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 detail the company’s new Peerium Computing Architecture for the AI era and its UnifiedBus (UB) interconnect. The announcement marks a significant leap in computing scale, as Huawei claims the architecture can make one million processors operate as a single computer, potentially transforming how large AI models are trained and deployed.
The Peerium architecture introduces Nested BSP (Nested Bulk Synchronous Parallel), extending the Turing parallelism paradigm, and features unified memory addressing and peer-to-peer interconnect to achieve strong scaling to the million-processor level. It overruns the traditional von Neumann single-machine architecture and the master-slave design that has dominated for decades. To support this, Huawei invented UnifiedBus, an open high-speed bus that scales without limit to connect CPUs, NPUs, memory, SSDs, interface cards, and switches, enabling peer-to-peer interconnect across compute, storage, and networking.
The first product built on Peerium is the Atlas 950 SuperPoD, powered by the Ascend 950 chip. A SuperCluster with 256,000 computing cards is currently being deployed and tested. Dr. Liao explained that this scale is necessary to train foundation models in the 10 to 40 trillion parameter range, which several Chinese frontier AI labs are targeting over the next two years. He also noted that China’s national power grid plans make 200,000 cards a conservative number for a single data center campus.
Eric Xu highlighted that the Ascend 950PR, primarily for AI inference, has been launched, while the Ascend 950DT for training is under testing with large-scale supply expected by late 2026 or early 2027. He stated that testing results for training are promising and that many AI model trainings will rely on SuperPoDs using the 950DT starting next year. However, he emphasized that supply capacity, not persuasion, will determine adoption. “It’s not about how hard we are going to push model providers, but about how much supply capacity we will have,” Xu said.
On the global AI race, Xu noted that Chinese models are largely open-source and transparent, while US providers have more computing power and may perceive risks differently. He stressed the need to balance AI development with risk management, ensuring “AI for good, not AI for evil.” Regarding international expansion, Xu said Huawei lacks capacity to fully serve markets outside China, though limited testing and supply are ongoing in countries with strong demand.
Xu also claimed that Ascend has surpassed NVIDIA in market share in China, though he acknowledged data collection is difficult. He framed China’s push for chip self-sufficiency as inevitable, stating that even if chips are less advanced, assured supply is crucial. “The path forward is undoubtedly to push for full self-sufficiency in terms of chips and the entire semiconductor value chain,” he said. The full Q&A session is available via the AP Multimedia Newsroom and was distributed by Media OutReach Newswire.
