Huawei plans big 950PR run
Huawei’s new 950PR AI inference chip — CUDA‑compatible and aimed at inference workloads — is reportedly lined up with ByteDance and Alibaba and targeted to ship about 750,000 units this year, though that’s still a small fraction of global Nvidia output. (x.com)
Huawei told partners it sent customer samples in January and indicated mass production is scheduled for next month, with broader shipments targeted for the second half of 2026. (finance.yahoo.com) The company debuted the Atlas 350 PCIe accelerator card paired with the Ascend 950PR at its China Partner Conference on March 20, 2026. (trendforce.com) Trade reporting and vendor briefings list the Atlas 350 at about 1.56 petaflops of FP4 inference compute and cite vendor claims of up to a 2.8× throughput advantage versus Nvidia’s H20 in China-specific tests. (trendforce.com) Huawei’s published roadmap places the 950PR inside an Ascend 950 family that includes a 950DT companion for training/decoding and later 960 and 970 generations, and Huawei has said the 950 series will employ its in‑house HBM memory technology. (huaweicentral.com) Local reporting gives an initial price band of roughly 50,000–70,000 yuan per Ascend 950PR unit in early order discussions. (news.rthk.hk) Reuters notes the sourcing for these details came from unnamed industry contacts who declined to be identified and that Huawei did not respond to Reuters’ requests for comment. (finance.yahoo.com)
