Huawei’s rotating chairman Eric Zhijun Xu at Huawei Connect 2026 in mid-September sat down with journalists alongside Liao Heng, Huawei’s chief scientist for semiconductors, for an extensive discussion of the company’s AI computing strategy. Multiple Chinese media outlets, including Guancha, published the same transcript on Tuesday, Oct. 6.
Much of the conversation was highly technical, covering Huawei’s new computing architecture, chip interconnection technologies, and its efforts to compete with Nvidia despite restrictions on access to advanced semiconductor manufacturing.
But five remarks stand out for a broader audience, without having to delve into the technological details.
1. China may eventually share American companies’ concerns about AI risks—but for now, it needs to accelerate.
Asked about calls by some leading American AI developers to slow down AI development, Xu offered an interesting explanation for why Chinese companies may see the risks differently. American frontier developers possess considerably greater computing resources and may have reached levels of AI capability that Chinese developers have yet to experience.
“Several leading American model developers have access to much greater computing power. Exactly how far they have progressed is something that perhaps only they themselves know. The AI risks they are experiencing may not yet be felt nearly as strongly in China.
For now, China still needs to accelerate its own pace of development. Once we reach the point where we can genuinely perceive those risks, perhaps we will feel the same way as the leading American model developers. But I have always believed that we must both advance AI development and manage its risks. We must ensure that AI is used for good rather than for evil.”
Xu’s point is not that those risks are imaginary or should be disregarded. Rather, he suggests that China may come to share American frontier developers’ concerns as its own AI capabilities advance.
2. The global AI hardware supply crunch may not end until 2029—and China could take longer.
The rapid expansion of AI infrastructure has strained global supplies of chips, memory, optical networking equipment, and other critical components. Xu believes the industry was not prepared for the speed of growth and predicts that supply will take several more years to catch up with demand.
“The reason optical module manufacturers and memory companies can charge such high prices today is not that their products are necessarily exceptionally good. It is because demand vastly exceeds supply.
Therefore, these bottlenecks will not be fully resolved until supply and demand reach equilibrium both globally and in China. I expect the global supply-demand imbalance to be largely resolved around 2029. China may take a little longer.”
His forecast is particularly notable given China’s additional constraints on access to advanced semiconductor technologies. It also suggests that the current boom in AI infrastructure investment may continue to face supply bottlenecks for several years.
3. Nvidia’s CUDA software advantage is becoming less formidable, according to Huawei’s semiconductor chief scientist.
Nvidia’s competitive strength extends well beyond its chips. Its CUDA software ecosystem has long made it difficult for AI developers to switch to alternative hardware, including Huawei’s Ascend processors.
But Liao Heng argues that advances in AI programming are beginning to erode this advantage.
“Today, with the emergence of new compiler languages, these tools are gradually replacing aspects of the CUDA-based development approach and weakening CUDA’s ecosystem barriers. Frontier laboratories no longer attach the same importance to CUDA that they did two years ago.
This represents the gradual dismantling of barriers at the software level. I believe we are approaching, or may even have reached, a point of balance. The gap has narrowed substantially.”
Liao also said that the overall gap between Nvidia and Ascend had narrowed considerably, including in how familiar developers find the programming environment. This is Huawei’s assessment rather than an independently established parity between the two ecosystems, but it addresses what has historically been one of Nvidia’s most important competitive advantages.
4. Huawei believes Ascend may already have overtaken Nvidia in China, but cannot meet domestic demand.
U.S. export restrictions have created opportunities for Chinese AI chipmakers, particularly Huawei. Xu made a striking claim about Ascend’s position in China, while acknowledging that the company’s manufacturing capacity remains insufficient.
“At present, our Ascend supply is still far from sufficient to meet domestic demand, so we have no plans for a comprehensive expansion into overseas markets.
There are indeed a small number of countries with particularly strong needs, and we have conducted tests and supplied products to them, but the quantities have been very small.
It is difficult to calculate Nvidia’s market share in China. However, based on the data we are able to gather, Ascend should already have surpassed Nvidia.”
Xu did not specify whether his comparison was based on shipments, revenue, computing capacity, or another measure, so the market-share claim cannot be independently assessed from his remarks alone. Nevertheless, it is noteworthy that Huawei now believes Ascend has surpassed Nvidia in its home market.
5. Huawei will prioritize Chinese customers, while serving a limited number of overseas buyers seeking alternatives.
Asked about reports that Malaysia was considering deploying sovereign AI infrastructure based on Huawei’s Ascend 910C, Xu acknowledged that geopolitical considerations and the desire to diversify suppliers could make Huawei attractive to some overseas customers.
But he made clear where the company’s priorities lie.
“As for balancing domestic and overseas demand, our principle is very straightforward. Our primary focus is the domestic market. We give priority to Chinese customers who urgently need computing power.
Overseas, we serve a small number of customers who cannot purchase alternative solutions, or who are determined to have another option.”
In other words, Huawei is not currently positioning Ascend for a broad international sales push. Instead, its limited overseas supplies are aimed primarily at customers who cannot obtain competing products or who want an alternative supplier, while China’s enormous domestic demand takes precedence.
The complete Q&A covers Huawei’s broader technological strategy, including its new Peerium computing architecture, UnifiedBus interconnection technology, plans for future Ascend chips, and efforts to improve the efficiency of large-scale AI training. It is very long and technical, so I will not put out a translation.
One other Huawei development during China’s just-concluded National Day holiday deserves attention. On October 5, Huawei and Qualcomm announced a multiyear patent cross-licensing agreement covering 5G, computing, AI, and networking, alongside Qualcomm’s planned purchase of certain Huawei patents in the United States.
Some Chinese media reports claimed that Qualcomm would become a net payer to Huawei under the licensing agreement—meaning it would pay Huawei more in royalties than it receives. Qualcomm subsequently disputed this, saying the terms were confidential and the reports incorrect.
Just over a month earlier, on August 26, Huawei and HP had announced a global patent cross-licensing agreement covering Wi-Fi technology, although HP subsequently emphasized that the deal settled a patent dispute rather than represented a broader strategic or commercial partnership.
There is something striking about a Chinese company subjected to years of U.S. sanctions continuing to conclude patent deals with some of America’s leading technology companies. Whatever the financial terms, these agreements offer tangible commercial recognition of the value of Huawei’s intellectual property.
They also present an interesting contrast with the ongoing federal criminal trial in Brooklyn, where U.S. prosecutors have accused Huawei of stealing American technology to build its telecommunications empire—allegations Huawei denies. The patent agreements offer a rather different perspective on Huawei’s technological capabilities.