The Hangzhou-based AI developer is expected to use the chips primarily for inference, the process of running trained models and generating responses, rather than for the more computing-intensive task of training frontier systems. DeepSeek continues to rely on Nvidia accelerators for training, underscoring the performance gap that still separates domestic processors from the US company’s leading hardware in some workloads.
The planned installation forms part of a roughly gigawatt-scale computing project in Ulanqab, about 350 kilometres northwest of Beijing. DeepSeek aims to bring at least part of the facility into operation by late 2027 or early 2028, although the schedule depends heavily on Huawei’s ability to manufacture and deliver enough processors.
Neither DeepSeek nor Huawei has publicly confirmed the chip order. The scale and timing therefore remain subject to change, particularly because supplies of high-end memory and other advanced components are expected to constrain production of the Ascend 950DT. Fulfilling an order of this size could take more than a year while Huawei also serves other domestic and overseas customers.
The order would be unusually large even by the standards of China’s accelerating AI build-out. Huawei has projected that its Atlas 950 SuperCluster can combine more than half a million 950DT processors, but supplying individual customers at that scale requires sufficient yields, memory and packaging capacity. DeepSeek’s proposed allocation alone would absorb a sizeable share of near-term output, potentially forcing deliveries to be staged as manufacturing volumes ramp through 2027.
Huawei has said the Ascend 950DT is designed for both the decode stage of AI inference and model training. The company’s published roadmap gives the processor 144 gigabytes of proprietary high-bandwidth memory, memory-access bandwidth of 4 terabytes per second and interconnect bandwidth of 2 terabytes per second. Huawei has scheduled the chip for availability in the fourth quarter of 2026.
DeepSeek’s decision to reserve the processors mainly for inference would reflect a broader pattern among China-based AI developers, which are increasingly adapting models to domestic hardware while retaining Nvidia systems for demanding training jobs where available. DeepSeek has already optimised models for Ascend processors, helping reduce dependence on imported chips for serving large volumes of user queries.
The planned cluster would also advance Beijing’s effort to build a more self-reliant AI computing ecosystem as US export controls restrict China’s access to Nvidia’s most advanced processors. Authorities have encouraged technology companies to adopt domestic chips, while access to some Nvidia products remains tightly controlled by both Washington’s licensing regime and Beijing’s own approvals.
DeepSeek has sought government support in securing a larger and faster allocation of Huawei processors for the project. The request highlights the pressure on domestic supply as AI laboratories, cloud providers and internet groups compete for scarce accelerators capable of supporting increasingly large models and growing inference demand.
The Ulanqab facility is designed to provide about one gigawatt of computing capacity. The proposed 160,000-chip Huawei deployment would account for only part of that total, leaving open the question of which processors DeepSeek will use for the remainder of the site.
Huawei is simultaneously expanding the scale of its own AI infrastructure products. Its Atlas 950 SuperPoD, built around Ascend 950DT chips, is designed to connect as many as 8,192 processors into a single logical system. Huawei has also outlined an Atlas 950 SuperCluster using more than 520,000 Ascend 950DT chips, illustrating the company’s push to compete at data-centre scale rather than solely at the level of individual accelerators.
For DeepSeek, the planned purchase would represent a substantial increase in dedicated domestic computing capacity as it spends heavily on infrastructure. The company has been expanding server access, chip procurement and data-centre capacity while developing successive generations of large AI models.
Follow Arabian Post
Select Arabian Post as your preferred source on Google and MSN News for trusted business news and Arab politics and updates.
