DeepSeek plans the largest known Huawei chip cluster with 160,000 processors in Inner Mongolia

Chinese AI company DeepSeek plans to install at least 160,000 Huawei Ascend 950DT accelerators in a data centre in Ulanqab, Inner Mongolia. That would make it the largest known cluster of Huawei chips in the world. The facility is intended entirely for inference, running AI models, rather than training them. For training, DeepSeek will continue to rely on Nvidia hardware for the foreseeable future.

At a street price of approximately $16,000 per chip, the estimated hardware cost for this cluster is around $2.56 billion. The planned data centre will have a compute capacity of roughly one gigawatt, comparable to the electricity consumption of around 750,000 households. DeepSeek aims to bring part of that capacity online in late 2027 or early 2028.

The timeline is under pressure. Huawei can only produce the Ascend 950DT in limited quantities, and full delivery of all 160,000 chips is expected to take more than a year from September 2026 onwards. This means full commissioning of the cluster will likely slip to 2028 or later.

What the Ascend 950DT chip does and what limits it

The Huawei Ascend 950DT is built on the third generation of Huawei's DaVinci v3 architecture and uses chipmaker SMIC's N+3 manufacturing process, a technology in the 7-nanometre class. Each chip features 144 GB of HiZQ 2.0 High Bandwidth Memory with a memory bandwidth of 4.0 TB/s and a GPU-to-GPU interconnect speed of 2.0 TB/s. The chip also supports the lower-precision formats FP8 and FP4, which are common in large-scale inference.

Huawei unveiled the Ascend 950DT approximately one year ago. Production capacity has since remained a bottleneck: Huawei is reported to manufacture several hundred thousand units per year. An order of 160,000 chips therefore places a significant claim on available output, particularly if other customers are also drawing on the same chip generation.

DeepSeek's choice of inference rather than training is notable. Inference places different demands on hardware than training: higher throughput at lower latency matters more than raw compute power for learning model parameters. Using Huawei hardware for inference also gives DeepSeek greater independence from Nvidia in the part of the infrastructure that is directly visible to end users.

DeepSeek as a company and its financial position

DeepSeek was founded in 2023 by Liang Wenfeng, who also led the quantitative hedge fund High-Flyer. DeepSeek is financed and directed from that fund. The company is headquartered in Hangzhou and had only 160 employees in 2025, which is small for an organisation with these infrastructure ambitions.

In May and June 2026, DeepSeek closed its first external funding round. The company raised more than 50 billion yuan, equivalent to over $7.4 billion, at a valuation of between $52 billion and $59 billion. Liang Wenfeng personally contributed around 20 billion yuan. Investors in this round include Tencent, CATL, JD.com, NetEase, IDG Capital, Monolith Management and China's state AI investment fund. That state fund holds voting rights and no lock-up period, distinguishing its position from that of the commercial investors.

At the same time, DeepSeek is in discussions about a pre-IPO round, targeting a valuation of approximately 500 billion yuan. Prospective investors in this round include CPE, Legend Capital and Stony Creek Capital. A stock market listing could potentially take place as early as 2027. Liang Wenfeng's net worth was estimated at $36 billion in July 2026.

What this reveals about the broader Chinese AI chip strategy

DeepSeek's decision to use Huawei hardware for inference at this scale reflects a broader shift in China. US export restrictions limit access to Nvidia's most advanced chips, forcing Chinese AI companies to seek or develop alternatives for large portions of their infrastructure. Huawei positions the Ascend line as the primary domestic alternative, but production capacity at SMIC acts as a structural constraint on deployment.

For DeepSeek, the inference infrastructure is directly tied to the commercial viability of its models for external users. The company gained wide recognition following the launch of DeepSeek-R1 in January 2025, a model that attracted attention for its performance relative to training costs. Further scaling of inference capacity is a logical next step for a company working towards an IPO and looking to serve a growing user base.

For investors and policymakers in Europe, the scale of this project serves as a reference point. A cluster of 160,000 chips with a target capacity of one gigawatt exceeds virtually all known European data centre projects for AI inference. The combination of state-backed financing, established tech investors and proprietary chip solutions gives the Chinese AI ecosystem a scaling capacity that is difficult to compare directly with the more fragmented situation in Europe, where chip access, data capacity and coordination between public and private parties are all still very much works in progress.