Huawei has unveiled a computing architecture that it says can make one million processors operate as a single computer, a development that could significantly alter the economics and scale of artificial intelligence training and inference. The Peerium Computing Architecture, introduced at HUAWEI CONNECT on September 17, 2026, underpins the new Atlas 950 SuperPoD, which is built on the Ascend 950 chip. According to a roundtable discussion between Huawei Rotating Chairman Eric Xu and Dr. Liao Heng, Chief Scientist of HiSilicon, the architecture 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 and departs from the master-slave design that has dominated computing for decades.
The implications for AI development are substantial. Training frontier foundation models with tens of trillions of parameters requires massive, tightly coupled computing infrastructure. Huawei's approach aims to provide that scale within a single system image, reducing the complexity of managing sprawling clusters. Central to this is UnifiedBus, an open-protocol high-speed bus that connects CPUs, NPUs, memory, SSDs, interface cards, and switches without a scaling limit. By enabling peer-to-peer interconnect across compute, storage, and networking, UnifiedBus fundamentally overruns traditional hierarchical designs. A SuperCluster with 256,000 computing cards is currently being deployed and tested.
Dr. Liao explained that the 256,000-card figure is not arbitrary. It aligns with the memory capacity needed to train 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 power delivery constraints in Chinese data centers make 200,000 cards a conservative estimate for a single autonomous zone. Eric Xu added that the Ascend 950PR, focused on inference, is already shipping in limited volumes, while SuperPoDs based on the Ascend 950DT for training are under testing and will see large-scale supply starting late this year or early next year. He indicated that testing results for training are promising and that widespread adoption by Chinese model providers is expected next year, constrained mainly by supply capacity.
On the global stage, Xu said Huawei does not plan a fully-fledged international expansion because it cannot meet domestic demand. Still, several countries with strong demand are receiving limited supply for testing. He also claimed that Ascend has surpassed NVIDIA in China's market, though precise data is difficult to collect. Regarding calls by some U.S. AI model providers to slow development, Xu noted that Chinese models are largely open-source and transparent, and that the market perception of AI risk is weaker in China. He emphasized the need to balance development and risk management, with the bottom line being AI for good, not evil.
China's push for chip self-sufficiency, Xu said, is an inevitable path forward. He argued that as a country of 1.4 billion people and an industrial powerhouse, China will not accept a future in which others decide its access to chips. Even if its chips are less advanced, assured supply is critical. For Huawei and the domestic industry, full self-sufficiency across the semiconductor value chain is the goal, and Xu believes it 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.
