DeepSeek는 수출 규제에 따라 엔비디아 칩의 대안으로 중국 AI 개발사들이 적극적으로 모색하는 가운데 화웨이 가속기에 대규모 투자를 진행하고 있다.

Chinese AI firm DeepSeek is planning a major expansion of its computing infrastructure, proposing to deploy at least 160,000 Huawei Technologies AI accelerators at a large data centre in Inner Mongolia. If completed, the installation would become one of the largest publicly known clusters built around Huawei’s Ascend processors. The project is designed to operate at a gigawatt scale, with the 160,000 chips representing only a portion of the facility’s total computing capacity. DeepSeek is reportedly pursuing billions of dollars in funding to support this broader infrastructure rollout.

The cluster will utilize Huawei’s next-generation Ascend 950DT accelerators to run DeepSeek’s artificial intelligence models. However, people familiar with the matter indicate that the deployment schedule could extend beyond a year. Huawei faces significant production bottlenecks, including shortages of advanced components such as high-performance memory, which are expected to limit the annual output of Ascend 950DT processors. Despite DeepSeek’s efforts to secure additional supplies of Huawei’s high-end processors, meeting immediate demand quickly remains challenging.

The planned deployment will focus primarily on inference—the execution of AI models after training—rather than the training phase itself. Although Huawei positions the Ascend 950DT for demanding workloads, DeepSeek does not currently plan to use these chips for model training. The company has previously tested Huawei hardware for training purposes but continues to rely heavily on Nvidia accelerators for this critical stage. To accelerate access to domestic silicon, DeepSeek has reportedly sought enhanced support from Chinese authorities.

This initiative aligns with Beijing’s intensified efforts to reduce reliance on Nvidia processors following stringent US export restrictions. These limitations have curtailed Chinese firms’ access to Nvidia’s most advanced AI accelerators, placing greater pressure on domestic manufacturers like Huawei to close the performance gap. Huawei’s Ascend family has emerged as a cornerstone of China’s domestic AI hardware strategy. The Ascend 950DT, introduced as a more advanced iteration, is positioned as a higher-performance solution for AI workloads and is widely viewed as competing with earlier generations of Nvidia processors.

Several Chinese technology companies are already scaling up their adoption of locally developed AI chips. Meituan, for instance, has reported training an AI model using tens of thousands of domestic processors, while other AI firms are constructing data centres optimized around home-grown silicon. DeepSeek’s proposed 160,000-chip installation marks a substantial leap in scale compared to earlier Chinese AI computing clusters, illustrating how the domestic ecosystem is adapting to technology restrictions through large-scale infrastructure built around local processors. Nevertheless, Nvidia hardware remains highly sought after, with companies continuing to pursue access through approved imports and overseas computing resources. The evolving equilibrium between domestic chips and Nvidia accelerators will likely serve as a key metric of China’s technological progress amid its broader strategic competition with the United States.