Huawei unveiled the Atlas 960E SuperPoD, designed for artificial intelligence models with up to 10 trillion parameters, at HUAWEI CONNECT 2026 on September 17. The system brings together 4,096 artificial intelligence processors and uses NPO optical interconnect technology for data transfer between processors. Data is transported by light rather than electricity; the Atlas 960E contains approximately 5,500 Hi-ONE optical units. According to Huawei, a traditional architecture could require approximately 48 thousand optical modules for a similar interconnect. By connecting multiple systems, the infrastructure could scale to 512 thousand processors, and to up to 1 million artificial intelligence processors with a different connection arrangement. The liquid-cooled system is said to be capable of saving more than 550 kilowatts of energy through changes to its interconnects. The Atlas 960E SuperPoD is planned to be launched in the third quarter of 2027.
Why it matters
This development shows that not only the number of processors but also the flow of data between processors has become a determining factor in artificial intelligence infrastructure. The transport of data by light and the difference in the number of optical units stand out as factors that could affect the cost and energy use of interconnect infrastructure for very large models. The scalable architecture aims to make it possible to expand the infrastructure in different configurations for needs that exceed the capacity of a single system. For this reason, the announcement is important for data centers running large artificial intelligence models and for those concerned with the energy consumption of these centers. However, how the announced energy savings and scaling capacity will perform in practice can be assessed when the product is launched.
Background
Huawei is not a new name in the FikirPilot archive: we have published 5 news articles mentioning this name in the last 90 days; the latest is dated September 15, 2026.