DeepSeek Reportedly Plans a 160,000-Chip Huawei Inference Cluster in Inner Mongolia

DeepSeek is reportedly preparing to install at least 160,000 of Huawei’s Ascend 950DT accelerators at a data center under construction in Ulanqab, Inner Mongolia. The important distinction is purpose: the chips are intended principally for inference—running already trained models to answer users’ requests—rather than for training future frontier models. 114

If the plan is completed, it could create one of the largest known clusters of Huawei AI accelerators. It is also a significant escalation from a reported 10,000-chip Ascend cluster that began operating earlier, though the 160,000 accelerators would represent only one portion of the larger campus. 11214

What DeepSeek is building

The reported site is designed around roughly one gigawatt of capacity at full build-out. One published account compares that level of electricity use with the demand of about 750,000 homes, an illustration of the infrastructure scale involved rather than a direct measure of the chip cluster’s consumption alone. 3

A project of that size is not simply an order for accelerators. It requires enough electrical capacity, cooling equipment, networking, servers and operational systems to keep a large model-serving fleet available under sustained traffic. The reported 160,000 Ascend chips would occupy only part of the eventual gigawatt-scale facility. 14

Why inference, not training, is the key detail

Inference is the production phase of AI: a trained model receives prompts and generates responses. It requires capacity that can reliably handle ongoing demand, rather than a finite training run.

DeepSeek reportedly does not currently plan to use the Ascend 950DT deployment for training. Instead, training is expected to remain on Nvidia hardware, while the Huawei cluster runs deployed models. 18

That split matters because it makes this an inference-first domestic-hardware strategy, not evidence that DeepSeek has replaced overseas hardware across its entire AI stack. The project’s near-term proof point will be whether a very large Ascend deployment can serve real model traffic efficiently and reliably. 114

Delivery is the central uncertainty

The rollout depends on Huawei’s ability to manufacture and deliver the accelerators. Reports say component constraints, particularly high-end memory, could limit 950DT production to the low hundreds of thousands of units in 2026. A single order of at least 160,000 chips would therefore compete with demand from other customers and could take more than a year to fulfill. 1723

DeepSeek reportedly wants part of the capacity operating by late 2027 or early 2028, subject to production availability. 58

What the plan signals for China’s AI infrastructure

The proposal aligns with China’s broader effort to build more AI computing capacity on domestic hardware as access to leading foreign accelerators faces pressure. 1

But the announcement is not the same as a completed, production-ready cluster. At this scale, success depends on more than accelerator supply: networking, power delivery, cooling, systems integration and ongoing component availability must all work together. The eventual result will show whether Ascend can move from smaller deployments to sustained, large-scale model serving—and how quickly domestic infrastructure can take on that workload. 114