Huawei Accelerates Ascend 960 AI Chip to Q1 2027, Unveils Ascend 960 SuperPoD at Huawei Connect

At Huawei Connect 2026 in Shanghai on September 17, rotating chairman David Wang announced Huawei is pulling forward commercial launch of its next-generation Ascend 960DT AI training chip to Q1 2027 — three quarters ahead of its prior roadmap — with the inference-focused Ascend 960PR following in Q3 2027, one quarter early. Wang said Ascend 960 development 'exceeded expectations,' with performance doubling as planned, and committed Huawei to a one-generation-per-year Ascend cadence, previewing the Ascend 970 (2028) and Ascend 980 (2029) built on the company's 'Tau (τ) Scaling Law' architecture. Huawei also unveiled the Atlas 960 SuperPoD, which it describes as the industry's first cluster built on Near-Packaged Optics: it links 4,096 Ascend 960 NPU cards for up to 8 exaflops of FP8 compute and 1 petabyte of HBM capacity, using 5,500 Hi-ONE optical units in place of roughly 48,000 traditional 800G modules to cut power draw by over 550 kilowatts and push system availability to 99.8%. Analyst Rui Ma noted the announced SuperPoD configuration (4,096 cards) is smaller than the 15,488-chip cluster Huawei had previously floated for the platform. Separately, Huawei Cloud confirmed its existing Ascend 950 AI-cluster cloud service — a 1,024-card system delivering up to 1 EFLOPS FP8 / 2 EFLOPS FP4 — goes commercial in China September 30 and expands globally November 30. The announcements land one week before Xi Jinping's planned September 24 Washington visit, underscoring Beijing's confidence in its domestic AI hardware stack even as US export controls persist.