At HUAWEI CONNECT on September 17, 2026, Huawei's Rotating Chairman Eric Xu and Dr. Liao Heng, Chief Scientist of HiSilicon, detailed the company's Peerium Computing Architecture and UnifiedBus (UB) technology during a roundtable with international media. The announcement, summarized by Media OutReach Newswire, marks a significant step in computing architecture for the AI era, enabling one million processors to operate as a single computer.
Huawei introduced the Ascend 950 chip and the Atlas 950 SuperPoD, which is built on the Peerium architecture. Peerium 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 BSP (Nested Bulk Synchronous Parallel) and overruns the traditional von Neumann single-machine architecture and master-slave design. The key interconnect technology, UnifiedBus, is an open-protocol high-speed bus that connects CPUs, NPUs, memory, SSDs, interface cards, and switches without limit, enabling peer-to-peer interconnect across compute, storage, and networking.
Dr. Liao Heng explained that a SuperCluster with 256,000 computing cards is currently being deployed and tested. He noted that this scale is necessary for training foundation models in the 5 trillion parameter range, and that over the next two years, six or seven frontier AI labs in China will aim to train models of 10 to 40 trillion parameters. The 200,000-card number is also a conservative fit for China's power grid design in a single data center campus.
Eric Xu stated that the Ascend 950PR is already available for AI inference, while SuperPoDs based on the Ascend 950DT for AI training are under testing, with large-scale supply expected by the end of 2026 or early 2027. He added that testing results for training with the 950DT are promising, and that many AI model training workloads will likely shift to these SuperPoDs next year. However, supply capacity remains a constraint, and Huawei is in extensive dialogue with AI model providers.
On international expansion, Xu said Huawei has no plans for a full-scale push outside China due to insufficient capacity to meet domestic demand. Limited testing and supply are occurring in a few countries with strong demand. He also claimed that Ascend has surpassed NVIDIA in market share in China, based on collected data. Regarding US calls to slow AI development, Xu emphasized the need to balance AI progress and risk management, ensuring AI for good. He also underscored China's push for chip self-sufficiency as an inevitable path, given the country's size and industrial needs.
The full Q&A session is available at this link. This development could significantly impact the AI industry by enabling the training of ultra-large models, reducing reliance on traditional architectures, and intensifying the global AI race. It also highlights China's accelerating efforts toward semiconductor self-sufficiency.
