DeepSeek은 내몽골에 추론 클러스터를 구축하기 위해 화웨이 애센드 950DT 칩 16만 개를 배치할 계획이다.

Hangzhou-based AI lab DeepSeek plans to install at least 160,000 of Huawei’s next-generation Ascend 950DT accelerators at a data center under construction in Inner Mongolia, according to people familiar with the matter cited by Bloomberg. If completed at that scale, the deployment would rank among the largest publicly described clusters built around Huawei AI silicon, marking a significant step-change from earlier Ascend installations measured in the low tens of thousands of cards.

The company intends to use the 950DT cards primarily for inference—serving production models to users—rather than for the heavier work of training new systems. While DeepSeek has experimented with Huawei hardware for training in the past, it continues to rely on Nvidia accelerators for that stage. The Ascend 950DT is marketed for both decode and training workloads; however, DeepSeek’s near-term plan keeps the two jobs on separate stacks, reflecting where domestic accelerators currently excel for high-volume serving.

The project’s location and power ambitions are as notable as its chip count. The Inner Mongolia campus is designed as a gigawatt-scale facility, capable of drawing enough power at full utilization to supply roughly 750,000 homes. The region’s inexpensive western power and favorable cooling conditions align with the “eastern data, western compute” strategy widely adopted by Chinese operators for dense AI workloads. The planned 160,000 Ascend 950DT units will occupy only a portion of that capacity, leaving the accelerator mix for the remainder of the site undefined.

Delivery timing hinges on Huawei’s ability to manufacture the 950DT at volume. Shortages of high-end memory and intense competition from other customers are expected to cap annual 950DT output in the low hundreds of thousands this year, meaning fulfilling DeepSeek’s order could extend beyond a year. To accelerate delivery, DeepSeek has also sought policy support to secure a larger, faster allocation of Huawei’s top-tier chips amid fierce competition from other cloud and internet providers.

This build reflects a broader industry shift toward domestic AI accelerators for large-scale serving. Chinese firms have already deployed smaller Ascend clusters, while labs continue to hybridize their stacks—using local silicon for inference and imported or overseas-rented capacity for training when necessary. Neither DeepSeek nor Huawei commented on the reported plan. For observers tracking China’s AI compute landscape, the project serves less as a single purchase order and more as a stress test: whether Ascend’s supply chain, networking, power, and cooling infrastructure can sustain model traffic at true data-center scale.