Huawei used its Connect 2025 event to unveil SuperPod, a new approach to cloud AI infrastructure aimed at overcoming one of the industry’s most persistent challenges: scaling without performance loss. Rather than treating servers as separate machines that communicate through conventional networking, SuperPod allows thousands of processors to operate as if they were part of a single, unified system.
The design relies on Huawei’s UnifiedBus interconnect protocol, which links hardware components at a deep level. Company executives explained that this reduces the inefficiencies that typically emerge in massive clusters, where more servers often bring more complexity instead of more usable computing power. By tightening the connections between processors, SuperPod delivers consistent performance as deployments expand.
At the top end, the Atlas 950 SuperPod connects over 8,000 Ascend processors with interconnect bandwidth exceeding 16 PB per second. Huawei compared this to more than ten times the world’s peak internet traffic. Huawei plans to release a larger version, the Atlas 960, in 2027, scaling performance further and targeting workloads that may reach trillions of AI model parameters.
The company also introduced enterprise-oriented versions designed for more conventional data centers. For example, an Atlas 850 can run in an air-cooled facility, allowing organizations to avoid expensive liquid cooling systems in the near term.
One of the most conspicuous facets of Huawei’s tactic is, perhaps, its verdict to throw open the UnifiedBus 2.0 specs to the greater market. By making available the technical standards as well as the hardware designs by the end of 2025, Huawei’s intention is to get the partners and developers to tailor the architecture for the different industry scenarios.
Huawei has already shipped more than 300 SuperPod units this year to sectors including finance, energy, and manufacturing. Whether SuperPod will become a dominant model remains uncertain, but the announcement highlights the company’s bet that architectural innovation, not just processor advances, will shape the next phase of cloud AI infrastructure.
