DeepSeek Plans 160,000 Huawei Accelerators for Inner Mongolia AI Data Center

DeepSeek plans to equip a new data center in Inner Mongolia with at least 160,000 Huawei accelerators, a deployment that would rank among China’s largest known AI computing clusters. The facility is intended to support the startup’s next generation of models and deepen its reliance on domestic hardware, Bloomberg reported in separate coverage of the plan. The project places Huawei at the center of DeepSeek’s computing strategy as Chinese companies face tighter access to Nvidia’s advanced processors.

The proposed cluster follows DeepSeek’s work adapting its V4 model to Huawei technology. Reports identify the Ascend family as the relevant hardware, while coverage of the planned facility describes Huawei’s top accelerators without specifying the exact model. The scale of the order matters because it would turn chip compatibility into operating infrastructure, rather than a laboratory demonstration.

DeepSeek’s V4 program has become a major catalyst for demand for Huawei AI chips. The model’s software was optimized for domestic silicon, prompting ByteDance, Tencent and Alibaba to seek additional Ascend 950 supply after V4’s release, according to people familiar with procurement discussions cited by Reuters and republished by Finance.yahoo. Capacityglobal also described a scramble for Huawei processors, linking the rush directly to DeepSeek’s hardware shift.

The commercial signal is significant because large internet companies are evaluating Huawei chips for workloads beyond basic inference. Their interest reflects a need to secure capacity before DeepSeek’s adoption increases competition for a constrained supply of domestic accelerators. Electronicsforyou likewise described bulk orders from Alibaba, Tencent and ByteDance as evidence of rising confidence in China’s local AI stack, although that outlet provided fewer independently detailed facts.

Chinese accelerators have made their clearest recent advance in model training through post-training work on DeepSeek V4-Pro. A Huawei-linked research team completed full-parameter post-training for the 1.6-trillion-parameter model using at least 1,000 Ascend 910C chips, according to the Shenzhen municipal government as cited by South China Morning Post and Tom’s Hardware. The achievement addresses a harder problem than running finished models for users.

South China Morning Post framed the work as a step from inference support toward training, where Chinese chipmakers have faced greater technical and scaling challenges. Tom’s Hardware emphasized the size of the cluster and the claim that it handled full-parameter post-training. Together, the reports show why DeepSeek’s planned data center carries strategic weight: its value depends on coordinating large numbers of domestic chips, software, networking and power systems at production scale.

DeepSeek’s Huawei deployment advances China’s effort to reduce dependence on Nvidia, but it doesn’t establish parity between the two platforms. Nvidia remains the benchmark for accelerator performance, software maturity, interconnects and large-scale availability. Economic Times coverage explicitly questioned whether Huawei can match Nvidia in power and scale, while describing the shift as part of China’s response to export restrictions.

Huawei’s advantage is strategic fit. Domestic chips give Chinese AI companies a supply route shaped by local policy and procurement rather than access to restricted Nvidia products. That route carries constraints of its own, including limited production capacity and the engineering burden of adapting models to a different software stack. Capacityglobal reported that export-control pressures helped drive the pivot while also limiting the output of the chips being sought.

DeepSeek’s relationship with Huawei is becoming both collaborative and competitive. Huawei hardware supports the startup’s model development and planned infrastructure, while DeepSeek is also reported to be developing an inference chip of its own. That chip is designed for generating responses from trained models, rather than training new ones, according to Reuters-based coverage summarized by Memeburn.

The division points to a broader Chinese strategy: use Huawei and other domestic processors for large-scale training and infrastructure while designing specialized silicon for high-volume inference. Such specialization can reduce operating costs, but it also creates pressure to maintain compatibility across chips, compilers and model-serving systems. Nvidia faces a direct displacement challenge in China, Huawei gains a flagship customer and technical partner, and DeepSeek retains an incentive to control more of its hardware stack.

The next test is execution. DeepSeek must secure tens of thousands of additional accelerators, bring the Inner Mongolia facility online and operate a large Huawei-based cluster efficiently. Supply constraints, power requirements, networking performance and software reliability will determine whether the announced scale translates into useful training and inference capacity.

The model’s performance will provide the clearest measure of the strategy. Semiwiki’s Reuters-sourced account said the V4-Pro version surpassed other open-source models on world-knowledge benchmarks while trailing Google’s Gemini-Pro-3.1, according to DeepSeek. If that performance holds as deployment expands, the project will strengthen China’s domestic AI supply chain. If hardware bottlenecks persist, the facility will remain a powerful symbol with a slower operational payoff.