China hits milestone in AI and semiconductor self-sufficiency ahead of U.S. visit

China’s AI stack is getting less American. That is the real story here. The news is not that China suddenly “caught up” everywhere — it hasn’t. The news is that one of the biggest weak points in China’s AI system, dependence on Nvidia and other foreign suppliers, is starting to look less like a hard stop and more like an engineering problem China can route around. (csis.org) The immediate trigger is DeepSeek and Huawei. DeepSeek’s latest V4 model was built to run on Huawei’s Ascend chips, and that has kicked off a scramble by Chinese tech groups to secure more of those processors. Reuters reported in late April that demand for Huawei’s Ascend 950 jumped after the model launch, with companies like Alibaba, Tencent, and ByteDance rushing to line up supply. That matters because it is not just a lab demo — it is a sign that a domestic AI software company and a domestic chip company can now reinforce each other at scale. (money.usnews.com) Why is that a milestone? Because the old U.S. theory of pressure was pretty simple: if China could be denied the best training chips, the rest of the AI stack would slow down with them. That still works at the absolute frontier. Nvidia remains stronger, especially for top-end training. But China is no longer stuck waiting for perfect substitutes. It is building an “alternative stack” — chips, software tools, cloud infrastructure, and customers willing to optimize around domestic hardware even when it is less elegant or less efficient. (businesstimes.com.sg) That is the piece people miss. AI is not one magic chip. It is a whole chain. You need semiconductors, memory, networking, model software, developer tools, data centers, and enough users to justify the cost of tuning everything together. China has been trying to localize that chain for years, well before the current export-control era. CSIS notes that Beijing’s push to replace foreign semiconductor inputs was already embedded in policy documents like Made in China 2025, with explicit localization targets for equipment and chipmaking. The controls did not create that ambition — they accelerated it. (csis.org) So what changed in 2026? The center of gravity moved from “can China build domestic pieces?” to “can those pieces work together well enough to matter?” DeepSeek’s decision to spend months optimizing V4 for Huawei chips is a good example. Bloomberg described the delay as a strategic shift toward deeper integration with China’s domestic chip ecosystem, not just basic compatibility. Basically, China is moving from substitute parts to a more coherent platform. (bloomberg.com) The catch is that self-sufficiency is still uneven. China remains behind at the true leading edge of semiconductor manufacturing equipment, especially the most advanced lithography. Even supportive analyses say the domestic stack is often less advanced than foreign alternatives, and some Chinese AI progress has still relied on smuggled or previously acquired Nvidia hardware. So this is not autonomy in the pure sense. It is partial insulation. (csis.org) But partial insulation is enough to change the politics. If Chinese firms can keep releasing competitive models, keep buying domestic chips, and keep building data centers around them, then U.S. export controls lose some leverage. They still raise costs. They still slow frontier progress. But they no longer guarantee strategic dependence. That is why this matters ahead of a U.S.-China summit — Beijing can walk in arguing that the pressure campaign hurt, but did not freeze its AI future. (csis.org)