Huawei has introduced a computing architecture designed to make one million processors operate as a single computer, a scale that could redefine the infrastructure required for training the next generation of artificial intelligence models. The Peerium Computing Architecture, announced at HUAWEI CONNECT on September 17, 2026, achieves strong scaling to the million-processor level through nested parallelism, unified memory addressing, and peer-to-peer interconnect. It extends the parallelism paradigm of Turing with Nested Bulk Synchronous Parallel (Nested BSP) and overruns the von Neumann single-machine architecture, displacing the master-slave design paradigm that has prevailed for decades.
Because available industry technologies could not support this architecture, 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 overturning the traditional master-slave approach. 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.
The scale is not arbitrary. Dr. Liao Heng, Chief Scientist of HiSilicon, explained that large clusters are needed for training foundation models already in the 5-trillion-parameter range. Over the next two years, roughly six or seven frontier AI labs in China will aim to train models of 10 to 40 trillion parameters, matching the memory capacity and size of such SuperPoDs. Additionally, China's national power grid plan makes 200,000 processors a relatively conservative number for a single autonomous zone and data center campus.
Eric Xu, Huawei's Rotating Chairman, said the Ascend 950PR is already on the market primarily for AI inference, though supply volume is limited. SuperPoDs based on the Ascend 950DT for AI training are under testing, with large-scale supply expected at the end of this year or early next year. Xu noted that testing results for training with the Ascend 950DT are good, and he believes many AI model training workloads will run on these SuperPoDs starting next year. He emphasized that supply capacity, not persuasion, will determine adoption, and he hopes the value chain expands capacity faster.
On the global AI race, Xu observed that Chinese models are largely open-source and transparent, while leading U.S. providers have more computing power and may better understand their own model sophistication. He said the market feeling about AI risks might be weaker in China, and Chinese providers need to accelerate to reach a level where they feel those risks. Still, he stressed balancing AI development with risk management, with the bottom line being AI for good, not AI for evil.
Regarding international expansion, Xu stated that because Huawei cannot satisfy 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 added that based on collected data, Ascend has surpassed NVIDIA in China's market share. On chip self-sufficiency, Xu called it an inevitable path forward for a country of 1.4 billion people and an industrial powerhouse. Even if chips are less advanced, assured supply removes daily uncertainty, and he believes full self-sufficiency across the semiconductor value chain will become reality sooner or later.
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. The announcement was distributed via Media OutReach Newswire.
