Huawei unveils new AI SuperPoD and chip roadmap as Gulf AI push accelerates

Huawei has unveiled a new generation of AI computing technology, including what it calls the industry's first SuperPoD built on near-packaged optics (NPO), as the Middle East continues to scale up investment in AI infrastructure.

The announcements were made at Huawei Connect 2026 in Shanghai on 17 September. They include the Atlas 960E SuperPoD and Hi-ONE, an optical engine Huawei describes as the first ready for mass production with a built-in light source.

Huawei also said development of its next-generation Ascend 960 AI chips is running ahead of schedule, as the company moves towards releasing a new chip generation every year.

According to Huawei, the technology is designed to meet the demands of increasingly large AI models. The company said foundation models are approaching 10 trillion parameters, and it expects that figure to exceed 100 trillion by 2030.

The tech company noted that governments and businesses across the Middle East are moving from AI experimentation to deployment at scale, in line with national strategies, driving demand for computing capacity, energy efficiency and reliability.

Atlas 960E and Hi-ONE

Huawei said the Atlas 960E SuperPoD, powered by its UnifiedBus architecture and the new Hi-ONE optical engine, can scale to 4,096 NPUs (neural processing units). The company said this delivers 8 exaflops of FP8 computing performance and 16 exaflops at FP4 precision, with up to one petabyte of high-bandwidth memory capacity. It is not clear from public materials how the newly announced Atlas 960E relates to the larger Atlas 960 SuperPoD Huawei described in 2025.

The system uses 5,500 Hi-ONE units, which Huawei said removes the need for around 48,000 conventional 800G optical modules. According to the company, this cuts power consumption by more than 550 kilowatts, doubles fault-free operating time and achieves 99.8% system availability.

Huawei said Hi-ONE reduces power consumption in high-speed optical interconnects by around 66% and cuts signal transmission latency by about 90%, while its integrated design is claimed to improve reliability tenfold.

David Wang, deputy chairman of the board and rotating chairman at Huawei, said: “AI is advancing faster than any technological revolution we have seen before, and the demands it places on computing infrastructure are growing just as quickly. While models become larger and more complex, we need to rethink how computing systems are designed, connected and scaled.”

He added: “At Huawei, we are addressing this through continued innovation in systems and architecture, from chips and interconnects to SuperPoDs and SuperClusters, while making computing power more readily accessible. No single company can build an intelligent world alone, which is why we will continue to work with customers, partners and developers to build the foundations needed for this next stage of AI.”

Chip roadmap moved forward

Huawei also gave an update on its Ascend AI chip roadmap. The company said the Ascend 960DT chip is expected in the first quarter of 2027, three quarters ahead of its original roadmap, followed by the Ascend 960PR in the third quarter of 2027, one quarter ahead of schedule.

According to Huawei, the company is moving to a one-generation-per-year development cycle, with the Ascend 970 planned for 2028 and the Ascend 980 for 2029. Huawei said it expects compute specifications to keep doubling with each generation, though this is the company projection rather than a verified outcome.

“We're evolving our Ascend chip series on a one-generation-a-year cycle,” Wang said. “Thanks to the Tau (τ) Scaling Law, not only will their compute specifications continue to double, but you can also expect to see huge improvements across the board in terms of memory bandwidth, memory capacity, interconnect bandwidth, and more.”

Scaling beyond a single SuperPoD

Huawei said multiple Atlas 960E systems can be linked into SuperClusters of up to one million NPUs. The company also unveiled an upgraded TaiShan 950 SuperPoD and the OceanStor M900 context memory storage cluster, extending its infrastructure range across computing and storage.

Ascend now supports more than 90 third-party open-source projects, while its Kunpeng ecosystem has grown to more than 4.16 million developers and 7,200 partners globally, according to Huawei's press release. The company added that CANN, the software foundation of the Ascend ecosystem, has moved to sustained, community-driven open-source development.

Huawei Connect 2026 ran until 19 September in Shanghai under the theme “Advancing the Agentic World.”