Huawei Ascend 950PR Architecture Confirmed as DeepSeek V4 Foundation — Detailed Specs Published

Industry analysts confirmed on April 5 that Huawei's Ascend 950PR — not the previously-reported 910C — is the specific chip architecture chosen by DeepSeek for its forthcoming V4 frontier model. TrendForce and DIGITIMES published detailed technical breakdowns: the Ascend 950PR delivers 1–2 petaflops of compute with an in-house HiBL memory stack of 112GB at 1.4 TB/s bandwidth, paired with a 2 TB/s inter-chip interconnect for cluster scaling. The chip includes Huawei's proprietary DaVinci AI compute cores optimized for mixture-of-experts transformer architectures — directly aligned with DeepSeek's V4 model design. The 950PR represents the top of Huawei's 2026 compute roadmap and is physically distinct from the 910C used in the Shenzhen 10,000-card cluster. Key industry implication: DeepSeek V4's adoption of the 950PR rather than 910C signals that China's frontier AI development is tracking Huawei's most advanced commercial silicon, not merely using readily-available supply. ByteDance, Alibaba, and Tencent have placed aggregate orders for hundreds of thousands of 950PR units with H2 2026 delivery, creating the demand signal needed to fund Huawei's next-generation chip R&D cycle.