The Pulse
Huawei Expands Ascend Chip Push Against Nvidia
Huawei has introduced the Atlas 960 SuperPoD and outlined a multiyear Ascend chip roadmap at its Shanghai conference. The company is targeting Nvidia with larger AI systems, domestic hardware and new interconnect technology.
AI.info Team ·
Huawei says its next generation of AI infrastructure will connect as many as 1 million processors in a single system, turning a chip shortage into a test of how far Chinese engineers can scale domestic hardware. The company unveiled the Atlas 960 SuperPoD in Shanghai on September 17, advancing a strategy built around linking large numbers of Huawei Ascend processors rather than matching Nvidia only at the level of an individual chip.
The announcement came at Huawei’s annual conference just days before an expected September 24 meeting in Washington between U.S. President Donald Trump and Chinese President Xi Jinping. It also arrives after U.S.-led export controls restricted China’s access to some advanced AI processors and semiconductor manufacturing equipment, pushing Chinese companies toward locally designed alternatives.
Atlas 960 Doubles Down on System Scale
Huawei describes the Atlas 960 SuperPoD as a faster successor to the Atlas 950 system introduced months earlier. The new platform is designed for both AI training and inference, the stage in which a trained model generates predictions or decisions.
Huawei’s published roadmap for the Atlas 960 calls for a system with up to 15,488 Ascend 960 chips. The company says the platform will deliver 30 exaflops of FP8 computing power and 60 exaflops of FP4 computing power, with 4,460 terabytes of memory and 34 petabytes per second of interconnect bandwidth. Huawei lists commercial availability for the fourth quarter of 2027.
The company is also developing larger linked systems. Reuters reported that Huawei executive David Wang said its biggest planned superclusters could support up to 1 million AI processors, while more than 1,000 smaller linked systems had already been shipped to more than 370 customers. Wang did not identify the customers or provide a chip count for each smaller system.
Huawei’s Ascend Roadmap Runs Through 2029
Huawei’s current roadmap extends beyond the Atlas 960. The company plans to introduce the Ascend 970 and Ascend 980 chip series in 2028 and 2029, according to the Associated Press. The roadmap signals a shift from one-off product announcements toward a regular cadence of new processors and increasingly large computing systems.
Huawei’s earlier technical presentation described the Ascend 960 as offering twice the computing power, memory capacity, memory bandwidth and interconnect ports of the Ascend 950 series. The company also says the chip will support its proprietary HiF4 data format, designed for four-bit inference workloads.
Those performance figures are Huawei’s own claims, not independent benchmark results. They also describe complete systems rather than a direct comparison between individual Huawei and Nvidia processors, making software support, workload design and actual customer deployments important parts of the contest.
UnifiedBus Is Huawei’s Answer to Nvidia’s Lead
Huawei is placing as much emphasis on communication between chips as on the chips themselves. Its UnifiedBus interconnect is designed to let thousands of processors share data and operate as one logical machine. The company has said that its larger systems will use combinations of optical networking, Ethernet and its own interconnect protocols.
The approach reflects the limits imposed by export controls. If Chinese companies cannot consistently obtain the most advanced foreign accelerators, they can try to assemble more locally available processors into larger systems. That strategy only works if the chips exchange information quickly and reliably enough to keep the full system busy.
Nvidia still holds a major advantage in software and developer adoption. Chinese AI companies continue to use Nvidia processors, especially for training, even as domestic chips gain ground in inference and other workloads. Reuters reported that Wang put Huawei’s monthly active AI-chip developer base at 5,270, a small ecosystem compared with Nvidia’s global reach.
Washington Reads the Announcement as a Signal
George Chen, a partner and chair of digital practice at consultancy The Asia Group, said the timing carried political weight.
“The timing — less than two weeks before President Xi Jinping’s visit to Washington — underscores Beijing’s confidence and ambition in technology and innovation,”
George Chen, partner and chair of digital practice, The Asia Group, quoted by The Associated Press
Chen also told AP that Huawei’s move from the Atlas 950 to the Atlas 960 in roughly a year could increase concern in the United States about China’s progress in AI computing. “AI developments move so quickly that no one can be certain of holding the lead forever,” he said.
Trump has made the competition explicit. The AP report quoted him saying, “whoever wins in AI, wins,” as he argued that the United States must maintain its lead over China.
The Hard Part Starts After the Launch
Huawei’s announcement establishes an ambitious hardware target, but it does not by itself show that the company can match Nvidia in useful AI output. Huawei must deliver the processors at scale, keep large systems supplied with memory and power, support model developers and persuade customers that its software can replace tools built around Nvidia’s CUDA platform.
The company has already shown that U.S. restrictions can accelerate domestic engineering efforts. The next test is operational: whether Atlas systems can train and serve demanding models reliably, whether Chinese developers will build around Ascend by choice rather than necessity, and whether Huawei can turn very large processor counts into performance customers can measure.