China tech analyst Rui Ma spotted an odd pair of numbers in Huawei’s latest AI presentation, and the discrepancy says as much as the headline did.
On paper the company described an Atlas 960 SuperPoD built from 15,488 Ascend 960 chips. The version shown this week at Huawei Connect came in at 4,096. Her conclusion on X was that the accelerator is arriving far sooner while the machine wrapped around it has shrunk.
The accelerator is indeed early. Ascend 960DT, Huawei’s next-generation AI chip, moves from a third-quarter 2027 target to the first quarter. David Wang, the company’s rotating and acting chairman, delivered the change. A spokesperson added that each generation is doubling performance and that the 960 line is running ahead of schedule.
Scale-out remains the strategy. Huawei wants vast fleets of accelerators to act as one system through what it calls Peerium Computing Architecture. UnifiedBus supplies the glue, linking processors to memory, storage and network fabric. Rotating chairman Eric Xu said the design covers training and inference. The first products are the Atlas 950 SuperPoD and SuperCluster, with Huawei claiming up to 256,000 accelerator cards in one SuperCluster.
Washington is watching. Trump and Xi Jinping meet on September 24. Xu’s position is that China should accelerate rather than slow down, on the argument that a country cannot understand the risks of frontier systems until it has built its own.
