- New system links up to 4,096 AI chips in a single supernode
- Huawei says its optical interconnect technology could enable AI clusters of up to 1 million chips
Huawei has unveiled the Ascend 960 SuperPoD, a supernode designed to link thousands of AI processors and support large-scale training and inference, as China seeks to expand AI computing capacity through system-level innovation.
Huawei Deputy Chairman and Rotating and Acting Chair David Wang Tao (汪涛) announced the system on September 17 at Huawei Connect 2026 in Shanghai.
The company said the Ascend 960 is the industry’s first supernode to use near-package optics (NPO), a technology designed to reduce the power and bandwidth costs of connecting large numbers of processors.
A supernode essentially connects thousands of computing cards through high-speed links so they can work together as a single computing system.
The UnifiedBus protocol
Huawei said communication can account for more than 40% of training time in a 100,000-chip cluster under conventional architectures, as processors spend much of their time waiting for data.
Huawei’s answer is its self-developed UnifiedBus interconnect protocol, which it says can reduce communication latency to 2 microseconds.
The Ascend 960 SuperPoD can accommodate up to 4,096 computing cards, delivering 8 exaflops of FP8 computing power and up to 1 petabyte of HBM memory, according to Huawei.
Optics take center stage
The system’s key change is its NPO optical interconnect technology.
Huawei says 5,500 Hi-ONE optical engines can replace about 48,000 800G optical modules, cutting power consumption by more than 550 kilowatts and raising system availability to 99.8%.
Huawei plans to connect multiple Ascend 960 SuperPoDs through its UnifiedBus network.
The company says the resulting Ascend SuperCluster can scale to 512,000 chips, while a multi-rail topology could support clusters of up to 1 million chips.
A chip family
Wang also said development of the Ascend 960 chip family is ahead of schedule.
The 960DT is now expected to be ready in the first quarter of 2027, three quarters earlier than initially planned, while the 960PR is scheduled for the third quarter, one quarter ahead of plan.
More than 1,000 Ascend 910C supernodes have already been deployed, while the Ascend 950 Atlas 950 SuperPoD has entered commercial deployment, Huawei said.
“Computing power is at the core of Huawei’s AI strategy,” Wang said. “We will focus on AI infrastructure and pursue system-level innovation around supernodes and clusters.”
China pushes beyond the chip
For the global AI industry, the significance of the Ascend 960 lies less in any single processor than in Huawei’s attempt to solve the scaling problem at the system level.
As restrictions on access to advanced chips constrain China’s ability to compete solely through semiconductor performance, Huawei is investing in faster interconnects, optical technology and large-scale system architecture to link thousands of processors into a single computing platform.
Whether Huawei’s million-chip vision can translate into reliable commercial deployments remains a separate test.
But the launch underscores how the AI infrastructure race is expanding beyond individual chips to the networks that connect them.
Header image credit: Huawei
