China is moving quickly to develop home-grown artificial intelligence chips, raising serious concerns for Nvidia, the U.S. company that has dominated the global GPU market. Chinese firms, backed by billions in government funding, are attempting to build chips that can train and run AI models without relying on American sources. Many in the industry argue that this effort is less about genuine innovation and more about stealing and copying existing designs. The signs coming from leading Chinese startup DeepSeek suggest that Beijing is closer than ever to reaching that goal.
Why China Is Racing to Build Its Own Chips
The motivation for China’s push is clear. U.S. export restrictions have blocked Chinese companies from acquiring Nvidia’s most advanced processors. This has put Chinese AI firms at a disadvantage, forcing them to look inward. DeepSeek revealed in a recent announcement that its latest V3.1 model was built “for the home-grown chips to be released soon.” The one-line note, posted on the company’s WeChat account, hinted that a new class of domestic processors is about to be unveiled, though the company refused to name a supplier or provide details.
By tailoring its software to work with Chinese chips rather than Nvidia’s, DeepSeek is helping create what analysts describe as a self-sufficient AI “stack.” This would allow China to cut dependence on foreign suppliers, a major strategic goal for Beijing. As one report explained, the announcement “hints that China has made key progress in building a self-sufficient AI stack consisting of domestic technologies, a development that could help the country shrug off US chip export restrictions.”
What the New Chips Are Based On
DeepSeek’s announcement was not just about hardware but also about the software techniques being used to support these chips. The company explained that its model was trained “using the UE8M0 FP8 scale data format to ensure compatibility with microscaling data formats.” FP8, or floating-point 8, is a method for reducing precision in calculations, which in turn speeds up AI training and inference.
The company went further by describing the efficiency benefits of this system. According to DeepSeek, the UE8M0 format could “cut memory use by up to 75 per cent.” This is significant because it lowers the hardware requirements needed to run large models. Such techniques closely resemble strategies that Nvidia pioneered, which has led many observers to conclude that China is building on American breakthroughs rather than investing billions to develop its own.
The Role of Chinese Firms Like Huawei
DeepSeek is not alone in this effort. Local chip developers such as Huawei Technologies and Moore Threads have rushed to adapt their processors to run Chinese AI models. Reports note that “Chinese AI infrastructure start-ups, including SiliconFlow, have already moved to use Huawei chips to power DeepSeek systems.”
Joint research published in June even claimed that Huawei’s Cloud Matrix 384 architecture, combined with its Ascend chips, could “run DeepSeek’s R1 models more efficiently than a system using Nvidia’s H800 chips.” If true, it suggests that Chinese firms may not only be copying Nvidia’s designs but also fine-tuning them for specific uses in a way that could make them more efficient for local markets.
Why This Threatens Nvidia
For Nvidia, the stakes could not be higher. Its H800 and H20 chips have been central to China’s AI boom, with DeepSeek admitting that its earlier V3 model was trained on “2,048 Nvidia H800 chips.” This means Chinese engineers had direct experience with Nvidia’s hardware at scale, giving them insights into how to reproduce and modify those designs.
Nvidia’s profit margins in China have been immense, with U.S. financial firms estimating that its GB200 NVL72 system delivered “a 77.6 per cent margin, the highest among eight tested systems.” But if Chinese chips can deliver similar performance at lower cost, Nvidia risks losing one of its largest and fastest-growing markets. Worse, much of the technology enabling these chips was not developed through expensive research, but instead through reverse-engineering.
China’s strategy fits into its broader push for technological independence. The government has already invested over $300 billion into its semiconductor industry and is expected to raise domestic AI chip production to 55 percent by 2027. Export controls by Washington have only sped up this effort. As one analysis put it, “limited access to advanced foreign chips is creating conditions where Chinese firms must rapidly improve domestic alternatives, potentially accelerating innovation timelines that might have taken decades under normal market conditions.”
What makes this development alarming for Nvidia and other U.S. firms is that China is not starting from scratch. By training earlier models on thousands of Nvidia chips, companies like DeepSeek have gained deep knowledge of how Nvidia’s architecture works. From there, it is a much shorter path to designing similar processors with small tweaks that allow them to bypass sanctions.
DeepSeek’s recent hints about “next generation” home-grown chips underline how serious this challenge has become. In just two years, the company has gone from releasing its R1 reasoning model to announcing a V3.1 system built specifically for Chinese processors. That pace of development is unusual and suggests state support and strategic coordination.
For now, major Chinese tech companies like Tencent and Baidu still rely heavily on Nvidia’s chips, citing performance issues with some domestic alternatives. But if DeepSeek and its partners succeed, it will only be a matter of time before China reaches performance parity. When that happens, the U.S. may find that its strategy of cutting off Nvidia sales has backfired, giving Chinese firms an incentive to accelerate their own chip programs through methods that critics say are closer to piracy than to innovation.
ACZ Editor: There is no possible way that China would have any aspect of artificial intelligence, neither hardware nor software, without the theft of designs and other intellectual property. This is potentially a multi-trillion dollar industry. Stolen.
