DeepSeek is reportedly planning an Inner Mongolia data center built around at least 160,000 of Huawei’s next-generation Ascend 950DT accelerators, a scale that would make the facility one of the largest known collections of Huawei AI chips. The plan was first reported by Bloomberg, citing people familiar with the matter, and subsequently summarized by TechNode. Neither DeepSeek nor Huawei confirmed the plan when Bloomberg sought comment, so the proposed cluster should be treated as a reported project rather than a completed deployment.
The reported design matters because it is aimed at inference, the process of responding to user prompts and running models after training. That focus distinguishes the plan from the more visible competition over the biggest training clusters. China’s AI companies need sustained capacity not only to create models but also to serve increasingly large volumes of users and enterprise customers. If the project proceeds on the reported scale, it would make inference infrastructure a much more visible part of DeepSeek’s strategy.
The new report also extends a story EastFrontier has followed since DeepSeek began hiring for its own AI data-center infrastructure. At that time, the company’s infrastructure ambitions were visible but the eventual hardware scale was not. The current report puts a specific, if unconfirmed, number on the potential demand for Huawei accelerators.
A 160,000-Chip Plan Built for Model Serving
Bloomberg reported that DeepSeek intends to deploy at least 160,000 Ascend 950DT chips at a facility under construction in Ulanqab, Inner Mongolia. The people cited by the outlet said the chips would be used to operate models, rather than to train them. That distinction is central to the reported architecture because Huawei designed and marketed the 950DT for the more demanding work of training.
Inference has different operational pressures from training. A training project can consume a large block of computing time while a model learns from data. Inference must respond continuously to requests from developers, consumers, and business customers. For a company with public APIs, chat products, and potential enterprise services, the ability to deliver responses at scale can be as commercially significant as the ability to train a new model.
The Ulanqab location also fits a broader pattern in Chinese AI infrastructure. Inner Mongolia has become an important location for data centers and computing projects because of available power resources and industrial development. EastFrontier recently examined an Inner Mongolia effort linking green computing with AI service exports. The DeepSeek plan, if realized, would add a company-specific example to that regional infrastructure story.
However, the reported 160,000-chip figure is not a purchase order publicly disclosed by either company. Bloomberg’s sources said that installation timing would depend on Huawei’s production capability. The report also said that high-end-memory constraints could restrict Huawei’s 950DT output to the low hundreds of thousands in 2026. These are source-based estimates, not official production guidance.
Huawei Supply Limits Shape the Reported Timeline
The proposed cluster highlights a problem that will affect any domestic AI-chip buildout: securing enough accelerators is only part of the challenge. Bloomberg reported that Huawei faces shortages of high-end components, including top-tier memory, and that fulfilling the full DeepSeek requirement could take more than a year. Huawei must also balance demand from other customers and work on limited overseas shipments, according to the people cited.
This constraint makes the story about supply as much as demand. A company can identify a data-center site, prepare power, and seek financing, but still be unable to bring capacity online quickly if accelerator production is limited. For DeepSeek, a large planned installation would create a dependency on Huawei’s ability to manufacture, allocate, and deliver chips over time.
The report also places the 950DT in a different role from the one Huawei originally emphasized. Using a training-oriented chip for inference may reflect DeepSeek’s need for large-scale serving capacity, the availability of hardware, or both. The sources did not provide a full technical explanation, so it would be unsafe to infer performance comparisons with Nvidia products or to claim a specific efficiency advantage.
DeepSeek’s prior work with Huawei hardware adds context without proving the new plan. EastFrontier covered a Huawei-led claim of full-parameter post-training for DeepSeek V4 Pro on Ascend 910C. That was a separate technical claim involving a different Ascend generation and a post-training task. The latest report concerns a proposed future deployment and should not be described as evidence that DeepSeek has already shifted all of its training or inference work to Huawei chips.
Inference Capacity Becomes a Strategic Asset
DeepSeek’s reported cluster plan points to a change in what counts as AI infrastructure leadership. Training capacity still attracts attention because it can determine how often a company can develop and refresh frontier models. But inference capacity determines whether a model can become a dependable service. A model with strong benchmark scores cannot generate recurring revenue or broad public adoption if its serving system is too limited or too expensive.
The proposed scale would also signal confidence in continued demand for DeepSeek’s models. Building or reserving gigawatt-scale infrastructure is a major commitment. Bloomberg reported that the wider data-center project targets power consumption on that scale. Yet demand expectations, hardware deliveries, and financing terms could all change before the facility is completed.
This is why the wording around the plan matters. DeepSeek has not publicly announced the 160,000-chip installation. Huawei has not confirmed supply. The project’s scale, timetable, and technical configuration therefore remain subject to the uncertainty inherent in anonymous-source reporting. Investors and customers should read it as evidence of the direction of DeepSeek’s infrastructure ambitions, not as proof that the cluster is operating today.
Still, the report provides a useful view of China’s AI buildout. It suggests that domestic accelerators may be deployed in volumes once associated primarily with foreign suppliers, and that large Chinese model companies are designing infrastructure around the operational challenge of serving AI at scale. Whether DeepSeek can execute on the reported plan will depend on chip supply, power, software optimization, and the demand generated by its model ecosystem.
For Huawei, the project would be an important test of whether its next-generation accelerator line can move beyond individual technical demonstrations into the backbone of a major model company’s inference service. For DeepSeek, it would be a bid to turn model popularity into durable infrastructure. The current evidence supports neither conclusion as a completed fact, but it makes both questions harder to ignore.