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 new Peerium Computing Architecture and UnifiedBus (UB) technology. The announcement matters because it signals a fundamental shift in how massive AI computing systems are designed, moving away from the master-slave paradigm that has dominated for decades.
The Peerium architecture enables one million processors to function as a single computer through nested parallelism, unified memory addressing, and peer-to-peer interconnect. It extends the Turing parallelism paradigm with Nested Bulk Synchronous Parallel (Nested BSP) and overruns the von Neumann single-machine architecture. To support this, 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. This achieves peer-to-peer interconnect across compute, storage, and networking, fundamentally overturning traditional master-slave designs.
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 the 256,000-card scale is not arbitrary: it supports training foundation models already in the 5-trillion-parameter range, and over the next two years, six or seven frontier AI labs in China will aim for models of 10 to 40 trillion parameters. The scale also aligns with China's national power grid plans for data centers, where 200,000 cards is a relatively conservative number for a single autonomous zone and campus.
Eric Xu noted that the Ascend 950PR, mainly for AI inference, is already launched but supply is limited. The Ascend 950DT, used for training in SuperPoDs, is under testing with large-scale supply expected by late 2026 or early 2027. He said testing results for training are "pretty good" and predicted that starting next year, much AI model training will be based on Ascend 950DT SuperPoDs. However, he emphasized that the constraint is supply capacity, not persuasion of model providers, and hoped the value chain would expand faster.
On AI risks, Xu observed that Chinese models are largely open-source and transparent, while US model providers have more computing power and perhaps a stronger sense of risk. He believes China needs to accelerate its pace to reach a level where it also feels those risks, but always balancing development with risk management, ensuring "AI for good, not AI for evil."
Regarding international expansion, Xu said Huawei lacks capacity to fully satisfy demand in China, so it has no plans for a full-scale international rollout, though limited testing and supply occur in countries with strong demand. He added that based on collected data, Ascend has surpassed NVIDIA in China, though exact market share data is hard to gather. On chip self-sufficiency, he called it an inevitable path for China, a country of 1.4 billion and an industrial powerhouse, to ensure supply security and push for full self-sufficiency across the semiconductor value chain.
The full Q&A session is available at https://www.apmultimedianewsroom.com/multimedia-newsroom/huawei-pioneers-a-new-computing-architecture-for-the-ai-era-making-one-million-processors-work-as-one-computer-2, and the announcement was distributed via Media OutReach Newswire.
