such activities, highlighting the legal and ethical challenges emerging from this intense AI competition. This dynamic illustrates how the company that invests in and supplies compute to a rival can then be outperformed by its own protégé, challenging the existing export control framework.
