Huawei recently unveiled its groundbreaking SuperPoD AI infrastructure architecture at HUAWEI CONNECT 2025, a development poised to redefine the construction and scaling of artificial intelligence systems globally. Moving beyond traditional independent server approaches, the SuperPoD technology allows thousands of scattered AI chips across dozens of server cabinets to function cohesively as a single, massive logical machine. This innovative system enables unified learning, thinking, and reasoning capabilities across myriad processing units, marking a significant shift in how AI computing power is organized, scaled, and deployed across various industries.

At the heart of Huawei's revolutionary infrastructure is UnifiedBus (UB) 2.0, an advanced interconnect protocol. Yang Chaobin, Huawei’s Director of the Board and CEO of the ICT Business Group, highlighted that the SuperPoD architecture, built upon UnifiedBus, deeply interconnects physical servers, allowing them to operate like a single logical entity. This protocol effectively addresses historical limitations in large-scale AI computing, particularly concerning the reliability of long-range communications and bandwidth-latency challenges. Traditional copper connections offer high bandwidth over short distances, while optical cables, though supporting longer ranges, have suffered from reliability issues that intensify with increased distance and scale.

Eric Xu, Huawei’s Deputy Chairman and Rotating Chairman, emphasized the criticality of solving these fundamental connectivity challenges for the company's AI infrastructure strategy. He detailed the breakthrough solutions embedded within the OSI model, explaining that reliability has been integrated into every layer of the interconnect protocol—from the physical and data link layers up to the network and transmission layers. This includes 100-nanosecond-level fault detection and protection switching on optical paths, ensuring that any intermittent disconnections or faults are imperceptible at the application layer, thus guaranteeing robust system operation.

The Atlas 950 SuperPoD stands as the flagship implementation of this architecture, integrating up to 8,192 Ascend 950DT chips. Xu revealed that this configuration delivers 8 EFLOPS in FP8 and 16 EFLOPS in FP4, alongside an interconnect bandwidth of 16 PB/s, which he noted is more than ten times higher than the entire globe’s total peak internet bandwidth. Occupying 160 cabinets across 1,000m², with 128 compute and 32 communication cabinets linked by all-optical interconnects, the system boasts a memory capacity of 1,152 TB and maintains a remarkable 2.1-microsecond latency across the entire system. Future developments include the Atlas 960 SuperPoD, slated to incorporate 15,488 Ascend 960 chips within 220 cabinets covering 2,200m², promising 30 EFLOPS in FP8 and 60 EFLOPS in FP4, with 4,460 TB of memory and 34 PB/s interconnect bandwidth.

Beyond AI, the SuperPoD concept extends to general-purpose computing through the TaiShan 950 SuperPoD, powered by Kunpeng 950 processors. This system is designed to address enterprise needs for replacing legacy mainframes and mid-range computers, particularly in the finance sector. Xu articulated that the TaiShan 950 SuperPoD, combined with the distributed GaussDB, offers an ideal and permanent alternative to mainframes, mid-range computers, and Oracle’s Exadata database servers.

A significant aspect of Huawei’s strategy is the release of UnifiedBus 2.0 technical specifications as open standards. This decision acknowledges the existing constraints in semiconductor manufacturing process nodes within the Chinese mainland and aims to foster a sustainable computing power ecosystem using practically available process nodes. Yang emphasized Huawei's commitment to an open-hardware and open-source-software approach, which encourages partners to develop industry-specific SuperPoD solutions, thereby accelerating developer innovation and nurturing a vibrant ecosystem. Huawei plans to open-source various hardware components, including NPU modules, blade servers (air-cooled and liquid-cooled), AI cards, CPU boards, and cascade cards. On the software front, the company has committed to fully open-sourcing CANN compiler tools, Mind series application kits, and openPangu foundation models by December 31, 2025.

The market deployment validates these technical claims, with over 300 Atlas 900 A3 SuperPoD units already shipped in 2025. These have been deployed for more than 20 customers across diverse sectors such as the Internet, finance, carrier services, electricity, and manufacturing. This initiative holds substantial implications for the development of China’s AI infrastructure, creating an open ecosystem around domestic technology that navigates challenges posed by constrained semiconductor manufacturing. Huawei's approach facilitates broader industry participation in AI infrastructure development, reducing reliance on the most advanced process nodes. Globally, Huawei’s open architecture strategy introduces a compelling alternative to the proprietary, tightly integrated hardware and software models prevalent among Western competitors. While the long-term performance and commercial viability of this open ecosystem at scale remain to be fully demonstrated, the SuperPoD architecture represents more than just an incremental advance; it proposes a fundamental rethinking of how massive computational resources are connected, managed, and scaled. The open-source release of its specifications and elements will be a critical test of whether collaborative development can accelerate AI infrastructure innovation within a partner ecosystem, potentially reshaping competitive dynamics in the global AI infrastructure market.