HUAWEI has lifted the curtain on its most ambitious AI hardware plans to date. Speaking at its annual Huawei Connect conference in Shanghai, the company introduced SuperPoD Interconnect, a system designed to link together as many as 15,000 graphics processors, including its in-house Ascend AI chips. The approach mirrors Nvidia’s NVLink, which allows high-speed communication across clusters of graphic processing units (GPUs), but is tailored to Huawei’s hardware ecosystem. While individual Ascend chips lag behind Nvidia’s top-performing H100 and B200 accelerators in raw power, Huawei’s bet is that scale and clustering can compensate, providing customers with access to massive pools of compute needed to train and deploy next-generation AI systems. The announcement comes on the heels of Chinese regulators ordering big domestic tech firms to halt purchases of Nvidia chips, including the RTX Pro 600D servers that had been specifically customized for the Chinese market. That decision, coupled with ongoing U.S. restrictions on high-end AI semiconductors, has increased the urgency for China to build its own alternatives. Rotating chairman Eric Xu revealed Thursday that the company will follow a one-year release cycle for its Ascend chips, promising to double compute power with each iteration. The Ascend 910C, released earlier this year, will be followed by the Ascend 950 in 2026, and then by the 960 in 2027 and 970 in 2028. In parallel, Huawei is building large-scale “supernodes” — powerful data center systems that group thousands of chips into unified clusters. The Atlas 950, set for release in late 2026, will support over 8,000 Ascend chips, while the Atlas 960 in 2027 will scale up to nearly 15,500 chips. These supernodes are designed to succeed the Atlas 900, a smaller system using 384 Ascend 910C chips that Huawei had previously positioned as one of the world’s most powerful AI training clusters. Xu said the upcoming systems would exceed all rivals “across major metrics.” Nvidia remains far ahead in raw chip performance and in the breadth of its software ecosystem, particularly with CUDA, the programming framework that has become the industry standard. Huawei will need to convince developers and enterprises to adopt its Ascend AI stack, which competes with CUDA through its own CANN (Compute Architecture for Neural Networks) toolkit. Huawei also faces steep challenges such as lacking access to cutting-edge lithography, which is required for manufacturing advanced GPUs. Yet Huawei may have an opening: Chinese firms cut off from Nvidia hardware have little choice but to turn to domestic alternatives. (SD-Agencies) | 