Huawei's Ascend 960 Superpod Debuts Optical Interconnect as Chip Roadmap Extends to 2029

At Huawei Connect, rotating chairman Wang Tao detailed the Ascend 960 chip and superpod alongside a multi-year roadmap. The Ascend 960DT shipped three quarters early, roughly doubling single-chip compute to 2 PFLOPS FP8 and 4 PFLOPS FP4, with up to 288GB HBM and 9.6TB/s bandwidth. The Ascend 960 superpod scales to 4,096 NPU cards for up to 8 EFLOPS FP8, and Wang confirmed a one-generation-per-year cadence through Ascend 970 (2028) and Ascend 980 (2029). It is the first superpod deployed at commercial scale with near-package optics (NPO): about 5,500 Hi-ONE optical engines replace roughly 48,000 800G optical modules, cutting power draw over 550 kilowatts and lifting availability to 99.8%. CANN, Huawei's software stack, has also passed what Wang called an inflection point — over 5,200 monthly active developers, 61% external, plus official PyTorch support.

Wang framed Huawei's ambition as the compute foundation under China's model companies rather than a competitor — Pangu still serves Huawei's own products, but the pitch to outside labs is infrastructure, not rivalry. That pairs with an admission that China's leading-edge fabrication still hasn't closed the process-node gap, so Huawei pushes gains into packaging, optics, and system architecture instead of waiting on lithography — the same logic behind the Kirin 9050 Pro's Tau Scaling architecture in the Mate XT 2 trifold.

For Chinese model developers, native PyTorch support and a maturing CANN ecosystem cut the cost of training on Ascend rather than porting later, while the stated path to 512,000-card clusters — with a technical ceiling near 1 million — signals Huawei is building for frontier-scale domestic demand. Per the AI Market Watch index, Huawei-tagged coverage held roughly steady at 25 items over 90 days versus 31 prior (name-matched, pipeline-ingested sources only), even as these technical claims escalate.