Huawei unveiled the Peerium computing architecture for the AI era, introducing Nested BSP to break through von Neumann single-machine architecture limits and enabling millions of processors to function as a single computer via the Lingqu interconnect technology.

SHANGHAI, CHINA – SEPTEMBER 17, 2026 | Huawei today announced Peerium, a new computing architecture engineered for the artificial intelligence era. Designed to meet rapidly expanding AI compute demands, Peerium enables millions of processors to operate as a single, unified computer.

Peerium is founded on nested parallelism, unified memory addressing, and fully equal interconnects to achieve massive scalability. The architecture breaks from the traditional Turing paradigm by introducing Nested BSP, effectively transcending the von Neumann single-machine model and dismantling the long-standing master-slave hierarchy. This fundamental shift allows million-scale processor arrays to function cohesively as one expansive computational entity.

Lingqu serves as the critical interconnect technology powering the Peerium architecture. Operating on an open protocol, Lingqu functions as a high-speed bus capable of infinite scaling to seamlessly link CPUs, NPUs, memory, SSDs, network interface cards, and switches. By establishing truly equal interconnection across compute, storage, and networking resources, Lingqu ensures uniform data flow and eliminates traditional architectural bottlenecks.

The first-generation implementation of the Peerium architecture is the Atlas 950 super node. A cluster comprising 256,000 Atlas 950 cards is currently in deployment, while the NPO-based Atlas 960 system is under active testing.

“In the AI era, Huawei will continue to develop super nodes and clusters tailored to customers’ training and inference requirements based on our pioneering Peerium computing architecture,” said Xu Zhijun, Huawei’s rotating chairman. “Our goal is to make compute power ubiquitous and intelligence omnipresent.”