China is aggressively pushing to develop its own AI chip industry and overcome export restrictions by the US. This initiative has transformed China's semiconductor localization drive from a long-term goal into an urgent national security mandate. Domestic investments in chip design and manufacturing, particularly by companies like SMIC, are accelerating, even when immediate commercial returns are not guaranteed. The objective is to eliminate strategic vulnerabilities stemming from reliance on foreign technology.
Despite these efforts, China faces significant challenges, particularly in advanced lithography. The country's access to cutting-edge EUV lithography equipment is severely restricted by US export controls, forcing Chinese manufacturers to rely on older DUV-based systems. This limitation impacts the yield and density of advanced logic nodes, making them less efficient and more costly compared to those produced with EUV technology. For instance, SMIC's advanced processes are comparable to TSMC's N7 in cost but suffer from materially lower yields and density.
The White House has accused Chinese AI company Moonshot AI of illicitly accessing advanced Nvidia chips, specifically the GB300, in Thailand to train its AI models. This accusation comes amid heightened competition between the US and China for AI supremacy. Moonshot AI recently unveiled its Kimi K3 model, which reportedly closes the performance gap with leading frontier models like Anthropic's Fable 5 and OpenAI's GPT 5.6 Sol, bahkan surpassing them on some benchmarks. US officials are also considering legislation, like the Remote Access Security Act (RASA), to expand export controls to include remote cloud-based access to critical hardware and software, aiming to close loopholes that allow Chinese companies to use US compute power both domestically and abroad.
While China has made progress in some semiconductor equipment categories like deposition, CMP, and wafer cleaning, it still lags significantly in critical areas like advanced memory. Nvidia's H200 chips, for example, rely on HBM3e memory from suppliers like SK Hynix and Micron, which are subject to export controls, preventing China from replicating such bandwidth domestically. This disparity in memory technology has direct consequences for large-model training and inference. The ongoing export controls are accelerating a bifurcation into two distinct global AI chip markets, each with different cost, performance, and supply-chain dynamics.
