Sources: DeepSeek plans to deploy 160K+ Huawei Ascend 950DT chips in an Inner Mongolia data center, which would create one of the largest Huawei chip clusters
Context & Ripple Effects
DeepSeek’s reported chip plan extends its proposed 1 GW Inner Mongolia data center, for which at least some capacity was targeted for late 2027 or early 2028. It also follows DeepSeek’s reported use of Ascend chips for smaller R2-model training in 2025, while retaining Nvidia chips for its largest models.
The reported deployment would turn that mixed-compute approach into a far larger commitment to Huawei hardware. Earlier bulk orders tied to DeepSeek V4 were reported to have driven a 20% increase in Ascend 950PR prices, making supply allocation and system delivery central to Huawei’s opportunity.
First-order effects
- If the reported deployment proceeds, DeepSeek becomes a major planned customer for Huawei’s Ascend 950DT, tying a substantial portion of the Inner Mongolia site’s initial compute buildout to Huawei hardware.
- Huawei would gain a large prospective workload for its accelerator platform, while DeepSeek would need to integrate the chips into a data-center-scale training environment rather than use them only for smaller-model work.
Second-order effects
- A large DeepSeek commitment would put additional pressure on Huawei’s accelerator supply and delivery capacity; other prospective Ascend buyers would face a market already shown to react to bulk demand with higher 950PR pricing.
- DeepSeek’s data-center build shifts competition from accelerator procurement alone toward the combined availability of chips, facility capacity, and software needed to run large model-training workloads.
Third-order effects
- If similar deployments are executed, large AI builders will increasingly treat a second hardware source as an infrastructure strategy, not merely a hedge for selected workloads.
- The durable advantage in AI compute would move toward integrated stacks that can deliver accelerators and power-ready data-center capacity together at cluster scale.
The trend: AI infrastructure is moving toward second-source, integrated compute stacks in which chip supply, data-center capacity, and model workloads are planned as one system.