Huawei Unveils Atlas 960 AI System as China Expands Its AI Computing Push
Huawei has unveiled a new generation of AI computing technology as the Chinese technology company continues expanding its domestic infrastructure for training and running advanced artificial intelligence models.At Huawei Connect 2026 in Shanghai on September 17, the company introduced the Atlas 960E SuperPoD, a large-scale AI computing system built around its Ascend chip family and UnifiedBus interconnect technology. Huawei says the system is designed to connect thousands of AI processors while improving the efficiency of large AI workloads.
The announcement comes as demand for AI computing continues to grow in China, while restrictions on access to some advanced foreign chips have increased the importance of domestic alternatives.
Atlas 960E targets large-scale AI workloads
According to Huawei, a single Atlas 960E SuperPoD can connect up to 4,096 NPUs and deliver up to 8 EFLOPS of FP8 computing performance, with up to 1 petabyte of HBM capacity.The system uses Huawei's Hi-ONE optical interconnect technology and near-packaged optics, or NPO, which the company says can improve connectivity while reducing power consumption in large computing systems.
Huawei also introduced an upgraded computing architecture called Peerium, designed to allow very large numbers of processors to operate together as a single computing system.
The company says its architecture can eventually scale to one million processors, using UnifiedBus to connect computing, memory, storage and networking components.
Huawei accelerates its next AI chips
Huawei also revealed an updated roadmap for its Ascend AI processors.The company plans to make the Ascend 960DT available in the first quarter of 2027 and the Ascend 960PR in the third quarter of 2027. Huawei said the launches are being brought forward from its previous roadmap. It also plans to introduce the Ascend 970 and Ascend 980 in 2028 and 2029 respectively.
Reuters reported that Huawei currently cannot produce enough AI computing equipment to meet demand in China and is therefore limiting the scale of its international expansion. Huawei's rotating chairman Eric Xu said the company was discussing its Ascend 950DT chips with Chinese AI developers and expected wider adoption next year.
A different approach to AI computing
The growing importance of large computing clusters reflects a broader challenge in AI infrastructure.Training increasingly large AI models requires enormous amounts of processing power. Rather than relying only on improvements to individual processors, Huawei is focusing heavily on connecting large numbers of chips so they can operate together efficiently.
Reuters reported that Huawei's largest planned systems could eventually support up to one million AI processors, while the company has already deployed more than 1,000 earlier AI systems.
Nvidia remains a major competitor, particularly because of its widely adopted CUDA software ecosystem. Huawei's challenge therefore extends beyond chip performance to the broader software and developer ecosystem surrounding AI computing.
Why the announcement matters
Huawei's latest products highlight China's broader effort to expand domestic AI computing capabilities.US export controls have restricted Chinese companies' access to certain advanced AI chips and semiconductor technologies. In response, Chinese technology companies have increased investment in locally developed processors, computing systems and software ecosystems.
For Huawei, the Atlas 960 and its next-generation Ascend roadmap represent an attempt to scale AI infrastructure through both chip development and large interconnected computing systems.
The company says its Ascend ecosystem already has thousands of active developers, while its latest architecture is designed to support increasingly large AI workloads.
Whether Huawei can close the broader gap with Nvidia will depend not only on hardware but also on software, manufacturing capacity, developer adoption and access to global markets.
For now, the Atlas 960 launch shows that the competition to build the infrastructure behind advanced AI is increasingly being fought at the system level, not just at the individual chip level.
Sources