Huawei UnifiedBus Architecture at HUAWEI CONNECT 2026
Analysis based on 11 articles · First reported Sep 21, 2026 · Last updated Sep 21, 2026
The launch signals intensifying competition in AI data-center interconnect and accelerator hardware, pressuring rivals in high-speed networking and AI chip supply chains while highlighting Huawei's push for self-reliant, open compute ecosystems. It reinforces expectations of continued heavy capex in AI infrastructure, a positive for suppliers of optics, memory and advanced packaging but a competitive threat to incumbent interconnect vendors.
At HUAWEI CONNECT 2026 in Shanghai on September 21, 2026, Huawei Executive Director of the Board and CEO of the ICT Business Group Yang Chaobin unveiled new details of the company's UnifiedBus interconnect architecture and announced a series of computing products built on it. Huawei said traditional architectures waste compute as clusters scale, with as little as 20% of a 100,000-NPU cluster's capacity used by models, and that UnifiedBus combines more than ten interconnect protocols to raise interconnect bandwidth from the 100 GB/s level to the TB/s level and cut round-trip latency from 7 microseconds to 2 microseconds. The architecture features unified protocol and memory semantics, heterogeneous compute collaboration interconnecting CPUs, NPUs, memory and SSDs, tiered storage with global pooling, and optoelectronic interconnect for flexible scaling. Huawei announced UnifiedBus LinkBlade for in-cabinet links, LinkDevice supporting 176 ports at 1.6 Tbit/s each for 280 Tbit/s all-optical interconnection between cabinets, and the UBG switch with radix fan-out up to 1,024 to support a million-NPU SuperCluster. It launched a UnifiedBus-powered agentic AI SuperCluster comprising TaiShan 950 and Atlas 960 SuperPoDs, OceanStor M900 memory storage and a Xinghe UBG switch, plus UnifiedBus-based appliances with one to eight NPUs and Kunpeng and Ascend modules for small and medium-sized enterprises. Yang said Ascend is collaborating with open-source frameworks including DeepSeek Harness, OpenCode and openJiuwen on open-source intelligent management, inference acceleration and security agent plugins.
Set up alerts, explore entity relationships, search across thousands of events, and build custom intelligence feeds.
Open Dashboard