SHANGHAI, China – Huawei has accelerated development of its artificial-intelligence chips and unveiled a new generation of AI computing systems as China works to reduce its dependence on Nvidia and other foreign semiconductor suppliers.
At its Huawei Connect conference in Shanghai on Sept. 17, the Chinese technology company introduced the Atlas 960E SuperPoD and upgraded its AI computing architecture. Huawei also brought forward the timetable for its next-generation Ascend 960 chips, with the Ascend 960DT now scheduled for the first quarter of 2027 and the Ascend 960PR for the third quarter.
The announcements demonstrate how quickly Huawei is expanding beyond individual processors into large-scale computing systems designed to link thousands of AI accelerators.
Huawei is building beyond the chip itself
The Atlas 960E SuperPoD is designed around Huawei’s Ascend 960 processors and uses near-packaged optics, or NPO, for high-speed connections between components.
Huawei says the system can connect up to 4,096 AI processors. Its broader architecture, called Peerium, is intended eventually to connect much larger numbers of processors into a single computing system. Huawei says its long-term architecture could scale to as many as 1 million processors.
That reflects a central challenge in modern AI computing: performance increasingly depends not only on the capabilities of individual accelerators but also on how efficiently large numbers of processors, memory, storage and networking components work together.
Huawei said its Atlas 960 SuperPoD can improve training and inference performance for models with as many as 10 trillion parameters compared with the preceding Atlas 950 system. Those figures are company claims rather than independent benchmarks.
Ascend roadmap moves ahead
Huawei also said development of the Ascend 960 had progressed faster than expected.
■ Related News
The company now plans to release the Ascend 960DT in the first quarter of 2027 and the Ascend 960PR in the third quarter. The DT version is intended primarily for AI-model training, while the PR version is designed for inference, when trained models process requests and generate results.
Huawei also announced plans for the Ascend 970 in 2028 and Ascend 980 in 2029. The company said each generation is expected to improve computing performance as well as memory and interconnect capabilities.
The accelerated schedule comes as Chinese technology companies increasingly seek domestic alternatives to Nvidia’s AI processors.
Nvidia still has a major software advantage
Huawei’s hardware progress does not by itself mean that the company has matched Nvidia across the AI computing market.
Nvidia has built a substantial software ecosystem around CUDA, its widely used platform for programming and running applications on its processors. That software base is an important part of Nvidia’s position in AI computing and is difficult to reproduce simply by improving chip performance.
Huawei is attempting to address that problem through its own software ecosystem. The company said Ascend now supports more than 90 major open-source projects and that its computing platform has been used to train more than 40 AI models.
Huawei also said more than 5,200 developers were active each month on Ascend-related software.
Those figures come from Huawei and have not been independently audited.
Chinese demand is already exceeding Huawei’s capacity
Huawei’s challenge is not simply a question of technological development. The company is also struggling to produce enough AI computing equipment to meet demand inside China.
Rotating chairman Eric Xu said Sept. 17 that Huawei did not have sufficient capacity to satisfy Chinese demand and therefore was not planning a full-scale international expansion of its AI computing products. He said overseas supplies would remain limited.
Huawei has already deployed more than 1,000 systems based on its earlier Ascend 910C processors, according to company materials. Its newer Ascend 950 systems have also entered commercial use.
The supply constraint indicates that China’s effort to build a domestic AI-chip ecosystem is generating substantial demand, while also exposing bottlenecks in the production chain.
Advanced memory remains a constraint
High-bandwidth memory, or HBM, is one of those constraints.
HBM allows AI processors to move large quantities of data rapidly and has become a critical component of high-performance AI accelerators. U.S. export controls have restricted China’s access to some advanced HBM products, while Chinese chipmakers have faced higher costs and tighter supplies.
Reuters reported earlier in September that Huawei and other Chinese AI-chip companies had raised prices for current and upcoming processors amid higher HBM costs. Sources cited by the agency said Huawei’s Ascend 950DT was being offered at more than 250,000 yuan, although the company has not publicly confirmed those reported prices.
That creates a second challenge for China’s domestic AI industry: developing competitive processors is only one part of building a large-scale supply chain.
U.S. controls remain part of the technology race
The acceleration of Huawei’s chip program is closely tied to U.S. restrictions on advanced computing technology.
The U.S. Commerce Department’s Bureau of Industry and Security has imposed controls affecting advanced computing chips, semiconductor manufacturing equipment and high-bandwidth memory connected to China. In 2025, BIS specifically warned about certain Chinese advanced-computing integrated circuits, including Huawei’s Ascend 910B, 910C and 910D.
U.S. policy has also changed over time. In January 2026, BIS said applications to export Nvidia H200 and AMD MI325X chips to China would be reviewed case by case under specified security conditions.
The result is a technology market in which Chinese companies are simultaneously developing domestic alternatives while navigating restrictions on access to some foreign hardware, manufacturing technology and advanced memory.
China is building an alternative AI computing stack
Huawei’s latest announcements therefore extend beyond a new processor generation.
The company is attempting to build an integrated computing system encompassing AI accelerators, interconnects, optical networking, memory, storage and software. Its stated objective is to make large numbers of processors operate together efficiently rather than relying on a single high-performance chip.
China’s domestic AI industry is also not dependent on Huawei alone. Companies including Cambricon, MetaX and Iluvatar CoreX are developing or supplying alternative AI accelerators, while Chinese technology companies are increasing efforts to develop domestic computing capabilities.
The market remains constrained by manufacturing capacity, advanced memory availability and software compatibility.
For Huawei, the immediate milestone is the speed of its development. The company moved the Ascend 960DT launch forward by several months and is already planning subsequent generations through 2029.
That does not establish that Huawei has overtaken Nvidia. It does show that China’s leading chip companies are moving rapidly to narrow dependence on Nvidia’s technology and to build a domestic AI computing infrastructure capable of operating at increasingly large scale.
Reporting Credit: Huawei Technologies — Ascend 960 roadmap, Atlas SuperPoD architecture, UnifiedBus and Peerium computing systems; U.S. Department of Commerce, Bureau of Industry and Security — semiconductor, advanced-computing and HBM export-control framework.
