DeepSeek to Build 160,000-Chip Huawei Cluster

DeepSeek is preparing to establish a massive computing cluster in Inner Mongolia featuring at least 160,000 of Huawei's next-generation Ascend-950DT processors. According to reports, this facility will focus exclusively on running AI inference rather than training new models. For its computationally intensive training workloads, the Chinese AI company continues to rely on Nvidia hardware. If successfully deployed, this project will represent the largest known cluster powered by Huawei silicon to date.

However, realizing this ambitious infrastructure project will take time. Huawei is expected to face significant production constraints and memory chip shortages, meaning it could take more than a year to fully deliver the massive order. A critical bottleneck lies in high-bandwidth memory (HBM). While China's leading memory manufacturer, CXMT, has started producing its first small batches of HBM3E, the domestic industry remains three to five years behind global leaders like Samsung, SK Hynix, and Micron, which are already mass-producing advanced HBM4.

This massive deployment aligns with Beijing's strategic push to foster a self-reliant domestic semiconductor ecosystem without losing ground in the global AI race. For AI practitioners and system architects, DeepSeek's strategy highlights a pragmatic, hybrid approach to hardware constraints. By offloading inference workloads to domestic chips like the Ascend-950DT while reserving scarce Nvidia GPUs for training, developers can scale their user-facing applications locally. This blueprint demonstrates how organizations can navigate strict export controls and build resilient, multi-vendor infrastructure in a fractured global supply chain.