- Bloomberg reporting summarized by TechNode describes a planned cluster of at least 160,000 Huawei Ascend 950DT accelerators.
- The proposed Inner Mongolia deployment is intended for inference.
- Timing depends on chip supply; the report does not describe an already completed cluster.
A reported plan for inference capacity
DeepSeek plans to deploy at least 160,000 Huawei Ascend 950DT accelerators at a facility in Inner Mongolia, according to people familiar with the project cited in Bloomberg reporting and summarized by TechNode. The proposed installation is intended for inference: running models to answer requests.
The account says the company does not currently plan to use those processors for training. It also says deployment depends on Huawei's production capacity and describes a broader project targeting power consumption on a gigawatt scale. These remain reported plans, not verified measurements of an operating facility.
Inference and training put different demands on a system
Training updates a model's parameters using data and computation. Inference uses an already trained model to produce an output. A company can therefore use one hardware platform for serving some workloads while retaining another for model development; the reported cluster should not be taken as proof of a complete hardware migration.
Serving performance depends on the model, response length, concurrency and latency target, as well as the hardware. A large accelerator count cannot be converted directly into a number of users or requests per second without those workload assumptions.
The supply and infrastructure milestones
The proposed scale makes delivery sequencing important. Chips, networking, electricity, cooling and software must become usable together. Equipment delivered to a site is not the same as capacity ready for customer traffic, and a power target is not the same as a measured load.
The report does not disclose an equipment contract, a completion date, operating cost or a benchmark for the proposed installation. Its significance is a possible large source of demand for Huawei inference infrastructure. Confirmation of orders, staged commissioning and actual service capacity would be needed to turn that reported intention into a clear account of deployment.
Sources & context
Go to the original material. Company claims remain attributed to their sources.
01Updates & corrections
— Expanded with reporting details, source context and clearly attributed limitations.
