Infrastructure Market: Huawei Unveils Ascend 960 Chips, Claims System-Level Parity with Nvidia

At Huawei Connect in Shanghai, Huawei unveiled its next-generation AI accelerators—the Ascend 960DT and 960PR—scheduled for Q1 and Q3 2027, respectively, about three quarters earlier than planned. Designed for training and inference of advanced open-weight models, these chips power the Atlas 960E SuperPoD, a system connecting 4,096 processors with 8 EFLOPS FP8 and up to 1 PB of high-speed memory. Huawei emphasizes system-level integration and open-source ecosystems, with over 40 models pre-trained on Ascend, including GLM-5. However, per-chip performance still trails Nvidia's Rubin generation, though system-level parity is closer. DeepSeek has ordered at least 160,000 Ascend 950DT chips for inference but still trains on Nvidia hardware. Huawei's in-house HBM development addresses US export restrictions, and the company projects a 100,000-fold increase in AI token processing by 2035.