Huawei AI chip demand outruns supply as Ascend 960 targets Nvidia in 2027

Huawei Technologies has disclosed perhaps the clearest evidence yet that China’s domestic artificial-intelligence infrastructure push is running into a problem that is simultaneously encouraging and restrictive for the company: demand is exceeding its ability to supply hardware. Huawei rotating chairman Eric Xu said at the Huawei Connect conference in Shanghai on September 17 that the company does not currently have enough production capacity to meet Chinese demand for its AI computing products, limiting its ability to expand internationally on a large scale.

The disclosure comes as Huawei accelerates development of its next generation of Ascend processors. The company plans to make the Ascend 960DT available in the first quarter of 2027, three quarters earlier than previously expected, followed by the Ascend 960PR in the third quarter, also ahead of its earlier schedule. Huawei additionally unveiled a new Peerium Computing Architecture intended to allow processors at enormous scale to function as a single computing system, with UnifiedBus serving as a core interconnect technology.

Why is Huawei struggling to meet Chinese AI chip demand?

China’s AI industry has been forced to develop a more self-contained computing ecosystem as U.S. export restrictions limit access to Nvidia Corporation’s most advanced AI processors and restrict the semiconductor manufacturing equipment needed to produce leading-edge chips domestically. That has increased demand for Huawei’s Ascend products among Chinese AI laboratories, cloud providers and companies building large-scale computing clusters. Huawei has faced separate U.S. trade restrictions since 2019, making its expanding semiconductor portfolio both a commercial business and a strategic response to technological isolation.

Huawei said testing of the Ascend 950DT has delivered positive results and that the company is discussing deployments extensively with Chinese AI model developers. Eric Xu expects systems built around that chip to be used more widely for model training during 2027. The company has consequently chosen to prioritise the domestic market rather than pushing aggressively overseas while manufacturing remains constrained.

That supply constraint is one of the most consequential elements of Huawei’s announcement. A company struggling to find customers faces a demand problem, but Huawei currently faces the opposite issue: its addressable domestic market is growing faster than its ability to manufacture and deploy enough computing equipment. The resulting bottleneck could affect how quickly Chinese AI developers can increase model size, training runs and inference capacity.

Is Huawei really bigger than Nvidia in China’s AI chip market?

Eric Xu said he believes Huawei’s Ascend products now hold a larger share of China’s AI chip market than Nvidia, although Huawei did not provide independent market data supporting that estimate. The claim therefore reflects Huawei management’s view rather than an independently established market-share figure.

Even without accepting a specific market-share percentage, Nvidia’s position inside China has unquestionably become more complicated. U.S. restrictions have repeatedly limited which Nvidia products may be supplied to Chinese customers, creating opportunities for local alternatives that might otherwise have struggled against Nvidia’s CUDA software ecosystem and enormous installed base.

Huawei’s strategy is also increasingly about systems rather than individual chips. Advanced AI workloads require hundreds, thousands or eventually tens of thousands of processors to communicate rapidly with one another, meaning interconnect technology and software orchestration can partly compensate when individual accelerators do not match the performance of the world’s most advanced chips.

Why does Huawei want to connect as many as one million AI processors?

Huawei’s new Peerium Computing Architecture is designed around massive-scale parallel computing. The company says the system introduces nested parallel processing, unified memory addressing and peer-to-peer interconnection so that extremely large numbers of processors can behave more like one computing resource. Huawei describes UnifiedBus as the technology enabling that architecture.

The approach reflects an important reality in the semiconductor race. China may not always have access to the most advanced fabrication technology, but a weaker individual processor can become much more useful if enormous numbers of chips can be efficiently connected without excessive communication delays.

Huawei has therefore placed increasingly heavy emphasis on SuperPoD and large-cluster architectures. Reuters reported that the company ultimately wants systems capable of linking as many as one million AI processors, while the newly accelerated Ascend 960 generation is intended to strengthen both training and inference capability.

The economics are challenging because scaling with more processors increases power consumption, cooling requirements, networking complexity and hardware cost. Nevertheless, China’s need for computing capacity gives Huawei an incentive to solve those engineering problems rather than waiting for unrestricted access to leading U.S. accelerators.

What does Huawei’s new roadmap mean for Nvidia?

Nvidia still possesses major advantages in performance, developer tools and global customer relationships. Its CUDA ecosystem has accumulated years of software support that cannot be reproduced simply by manufacturing another processor.

Huawei’s opportunity is different because geopolitical restrictions have altered normal competitive dynamics. Chinese customers are being encouraged to develop domestic supply chains while access to U.S. technology can change with export-control policy. That makes supply security and technological independence part of purchasing decisions alongside benchmark performance.

Nvidia shares were trading higher on September 17 while the Huawei announcements circulated, demonstrating that investors were not treating the development as an immediate threat to Nvidia’s global growth trajectory. Huawei is privately held, preventing investors from gaining direct public-market exposure to its AI-chip expansion.

The more consequential competitive issue is likely to unfold over several years. If Huawei can make its chips good enough, produce them in much larger volumes and connect them efficiently into giant computing systems, Nvidia could face a structurally smaller opportunity in one of the world’s largest AI markets even while continuing to dominate elsewhere.

Why could manufacturing capacity decide China’s AI race?

Huawei’s roadmap demonstrates that designing competitive hardware is only one part of semiconductor independence. Producing enough chips, memory, networking components and complete systems is another challenge entirely.

U.S. restrictions on advanced semiconductor manufacturing equipment have made capacity expansion particularly difficult for Chinese companies. Huawei can accelerate chip roadmaps, but the commercial impact will ultimately depend on whether suppliers can manufacture enough working processors at competitive yields.

That makes the company’s admission that demand exceeds supply especially significant. China has customers prepared to consume domestic AI infrastructure, and Huawei believes its next products can support increasingly sophisticated models. The constraint is turning those designs into sufficient quantities of deployable computing capacity.

If Huawei succeeds, the global AI chip market could become increasingly divided into parallel ecosystems. Nvidia would remain dominant across much of the world while Huawei becomes progressively more embedded in China and selected overseas markets. The Ascend 960 generation will provide an important test of how quickly that transition can occur.