Huawei says its Peerium Computing Architecture is designed to let up to one million processors work together as a single computing system. In a roundtable summary distributed on 30 September, the company outlined how the architecture supports its Atlas AI systems and why chip supply is still limiting its expansion overseas.
The announcement centres on data-centre hardware for AI training and inference. Peerium was unveiled at Huawei Connect in Shanghai on 17 September 2026, and Huawei says a 256,000-card Atlas 950 SuperCluster is being deployed. That deployment should be distinguished from the architecture’s claimed million-processor scale: the release does not establish that a complete million-processor system is already operational.
How Huawei wants the processors to work together
Huawei describes three building blocks behind Peerium: nested parallelism, unified memory addressing and peer-to-peer interconnection. Put simply, the design aims to coordinate work and access to memory across a very large collection of processors.
The company calls its parallel-computing approach Nested BSP, short for Nested Bulk Synchronous Parallel. Its UnifiedBus interconnect provides a common open protocol linking CPUs, neural processing units, memory, SSDs, network cards and switches. Huawei says this lets computing, storage and networking components communicate as peers.
Huawei identifies the Atlas 950 SuperPoD and the larger SuperClusters built from it as the first generation of products using Peerium. These are claims about the company’s system architecture; the announcement does not provide independently verified performance results for a million-processor deployment.
Ascend training hardware is still ramping up
At the 17 September roundtable, rotating chairman Eric Xu said the Ascend 950PR was already being supplied primarily for AI inference, although volumes were limited. Systems using the Ascend 950DT, aimed at training, were undergoing testing, with large-scale supply expected at the end of 2026 or early 2027.
Xu said Huawei was discussing the platform with AI model providers and expected more training workloads to use Ascend 950DT-based systems in 2027. Those are Huawei’s expectations, rather than confirmed deployment commitments from customers.
No broad international rollout announced
For markets outside China, supply remains the immediate constraint. Xu said Huawei could not yet meet domestic demand and was not planning a full international expansion, although testing and limited supplies were taking place in some countries.
No Malaysian pricing, customer deployment or local availability date was supplied. For now, Peerium is a look at how Huawei is building larger AI computers by combining processors, memory and interconnects, with the scale and supply of working deployments still the key details to watch.
