Huawei Unveils New AI Infrastructure to Power the Agentic World

At HUAWEI CONNECT 2026 in Shanghai, Huawei's Deputy Chairman and Rotating Chairman David Wang delivered a keynote titled "Advancing the Agentic World, Building a Solid Silicon Foundation," outlining the company's strategy to meet the soaring computing demands of the intelligent world. The event, held from September 17 to 19 at the Shanghai World Expo Exhibition & Convention Center and Shanghai Expo Center, showcased Huawei's latest AI infrastructure innovations, which could have significant implications for industries relying on AI.

Wang highlighted that AI is advancing faster than any previous technological revolution. Foundation model parameters are approaching 10 trillion and are projected to exceed 100 trillion by 2030. In China alone, daily inference token consumption has surged to around 500 trillion, expected to reach quintillions (10^18) by 2030. On-device AI is also expanding, with smartphone models growing from three billion parameters in 2024 to 30 billion today, and heading toward hundreds of billions. These trends demand unprecedented scale, performance, and reliability in underlying systems.

To address these challenges, Huawei is focusing on AI infrastructure, particularly SuperPoDs and SuperClusters. A SuperPoD is defined as a computing system where multiple nodes are tightly coupled through high-speed interconnect protocols, featuring unified memory addressing across physical nodes, functioning like a single logical computer. Over 1,000 Atlas 900 A3 SuperPoDs have been deployed, and the Atlas 950 SuperPoD is in large-scale commercial use. According to Huawei's Markov Lab simulations, a 100k-NPU cluster built with 4k-NPU SuperPoDs can deliver a 2.75x increase in Model FLOPs Utilization compared to traditional 8-NPU servers.

Huawei unveiled the Atlas 960E SuperPoD, the industry's first NPO-based SuperPoD. It can scale up to 4,096 NPUs, delivering 8 EFLOPS of FP8 compute performance and up to 1 petabyte of HBM capacity. With 5,500 Hi-ONE units, it eliminates the need for 48,000 800G optical modules, cutting power consumption by over 550 kilowatts and doubling fault-free operating time to achieve 99.8% system availability. The Hi-ONE is the industry's first NPO product ready for mass production, offering 7.2 Tbit/s transmission capacity and a built-in light source. Huawei has submitted an implementation agreement on NPO to the Optical Internetworking Forum.

The upgraded TaiShan 950 SuperPoD supports up to 4,096 nodes with a unified memory pool of up to 256 TB, improving agent performance. For sandbox-intensive workloads, startup speeds for 100,000 sandboxes are 30 times faster than traditional servers, with 25% higher density. The OceanStor M900, a context memory storage cluster, provides petabyte-scale KV caching for agentic inference, extending SSD read/write lifespan by 16-fold. The new agentic SuperCluster uses UnifiedBus to consolidate multiple interconnect protocols, enabling peer-to-peer interconnect between subsystems and scaling up to one million NPUs.

Huawei is also committed to open ecosystems. The Kunpeng ecosystem has attracted over 4.16 million developers and more than 7,200 partners, supporting over 560 open-source projects. CANN has moved to community-driven open-source development, with external developers now comprising 61% of all CANN developers. Ascend supports over 90 leading third-party open-source projects, including PyTorch, Triton, vLLM, and veRL, and is the first official Chinese compute platform on PyTorch's website.

For more information, visit HUAWEI CONNECT 2026 online.