China’s stunning achievements in AI have one glaring weak spot: access to compute—the raw processing power that fuels AI and relies on large volumes of advanced semiconductors. The U.S. currently has a tenfold advantage over China in total compute capacity, a gap that may only widen over time. U.S. tech firms are pouring billions of dollars into new data centers and can reap the benefits of the latest chip advancements from Nvidia and AMD or their own self-developed AI chips.
Meanwhile, the performance and volume of foreign AI chips that Chinese firms can obtain have gone down over time due to increasingly stringent U.S. export controls. Chinese tech leaders such as Tencent, Baidu, and DeepSeek have called out compute constraints as a key bottleneck to faster AI development.
Huawei’s domestically produced AI chips, known as the Ascend series, might seem like the obvious solution to China’s compute challenges. But there’s a catch: Chinese tech firms don’t want to use Huawei’s chips, which lag behind their foreign counterparts, for training their AI models. In 2024, Chinese companies bought around 1 million Nvidia H20 chips compared with an estimated shipment of 450,000 Huawei Ascend 910B chips.
Only a handful of state-backed companies in China have used Huawei chips to train their models, including iFlytek, SenseTime, and China Mobile. Chinese companies are dragging their feet on switching to domestic AI chips despite pressure from Chinese central government agencies to do so.
Chinese AI developers overwhelmingly prefer using Nvidia chips—even severely performance-degraded ones—and go to great lengths to access them. Many of China’s top AI models today are still trained on Nvidia’s hardware, including DeepSeek’s V3 model and Moonshot’s Kimi K2 model. In anticipation of the U.S. ban on Nvidia’s H20 chips, ByteDance, Alibaba, and Tencent rushed to spend $16 billion to stockpile roughly 1.3 million to 1.6 million H20 units.
At the end of 2024, ByteDance had planned to spend $7 billion to access Nvidia chips on servers outside of China. Chinese tech companies have been scouring black markets across Asia as well as e-commerce sites to acquire banned Nvidia chips for as much as double their normal pricedouble the price. Chinese buyers have even resorted to buying Nvidia’s RTX gaming chips as substitutes, even though they are not designed for AI workloads, and smuggling hard drives full of data out of the country to train models on servers outside of China.
Why are China’s AI developers so reluctant to switch from Nvidia to Huawei, even as their access to Nvidia chips becomes increasingly constrained?
First, Nvidia’s degraded chips for sale to China still outperform Huawei’s chips in some important dimensions. Huawei’s Ascend 910B chips use older HBM2E memory technology, offering only two-thirds of the memory capacity and 40 percent of the bandwidth of Nvidia’s H20 chips.
Huawei’s newer Ascend 910C chips, which are ramping up production this year, offer 80 percent of the H20’s bandwidth but still use the older HBM2E memory standard that is two generations behind the most advanced AI chips. This gap in memory performance is particularly important given the rise of reasoning models and inference, where memory bandwidth plays a vital role.
A second key reason why Chinese tech companies can’t easily quit Nvidia is the same reason American tech companies can’t, either: CUDA. Nvidia’s parallel computing platform, launched in 2006, has accumulated and is tightly integrated with PyTorch, the dominant AI framework, creating a mature software ecosystem that locks developers into Nvidia’s AI systems.
For Chinese tech firms, switching away from Nvidia means rewriting code, abandoning this industry-leading infrastructure, and losing access to the applications in CUDA libraries built up over years by global developers. Huawei’s alternatives—its CANN platform and MindSpore framework, launched in 2018 and 2019—are newer, less mature, and plagued by technical issues including bugs, crashes, and overheating.
With a far smaller AI hardware user base than Nvidia’s systems, Huawei lacks the high-volume, real-world feedback from major customers needed to rapidly refine its chips and software. As a result, Huawei’s AI solutions are unable to take advantage of the kind of iterative optimization that made China a global leader in other industries.
While access to Nvidia chips is becoming increasingly difficult, the supply of Huawei’s chips remains both constrained and uncertain. U.S.-led export controls on semiconductor manufacturing equipment to China have limited the country’s chipmaking capabilities.
In particular, Huawei and SMIC have been struggling to ramp up production of advanced chips at the 7-nanometer process level or below. A lack of access to extreme ultraviolet lithography (EUV) machines from ASML and U.S. tools for key tasks such as etching and deposition have made it difficult for SMIC to manufacture advanced chips precisely and reliably, keeping its production yield far below industry leader TSMC.
While SMIC is making steady progress and Huawei is on track to sell over a million Ascend dies this year, Huawei also illegally procured more than 2 million of TSMC’s logic dies, a core chip component, for its Ascend 910B and 910C chips in 2024. Ironically, because Huawei is already heavily sanctioned, it faces little punishment for skirting export controls in this way.
Chinese companies are also wary of the additional commercial and geopolitical risks involved with Huawei, which has been a frequent target of the U.S. government for years. For example, the U.S. Department of Commerce warned in May that using Huawei chips “anywhere in the world” would violate U.S. export control rules before later adjusting its announcement.
