Huawei launches the Atlas 900 A3 SuperPoD

Why interconnect became the focus

Training and serving large AI models depends not only on how fast a single chip is, but on how quickly many chips can exchange data. When hundreds of processors work on one problem, the links between them can become the bottleneck: the machine spends time waiting for data instead of computing. A "super node" (SuperPoD) is an approach that treats a large group of chips as one machine rather than as a cluster of separate servers networked together.

What Huawei announced

Huawei officially launched the Atlas 900 A3 SuperPoD in March 2025. The system packs up to 384 Ascend 910C chips and works as a single logical computer delivering up to 300 PFLOPS, where one PFLOPS is a thousand trillion floating-point operations per second.^[1]

Huawei said the design lets the 384 chips be treated as one machine for learning, thinking and reasoning rather than as a cluster of separate servers. That distinction matters in practice: software written for one computer does not have to be rewritten to coordinate many independent nodes, and the whole system can be scheduled as a single unit.^[1]

CloudMatrix384 as a cloud service

The same hardware is offered through Huawei Cloud under the name CloudMatrix384, sold as a cloud service instance built on the Atlas 900 A3 SuperPoD. Cloud customers therefore rent the system by the instance rather than buying and operating the racks themselves.^[1]

Deployment

By September 2025, Huawei had deployed more than 300 of the systems to over 20 customers. Those customers were in internet service provider, telecom and manufacturing sectors, among others — that is, in businesses that both run large online services and apply AI to their own operations.^[1]

Huawei also said that, as of September 2025, the Atlas 900 A3 SuperPoD remained the world's largest SuperPoD.^[1]

Where it sits in the line

The SuperPoD line was extended in September 2025 with the Atlas 950 SuperPoD and Atlas 960 SuperPoD announcements, positioning the Atlas 900 A3 as the first commercially deployed member of a family rather than a one-off product.^[1]

Why it matters

The announcement is notable for two reasons. First, it shows a vendor organizing AI infrastructure around very large single-machine domains instead of conventional clusters of servers, a design choice that shifts attention to the interconnect as much as to the chips. Second, the cloud packaging and the reported customer count indicate that the system moved from announcement to paying users within months, across several industries rather than a single pilot customer. Together with the follow-on Atlas 950 and Atlas 960 announcements, it marks the point at which the SuperPoD idea became a product line.