At HUAWEI CONNECT on September 17, 2026, Huawei Rotating Chairman Eric Xu and Dr. Liao Heng, Chief Scientist of HiSilicon, held a roundtable with international media to discuss the company's Peerium Computing Architecture and UnifiedBus (UB). The announcement matters because it signals a fundamental shift in how massive computing systems are built, moving beyond the master-slave paradigm that has dominated for decades. By enabling one million processors to operate as a single computer, Huawei aims to address the escalating demands of training frontier AI models with tens of trillions of parameters.
The Peerium architecture achieves strong scaling through nested parallelism, unified memory addressing, and peer-to-peer interconnect. It introduces Nested Bulk Synchronous Parallel (Nested BSP), extending the Turing parallelism paradigm, and extends the von Neumann single-machine architecture. Because existing industry technologies could not support this design, Huawei invented UnifiedBus, a high-speed bus using an open protocol that scales without limit to connect CPUs, NPUs, memory, SSDs, interface cards, and switches. UnifiedBus enables peer-to-peer interconnect across compute, storage, and networking, fundamentally overrunning the traditional master-slave architecture. The Atlas 950 SuperPoD is the first product built on Peerium, and a SuperCluster with 256,000 computing cards is currently being deployed and tested.
The implications are significant for AI training infrastructure. Dr. Liao explained that 256,000 cards is not arbitrary: it reflects the memory capacity and scale needed to train foundation models expected to reach 10 to 40 trillion parameters within two years. He noted that six or seven frontier AI labs in China will pursue such models, and that 200,000 is a conservative number given power grid designs in single autonomous zones and data center campuses. For the Ascend 950, Eric Xu said the 950PR is already available primarily for inference, while SuperPoDs based on the Ascend 950DT for training are under testing, with large-scale supply starting late this year or early next year. He added that testing results for training are promising and that many AI model training workloads will likely rely on these SuperPoDs next year, though supply capacity remains the limiting factor.
On international expansion, Xu stated that because Huawei cannot meet domestic demand, there are no fully-fledged plans to enter other markets, though limited testing and supply are occurring in countries with strong demand. He also claimed that Ascend has surpassed NVIDIA in China based on collected data. Regarding calls by leading US AI model providers to slow AI development, Xu observed that Chinese models are largely open-source and transparent, while US providers have more computing power and may better understand their own sophistication. He said the sense of AI risk might be weaker in China but that a balance between driving development and managing risks is essential, with the bottom line being AI for good, not evil. On chip self-sufficiency, Xu called it an inevitable path for China, emphasizing assured supply even if chips are less advanced, and predicted full self-sufficiency in chips and the semiconductor value chain will become reality sooner or later.
The full Q&A session is available through the AP Multimedia Newsroom. The announcement was distributed via Media OutReach Newswire.
