The US has tightened export controls on advanced AI chips, closing a loophole that let Chinese companies buy the best Nvidia processors through overseas subsidiaries. This move changes the game for China’s AI chip industry.

For years, China relied on Nvidia’s powerful GPUs to run AI models. But US rules blocked direct sales of these chips to China. Some Chinese firms sidestepped this by ordering chips through affiliates in places like Malaysia or Singapore.

Now, the US Department of Commerce says that if a company is headquartered in China, any shipment to its subsidiaries anywhere needs a license. This ends the trick of routing chip purchases abroad to avoid restrictions.

China’s tech giants are responding by building their own AI chips. Instead of copying Nvidia’s general-purpose GPUs, they focus on custom ASICs—chips designed for specific AI tasks. These chips sacrifice flexibility but offer better speed and power efficiency for targeted workloads.

Huawei leads this shift with its Ascend series of neural processing units. It holds an estimated 62% of China’s AI chip market. Cambricon and Alibaba are also developing ASICs with different architectures to support AI workloads.

A Split in AI Chip Design

The US export controls pushed Chinese firms to rethink their chip designs. GPUs are versatile and ideal for AI research because they handle various tasks. But they are expensive and power-hungry.

ASICs, on the other hand, focus on doing one job very well. This suits China’s AI market, which prioritizes deploying AI applications to millions of users rather than experimenting with new models. Inference—the phase of running trained models—benefits from ASICs’ efficiency.

Huawei’s Ascend 950 and Cambricon’s Siyuan 690 ASICs now outperform Nvidia’s most powerful export-allowed chip, the H20, by 50% to 150% in some benchmarks. This shows China is closing the performance gap despite US restrictions.

The Software Challenge and Future Outlook

Hardware is only half the story. Nvidia’s CUDA software ecosystem dominates AI development globally. It makes programming GPUs easier and supports countless AI frameworks and models.

Chinese firms face a tough task breaking into this ecosystem. Huawei created its own software stack called CANN. Moore Threads developed MUSA to support its GPUs. Some AI startups are rewriting their code to run on these new platforms.

Experts say the line between GPUs and ASICs is blurring as AI models grow more complex. The future may combine flexibility with efficiency, mixing both chip types in AI systems.

US export rules keep tightening, aiming to maintain America’s edge in AI hardware. But China’s push for self-reliance is accelerating innovation in chip design and software. This could lead to two very different AI ecosystems, with China focused on custom chips optimized for large-scale deployment.

For Nvidia and AMD, the new rules reduce their potential market in China. Chinese companies must choose between applying for US licenses or switching to homegrown chips. The chip war is now as much about architecture and software as it is about supply chains.