DeepSeek plans to deploy at least 160,000 of Huawei Technologies’ Ascend 950DT accelerators at a data center it is building in Inner Mongolia, Bloomberg reported Friday, citing people familiar with the matter. CNBC TV18 carried the Bloomberg report in full; Outlook Business summarized the same figures. Bloomberg Television also ran a same-day segment on the planned cluster.
The chips would go mainly to run models, not train them, one of those people said, even though Huawei markets the 950DT for training as well. DeepSeek has tried Huawei silicon for training before but has so far relied on Nvidia accelerators for that step, according to the report. Neither DeepSeek nor Huawei confirmed the plan when Bloomberg asked.
Fulfillment is not immediate. Component shortages, including high-end memory, are expected to keep Huawei’s 950DT output in the low hundreds of thousands this year, and the company is balancing other customers plus small export volumes, the people said. Completing DeepSeek’s order could take more than a year. DeepSeek wants more of Huawei’s best chips than current capacity can deliver, the report said.
The Inner Mongolia project is framed as gigawatt-scale compute, with the 160,000 Ascend units only part of planned capacity. Bloomberg previously reported the site around Ulanqab in July as a build-and-lease mix aimed at late 2027 or early 2028 partial operations. If the Ascend cluster lands as described, it would be among the largest publicly discussed Huawei AI deployments as China pushes domestic chips under U.S. Nvidia export limits.
This is a hardware-supply story about inference capacity in China, not a model launch.