Huawei is significantly expanding its Ascend AI platform with new systems capable of handling AI models with up to 10 trillion parameters—far larger than most current systems. This expansion includes the Atlas 960 SuperPoD, a high-performance computing system that can house up to 15,488 Ascend 960 processors. The system is designed to deliver massive computing power, with 30 EFLOPS of FP8 computing and 60 EFLOPS using FP4, along with 4,460 TB of memory capacity. Its interconnect bandwidth reaches 34 PB/s, and Huawei expects it to achieve a training performance of 15.9 million tokens per second, which is crucial for training complex AI models. At the heart of this advancement is Huawei’s UnifiedBus, an interconnect architecture that efficiently links processors, memory, and storage in large AI systems. UnifiedBus is used in the Atlas 960E, a version of the SuperPoD that can connect up to 4,096 NPUs (neural processing units) and provide them with shared memory access. This system integrates UnifiedBus with Huawei’s Hi-ONE optical technology, which uses near-packaged optics (NPOs) to place optical components closer to processors. This design improves data transmission speeds and reduces energy use, allowing the system to move data between thousands of processors more efficiently than traditional methods. The Atlas 960E uses around 5,500 Hi-ONE units instead of 48,000 conventional 800G optical modules, significantly reducing power consumption by over 550 kilowatts in the interconnection system. This configuration provides the system with 8 EFLOPS of FP8 computing performance and supports AI models with up to 10 trillion parameters. Additionally, Atlas 960E systems can be connected into larger SuperClusters containing 512,000 NPUs, with the architecture eventually supporting configurations with millions of processors. Huawei is also accelerating the development of its Ascend 960 family of AI chips. The Ascend 960DT is scheduled for release in Q1 2027, and the 960PR is planned for Q3 2027—both arriving earlier than previously expected. Huawei reports that performance improvements have surpassed initial projections. David Wang, deputy chairman and rotating chairman of Huawei, noted that the company is now updating its Ascend chip series annually. Future releases, including the Ascend 970 and 980 chips in 2028 and 2029, are expected to follow the Tau (τ) Scaling Law, which predicts continued performance improvements and significant advancements in computing specifications.