Huawei is not only a chip supplier to Chinese tech companies but also a powerful competitor. Huawei is China’s second-largest cloud service provider and has developed its own open-source Pangu family of AI models. Other Chinese tech companies are jostling with Huawei to provide cloud services not just within China but globally as well—making Huawei’s chips an unpopular option for firms competing in the same space.
However, this could all change if the U.S. makes the wrong decisions.
While Huawei’s chips have lower bandwidth memory performance than Nvidia’s H20 chips, Huawei’s 910B and 910C chips already offer greater total processing performance (TPP) and better energy efficiency (TPP/watt) than the H20. Far from being a “powerful chip” as some have claimed, the H20 actually has worse computational performance and energy efficiency than Nvidia’s older A100 chips, launched back in 2020. Moreover, Huawei has been able to improve the computational performance of its Ascend chips, even after switching production from Taiwanese chip manufacturer TSMC to China’s own SMIC following U.S. export controls.
Perhaps more importantly, Huawei has been making significant progress at the level of AI computing systems. Huawei recently unveiled its CloudMatrix 384 system, made up of 384 of Huawei’s latest 910C chips and a novel all-optical networking approach. According to SemiAnalysis, Huawei’s new CloudMatrix system outperforms Nvidia’s state-of-the-art GB200 NVL72 system on key dimensions, such as compute power (how fast the chip can process large volumes of data), memory bandwidth, and integrated networking.
While Huawei’s new system is a lot more costly and energy-intensive than Nvidia’s counterpart, which may limit customer adoption, it marks a striking advancement in system-level performance, which may be even more important than individual chip performance for scaling up large AI compute clusters.
In a recent technical paper, Huawei has already proved that its new CloudMatrix system can be successfully used to train advanced AI models. The pricing and energy issues will likely be manageable for Huawei as it continues to invest heavily in R&D and receive significant state support.
As Huawei’s AI systems continue to improve, U.S. export control policies must be carefully calibrated to avoid pushing China’s AI industry too far. If China’s domestic AI chips continue to improve, while U.S. chips available in China are further downgraded by export controls, there will be a crossover point where the performance of Chinese chips clearly exceeds that of American chips available in China.
The crucial tipping point could be if China’s largest tech companies, such as Alibaba, Tencent, and ByteDance, throw their formidable resources toward working with Chinese AI chipmakers. This would kick off a positive feedback loop for China’s AI chipmakers, particularly Huawei, building up the software libraries and tools for creating a complete Chinese AI hardware-software ecosystem. Once this process is underway, it would also mark a point of no return for American AI chipmakers like Nvidia in the China market.
There are already some signs of this potential shift. DeepSeek and ByteDance are experimenting with using Huawei’s AI chips to run their AI models. Ant Group, a spinoff of Alibaba, is even testing the use of Huawei’s chips for model training. Huawei’s Ascend developer community has grown nearly tenfold in the past four years, though it still remains far smaller than Nvidia’s.
Other Chinese AI chipmakers besides Huawei are making progress as well, including Cambricon, Biren, Moore Threads, Enflame, and Hygon. Cambricon saw its first-quarter revenue surge more thanforty‑foldfortyfold last year and received a large order for its AI chips from ByteDance, Cambricon’s revenue is forecasted to grow 3.7 times to 5.5 billion yuan this year according to Goldman Sachs. Nvidia’s CEO, Jensen Huang, has said that Nvidia’s market share in China has declined from 95 percent to 50 percent—a claim supported by other credible analysis.
The U.S. needs a more sophisticated approach to export controls. The reversal of the H20 chip ban by the Trump administration was a step in the right direction. At the same time, the White House’s new AI action plan correctly recognizes that winning the AI race with China depends on making the U.S. tech stack, including its AI chips, the dominant platform for global AI development.
Semiconductor export controls are not as simple as tightening the valve on a tap. China’s AI chip dilemma is not just a hardware problem but an ecosystem one. Huawei now has access to many of the key resources it needs to develop advanced AI chips, including financing and talent. But it’s missing a large and dedicated customer base that is committed to co-refining the software and hardware Huawei offers.
A smart approach to export controls would focus on setting a performance threshold for AI chips that can be sold to China based on a window between U.S. and Chinese hardware capabilities. The performance threshold should be high enough to outperform China’s domestic hardware options to ensure Chinese developers remain on U.S. platforms. At the same time, it should be low enough to maintain a significant performance gap with hardware systems available to American developers.
Ideally, this performance threshold would include a buffer, such as a 50 percent performance advantage over Chinese hardware systems on key metrics, in anticipation of improvements in Chinese hardware offerings. A regular yearly update, with ad hoc changes for unexpected developments, would likely be sufficient to adjust for advances made in Chinese AI chips while providing enough policy stability for industry participants.
The overarching policy goal is clear: Ensure the U.S. continues to lead the world in AI. By constraining China’s access to cutting-edge chips without pushing Chinese AI developers to make the leap to China’s own domestic chips, the U.S. can use export controls to help make this a reality.
The views expressed by the authors do not represent those of their affiliated institutions.