Citing people familiar with the matter, Bloomberg reports that DeepSeek plans to deploy at least 160,000 Huawei Ascend 950DT accelerators in a data centre under construction in Ulanqab, Inner Mongolia, one of the largest Huawei AI chip clusters known. DeepSeek and Huawei have not commented.
The purpose is clear. The chips will run already-trained models, that is, inference, not training. On the training side the company still relies mainly on Nvidia accelerators, where peak compute and a mature software stack matter most.
Inference is about throughput, latency and unit cost, and it is the API and online-serving load that burns power and cards every day, far more often than pretraining. By pushing the domestic cluster first onto the part used daily, DeepSeek localises the serving layer while keeping its steadiest training tools on Nvidia.
Public reports say DeepSeek raised about RMB 50 billion in June 2026 and has since reopened funding talks, with infrastructure expansion a stated use of proceeds.
Editor’s note: This is an adapted translation of the original Sohu IT report. It has been trimmed and restructured for readability for an international business audience.
Translated and adapted from Sohu IT (https://www.sohu.com/a/1072541857_362225).