Overview

• At Huawei Connect, rotating chairman Eric Xu unveiled UnifiedBus, a high-speed interconnect the company says can link up to 15,488 Ascend processors per SuperPod and scale to roughly 1 million accelerator cards in super clusters.
• Huawei claims UnifiedBus can move data up to 62 times faster than Nvidia’s upcoming NVLink144, positioning networking throughput as the core lever for large-scale AI training.
• The company published an Ascend roadmap with two Ascend 950 variants in 2026, an Ascend 960 in 2027 with roughly double the performance, and an Ascend 970 in 2028 featuring a 4 Tbit/s interface.
• Huawei acknowledged its single-chip performance trails Nvidia and is betting on brute-force clustering, self-designed high-bandwidth memory, and its networking expertise to deliver competitive aggregate compute.
• Analysts highlighted persistent constraints, with Bernstein estimating a next-generation Ascend 950 at about 6% of Nvidia’s forthcoming VR200 performance and Jefferies flagging manufacturing limits tied to advanced-node tools and yields.