DeepSeek plans to deploy at least 160,000 Huawei AI chips, using Ascend 950DT accelerators in a new Inner Mongolia data centre to reduce reliance on Nvidia.

DeepSeek plans to deploy at least 160,000 Huawei AI chips

DeepSeek plans to deploy at least 160,000 Huawei AI chips at a new data centre in Inner Mongolia, according to a Bloomberg report. The deployment would use Huawei’s next-generation Ascend 950DT accelerators and could form one of the largest known clusters of Huawei AI chips to date.

The move is part of China’s broader push to build AI infrastructure with domestic semiconductors amid US export restrictions that limit Chinese companies’ access to Nvidia’s most advanced GPUs. DeepSeek, a Chinese AI startup, is aiming to reduce its reliance on foreign hardware for running its models, even as it continues to depend on Nvidia for training.

What the deployment involves

The planned facility is located in Inner Mongolia and is being built as part of a larger, gigawatt-scale data centre project. DeepSeek intends to use the Ascend 950DT chips primarily for inference — running trained AI models — rather than for training new models from scratch.

Key points from reports include:

• At least 160,000 Ascend 950DT chips are planned for the site.
• The chips will be used to operate DeepSeek’s AI models.
• The installation timeline depends on Huawei’s ability to manufacture and deliver the chips.
• Component shortages, especially in advanced memory, may limit Huawei’s 950DT output to the low hundreds of thousands this year.
• Fulfilling DeepSeek’s order could take more than a year.

The 160,000 chips are expected to represent only one portion of the overall data centre’s capacity.

Why DeepSeek is using Huawei chips

DeepSeek’s decision reflects both commercial and strategic considerations:

• US restrictions on Nvidia: American export controls have made it harder for Chinese firms to obtain Nvidia’s most powerful AI accelerators.
• Domestic supply chain: Using Huawei chips supports China’s goal of building a self-reliant AI hardware ecosystem.
• Cost and availability: Domestic chips may be more accessible and potentially cheaper over time, especially if production scales.
• Inference-focused workload: Running models (inference) can be more tolerant of hardware limitations than large-scale training.

Despite this shift, DeepSeek has not abandoned Nvidia entirely. The company has reportedly tried training earlier models on Huawei hardware but continues to rely on Nvidia GPUs for core training workloads.

Production and delivery challenges

Huawei faces several constraints in meeting DeepSeek’s demand:

• Limited output of the Ascend 950DT due to shortages of advanced memory and other components.
• Competing orders from other customers, including domestic and some overseas buyers.
• A production ramp-up that may not keep pace with DeepSeek’s planned deployment schedule.

Because of these factors, sources told Bloomberg that fully equipping the data centre with 160,000 or more chips could take more than a year. DeepSeek is also reported to be in talks to raise billions of dollars to fund the buildout.

Strategic significance for China’s AI ecosystem

If completed as planned, the Inner Mongolia cluster would be a major milestone for China’s domestic AI hardware ambitions. It would demonstrate that:

• Large-scale AI inference can be run on Chinese-designed accelerators.
• Domestic chips can support production AI services at significant scale.
• Chinese AI firms can partially decouple from Nvidia for inference workloads.

However, the continued use of Nvidia for training underscores that China still lags in the most demanding AI workloads. Training frontier models requires enormous compute, high-bandwidth memory, and mature software ecosystems that currently favour Nvidia’s platforms.

Summary: DeepSeek plans to deploy at least 160,000 Huawei AI chips in a new Inner Mongolia data centre, using Ascend 950DT accelerators mainly for inference. The project supports China’s push to reduce reliance on Nvidia, but component shortages mean full delivery could take more than a year, and training still depends on Nvidia GPUs.