Huawei’s Chips Are an Unpopular Alternative to Nvidia for Chinese Firms

China’s amazing achievements in artificial intelligence have one weak spot: arrival to the account – the power of raw treatment that feeds artificial intelligence and depends on large amounts of advanced semiconductors. The United States is currently ten times a feature of China in the total account capacity, a gap that may expand only over time. American technology companies flow billions of dollars in new data centers and can reap the benefits of the latest developments from NVIDIA and AMD chips or sophisticated artificial intelligence chips.

Meanwhile, the performance and size of foreign artificial intelligence chips that Chinese companies can have has decreased over time due to increasingly American export controls. Description of Chinese technology leaders such as Tencent, Baidu and Deepseek account restrictions as the main bottle neck of AI’s faster development.

Locally produced artificial intelligence chips from Huawei, known as the Ascending series, may seem a clear solution to China’s account challenges. But there is hunting: Chinese technology companies do not want to use Huawei chips, which are behind their foreign counterparts, to train artificial intelligence models. In 2024, Chinese companies bought about one million NVIDIA H20 chips compared to an estimated charging of 450,000 Huawei Ascends 910B.

Only a few state -backed companies in China have used Huawei chips to train their models, including IFlytek, Sensetime and China Mobile. Chinese companies withdraw their feet to turn into local artificial intelligence chips despite the pressure from the Chinese central government agencies to do so.

Chinese artificial intelligence developers prefer an overwhelming majority to use NVIDIA chips-even those that have been largely identified-and go to the extreme to reach them. Many of the best artificial intelligence models in China are still trained on NVIDIA devices, including the V3 Deepseek and the Kimi K2 model of Moonshot. In anticipation of the American ban on the H20 chips in NVIDIA, it rushed by BYTEDANCE, alibaba and TENCENT to spend $ 16 billion to store approximately 1.3 million to 1.6 million units H20.

At the end of 2024, Bytedance planned to spend $ 7 billion to reach NVIDIA chips on servers outside China. Chinese technology companies roam black markets throughout Asia as well as e -commerce sites to acquire banned NVIDIA chips by double the normal price. Even Chinese buyers have resorted to the purchase of RTX games in NVIDIA as alternatives, although they are not designed to the burdens of artificial intelligence work, and smuggling of hard -to -country hard drives to train models on servers outside China.

Why does artificial intelligence developers in China refrain from shift from NVIDIA to Huawei, even when their arrival in NVIDIA chips become increasingly restricted?

First, the deteriorating NVIDIA chips for sale to China are still outperforming Huawei chips in some important dimensions. HUAWEI’s ASCEND 9B chips use the oldest HBM2E memory technology, providing two -thirds of the memory capacity and 40 percent of the H -DIDIA H20 Potatoes width in NVIDIA.

The latest ASCEND 910C chips from Huawei, which discover production this year, provides 80 percent of the H -20 buttons but still uses the oldest HBM2E memory standard that represents two generations behind the most advanced artificial intelligence chips. This gap in memory performance is especially important given the appearance of thinking and inference models, as the memory domain width of the memory plays a vital role.

The second main reason that makes Chinese technology companies cannot easily give up NVIDIA is the same reason that American technology companies cannot, either: Cuda. The parallel computing platform for NVIDIA, which was launched in 2006, collected users and tightly integrated with Pytorch, the dominant AI frame, creating a mature environmental system for programs that secure developers in AI in NVIDIA.

For Chinese technology companies, the switching from NVIDIA means rewriting a code, giving up this leading infrastructure in this field, and losing access to applications in Cuda libraries that have been built over years by international developers. Huawei – Cann Platform and Mindspore Framework, which was launched in 2018 and 2019 – are the latest and less mature, and are affected by technical problems including insects, disruption, and fragmentation.

With the AI user base with much smaller NVIDIA systems, HUAWEI lacks large and realistic comments from the main customers who need to improve their chips and programs quickly. As a result, the AI solutions from Huawei are unable to benefit from the type of repetitive improvement that made China a global leader in other industries.

While reaching NVIDIA chips is increasingly difficult, the provision of Huawei chips is still restricted and uncertain. The US -led export controls have led to the semiconductor manufacturing equipment to China to reduce the capabilities of the country’s manufacturing.

In particular, Huawei and SMIC were rewarded to increase the production of advanced chips at the level of 7 nm or less. The lack of access to the ASML EUV and American tools for the main tasks such as drilling and deposit has made it difficult for SMIC to manufacture advanced and reliable advanced chips, while maintaining a much lower production revenue than the TSMC leader.

While SMIC makes a steady progress and Huawei on the right track to sell more than a million ascending this year, Huawei has been purchased illegally more than two million TSMC logic, which is composed of basic chips, for the rise of 910B and 910C in 2024.

