At Huawei Connect 2025, the company rolled out a three-year roadmap for its Ascend chips alongside the debut of its Atlas 950 supernode. Rotating chairman Eric Xu warned that China’s AI future can no longer hinge on single-chip performance, constrained as it is by advanced process limits. The way forward, he argued, is a “supernode + cluster” architecture — Huawei’s bet on system-level design to secure scalable and sustainable compute.