Huawei didn’t come to the 2026 Connect conference with incremental updates. The company announced the Ascend 960 SuperPoD, an ultra-large-scale AI computing cluster designed to operate as a single logical machine, alongside two supporting technologies: Near-Packaged Optics (NPO) and an upgraded UnifiedBus interconnect. Taken together, this is Huawei’s most serious attempt yet to position itself as a full-stack alternative to Nvidia in AI infrastructure.
The Ascend 960 SuperPoD matters because of what it represents structurally, not just technically. Nvidia’s dominance in AI compute isn’t just about the H100 or B200 chips in isolation. It’s about the full system: the NVLink interconnects, the NVSwitch fabric, the software stack, and the ability to pool thousands of GPUs into one coherent training environment. Huawei is clearly targeting that same systems-level thinking. A cluster that behaves as a single machine is exactly the architecture serious AI training workloads demand.
The addition of Near-Packaged Optics is worth paying attention to. Optical interconnects are becoming a real focus across the industry because copper just can’t move data fast enough at the distances and densities modern AI clusters require. Nvidia, Broadcom, and a handful of startups including Ayar Labs and Celestial AI are all working in this space. Huawei entering with NPO signals it understands where the bottleneck actually is: it’s rarely the chip, it’s the interconnect.
So who is this for? Primarily Chinese cloud providers, state-backed AI research institutions, and enterprises that are either cut off from Nvidia hardware due to US export controls or actively looking to reduce that dependency. Since the Biden-era restrictions on exporting advanced chips to China, there has been enormous pressure on Chinese tech companies to build domestic alternatives. Huawei’s Ascend line is the most credible response to that pressure so far.
The key unknowns are software maturity and real-world performance at scale. Nvidia’s moat isn’t just silicon, it’s CUDA and a decade of developer tooling built on top of it. Huawei’s CANN framework is improving, but adoption outside China remains almost zero. For global AI developers, that’s a real barrier.
Still, dismissing this as irrelevant would be a mistake. China’s AI infrastructure market is enormous, and Huawei is increasingly capable of serving it end to end, chips, interconnects, optics, and cluster management included. That’s a meaningful shift in how the global AI hardware market is splitting apart.