Chinese companies are also cautious about the additional commercial and geographical risks involved in Huawei, which was a frequent target for the US government for years. For example, the US Department of Commerce warned in May that the use of Huawei chips “anywhere in the world” would violate US export control rules before amending its announcement later.

Huawei is not only a chip -chip -technology supplier but also a strong competitor. Huawei is the second largest cloud service provider in China, and has developed the Pangu Open Open Openge family of artificial intelligence models. Other Chinese technology companies wander with Huawei to provide cloud services not only inside China but also worldwide – which makes Huawei chips an unpopular option for competing companies in the same space.

However, all this can change if the United States makes wrong decisions.

While HUAWEI chips have a NVIDIA H -DIDIDID frequency domain, Huawei and 910C chips already provide larger TPP performance and better energy efficiency (TPP/Watt) from H20. Away from being a “strong slice” as some claimed, the H20 actually has my worse and efficiency of energy than the old A100 chips in NVIDIA, which was launched in 2020.

Perhaps most importantly, Huawei has made great progress at the level of artificial intelligence computing systems. Huawei has recently unveiled the Cloudmatrix 384, which consists of 384 of the latest 910C chips in Huawei and the new visual network approach. According to Semianlysis, the new Huawei Cloudmatrix system outperforms the GB200 NVL72 system that revolves around NVIDIA, such as arithmetic energy (extent to which the chip speed of large volumes of data), the frequency range of memory, and integrated networks.

While the new Huawei system is more expensive and intense than NVIDIA, which may limit customer dependence, it represents amazing progress in the system level, which may be more important than the performance of individual chips to expand the large AI account groups.

In a modern technical sheet, Huawei has already proven that the new Cloudmatrix system can be used successfully to train advanced artificial intelligence models. Pricing and energy problems are likely to be manageable for Huawei, as they continue to invest intensively in research and development and receive great support for the state.

As AI’s systems continue to improve in Huawei, US export control policies must be carefully calmed to avoid pushing artificial intelligence in China away. If artificial intelligence chips in China continue to improve, while the US chips available in China are reduced due to export controls, then there will be an intersection point where the performance of Chinese chips clearly exceeds the width of the American chips available in China.

It can be the decisive turning point if it is the largest technology companies in China, such as Alibaba, Tencent and Bytedance, throwing its huge resources towards working with Chinese AI makers. This would start the positive feedback ring for the artificial intelligence maker in China, especially Huawei, to build libraries and software tools to create a complete ecosystem for Chinese devices. Once this process begins, it will also represent a point for the return of American artificial intelligence chips like NVIDIA in the China market.

There are already some signs of this potential transformation. Deepseek and Bytedance experience using Huawei’s AI chips to run artificial intelligence models. Ant Group, which is a collection of alibaba, is to test the use of Huawei Training chips. The Huawei Ascend Developer has grown nearly ten times in the past four years, although it is still much smaller than NVIDIA.

The Chinese artificial intelligence makers alongside Huawei also make progress, including Cambricon, Biren, Moore Threads, Enflame and Hygon. Cambricon has seen an increase in its revenues in the first quarter more than that in the past year and received a large demand for artificial intelligence chips from Bytedance. The CEO of NVIDIA, Jensen Huang, said that the market share in NVIDIA in China has decreased from 95 percent to 50 percent – a demand for the support of another reliable analysis.

The United States needs a more advanced approach to export controls. It was the opposite of the H20 chip ban by the Trump administration a step in the right direction. At the same time, the new artificial intelligence plan at the White House admits that winning the artificial intelligence race with China depends on making the American technology staple, including artificial intelligence chips, the dominant platform for developing global artificial intelligence.

Semiconductor export controls are not simple as the valve tightening on the tap. The artificial intelligence chip dilemma in China is not just a problem with devices but an ecosystem. Huawei now has access to many major resources you need to develop advanced artificial intelligence chips, including financing and talent. But it lacks a large and customized customer base committed to reshaping the program and the devices offered by Huawei.

The smart approach to export controls will focus on determining a performance threshold for artificial intelligence chips that can be sold to China based on a window between us and the capabilities of Chinese devices. The performance threshold should be high enough to outperform the options of local devices in China to ensure that Chinese developers remain on American platforms. At the same time, it should be low enough to maintain a large performance gap with the systems available for American developers.

Ideally, the performance threshold will include this temporary store, such as 50 percent performance feature on Chinese devices systems on the main standards, in anticipation of improvements in Chinese devices. Regular annual update, with changes for unexpected developments, is likely to be adaptive to the progress of Chinese artificial intelligence chips while providing enough stability in politics to participants in this field.

The goal of the comprehensive policy is clear: make sure the United States continues to lead the world in artificial intelligence. By restricting China’s access to advanced chips without pushing the Chinese artificial intelligence developers to jump to local chips in China, the United States can use export controls to help make this a reality.

The opinions expressed by the authors do not represent the opinions of their affiliate institutions.