DeepSeek's Massive Huawei Ascend Order Signals Inference Shift Amid US Export Controls

Bloomberg reported on September 4, 2026, that Chinese AI startup DeepSeek plans to deploy at least 160,000 of Huawei's next-generation Ascend 950DT accelerators at a roughly 1-gigawatt data center under construction in Ulanqab, Inner Mongolia—potentially one of the largest known clusters of Chinese-made AI chips.[1][2] The site, about 350 km northwest of Beijing, targets partial operation by late 2027 or early 2028, with the 160,000 chips forming only a portion of the planned capacity alongside owned and leased infrastructure.[3]

DeepSeek intends to use the 950DT chips exclusively for inference—running trained models to serve user queries—rather than the more demanding training phase, despite Huawei marketing the processor for both.[1] This aligns with prior unsuccessful attempts by DeepSeek to train models on older Ascend generations, leading it to rely on Nvidia hardware for training while shifting inference workloads domestically.[4] The move underscores how US export controls, tightened on advanced memory like HBM in late 2024, have effectively segmented the market: blocking frontier training but allowing China to scale inference on sovereign silicon.[5]

Production bottlenecks persist. Huawei faces component shortages, particularly high-bandwidth memory, limiting 950DT output to the low hundreds of thousands in 2026; fulfilling the order could take over a year amid competing domestic demand and limited exports.[2][6] Huawei has developed proprietary HBM (HiZQ 2.0) for the chip to circumvent restrictions, achieving 4 TB/s bandwidth, though yield and scale remain challenges.[7] DeepSeek has reportedly sought government intervention to expedite deliveries.[8]

This development highlights broader patterns in China's AI push under the 'Eastern Data, Western Compute' initiative, with Ulanqab emerging as a hub for energy-efficient, large-scale compute.[9] While inference represents the higher-volume, revenue-generating segment, training sovereignty remains elusive, preserving Nvidia's edge in frontier model development even as smuggling or overseas rentals supplement limited legitimate access.