Why does. The U.S. bar AI chip designers like Nvidia and AMD from selling their most powerful chips to China? Does China need these chips and why? Do they have alternatives?

Why the ban exists

An AI chip, in simple terms, is a processor built specifically to run the massive calculations that training and operating AI models requires. Think of it as the engine under the hood of every AI system, from a chatbot to a weapons-targeting algorithm. The most powerful of these, Nvidia's H100 and its successors, can do in hours what would take ordinary computers weeks.

The US government prevents Nvidia from selling some of its more advanced AI chips to China in order to prevent the Chinese military and surveillance agencies from accessing cutting-edge processing power. The logic is straightforward: whoever controls the most powerful chips controls the pace of AI development, and AI increasingly underpins everything from battlefield decision-making to facial recognition networks.

The restrictions, first announced in August 2022, targeted China's semiconductor and supercomputing industries and specifically impact Nvidia's A100 and H100 integrated circuits, as well as any systems incorporating them. The export controls are complex and address a variety of parameters, including the total processing performance of a chip, its "performance density," its interconnect bandwidth, and its memory bandwidth. In other words, it is not just one chip that is banned: any future chip that hits certain performance thresholds is automatically controlled.

These restrictions were initially imposed by President Joe Biden and were then expanded by President Donald Trump in 2025, though Trump has since reversed some of them. As of May 2026, Nvidia can sell its H200 chip and AMD can sell its MI308 chip to China, but those are not the companies' most powerful products, and the deal is tangled. The US Department of Commerce cleared approximately ten Chinese firms, including Alibaba, Tencent and ByteDance, to purchase Nvidia's H200 chip with a limit of 75,000 units per customer, but no chip deliveries have been made, as the deal remains in legal limbo due to the ongoing US-China tech rivalry and Beijing's new supply chain regulations.

Chinese AI firms have reportedly accessed advanced Nvidia computing power overseas, and lawmakers are now weighing closing that cloud-access gap. Cloud access means a Chinese company can rent time on a powerful chip sitting in a data center in Singapore or the Netherlands, without ever importing the chip itself. That is the loophole the controls are still chasing.

Does China need these chips?

Absolutely, and the evidence is visible in what Chinese companies did the moment access was cut off. The launch of China's DeepSeek model took the world by surprise, and it was powerful, but it ran on old Nvidia compute technology. That tells you something important: even China's most celebrated AI breakthrough relied on chips it was no longer supposed to have access to.

The gap is measurable. According to DeepSeek's own testing, Beyond Huawei's best current chip, the Ascend 910C, achieves roughly 60% of Nvidia H100 inference performance. Inference is the step where a trained model generates answers for users. For training, the heaviest compute task of all, the Ascend 910C delivers roughly one-third the BF16 throughput of Nvidia's B200, which is the flagship chip Nvidia sells in the US market.

What alternatives does China have?

Not none, but not enough yet. Chinese semiconductor firms delivered 1.65 million AI GPUs in 2025 out of a total of 4 million units, claiming 41% of the local AI server market, with Huawei alone shipping roughly 812,000 AI chips.

Huawei is the centerpiece of the domestic push. Huawei's upcoming Ascend 960 AI chips, to be commercially available in 2027, are being positioned as China's answer to Nvidia. A further upgrade, the Ascend 970, is already planned for 2028, with each new generation promising to double the computing power of its predecessor. But Huawei's AI chips are still trailing Nvidia's top offerings.

Beyond Huawei, at the high end, national champions are deploying massive clusters for foundational model training, a domain once exclusively held by Nvidia: Baidu is leveraging a 30,000-chip cluster of its own Kunlun processors to train its ERNIE large language models, while DeepSeek has publicly opted for Huawei's Ascend chips for some training tasks.

The manufacturing constraint is the hardest wall to climb. Wafer fab capacity remains a big constraint, with all certified Chinese chipmakers competing for limited production slots at SMIC, whose most advanced stable node is its N+2 process, roughly equivalent to 7nm, while Nvidia manufactures on TSMC's 4nm process. That gap in manufacturing precision translates directly into a gap in chip performance and energy efficiency that software alone cannot bridge.

Morgan Stanley estimates China's total AI chip market could reach $67 billion by 2030, with domestic supply covering roughly 76% of demand. That would be a dramatic shift, but it is still a projection, and it still leaves a meaningful gap at the frontier.