DeepSeek’s 160,000 Huawei Chip Plan Could Test China’s AI Supply Chain

DeepSeek is reportedly preparing a major Huawei-based AI deployment in Ulanqab, Inner Mongolia: at least 160,000 next-generation Ascend 950DT accelerators intended to run models at a data center planned at roughly one gigawatt of compute capacity. Neither DeepSeek nor Huawei has publicly confirmed the order or a delivery timetable, so the plan should be viewed as reported intent rather than an installed cluster. 45

What DeepSeek is reportedly building

People familiar with the project told reporters that DeepSeek intends to use Ascend 950DT chips primarily for inference—serving model requests—rather than for its principal model-training work. At the reported 160,000-chip scale, the installation could rank among the largest known clusters of Huawei AI accelerators. 58

The distinction between inference and training matters. A large inference fleet can support broad model availability and lower operating costs at scale, but it does not by itself demonstrate that Huawei hardware has displaced Nvidia across DeepSeek’s entire computing stack. Reporting on the plan says DeepSeek continues to use Nvidia hardware for training. 8

Why the full deployment may take more than a year

The critical constraint is supply, not simply the ability to finance or construct a data center. Reporting says DeepSeek is seeking more of Huawei’s leading chips than Huawei can currently provide, and that shortages of high-end memory and other components could hold Ascend 950DT output to the low hundreds of thousands this year. That makes a 160,000-unit commitment exceptionally large relative to available supply. 8

Three factors compound the risk:

• High-end memory availability: Advanced AI accelerators depend on high-bandwidth memory and other specialized components. A shortage in any one of those inputs can limit completed-chip shipments. 8
• Finite Huawei production capacity: The installation timetable reportedly depends on Huawei’s ability to manufacture the chips. Expanding allocation to one customer does not immediately expand total output. 5
• Competing buyers: Demand for Ascend 950 chips increased after DeepSeek released V4. Reuters reported that ByteDance, Tencent and Alibaba were among major Chinese internet firms seeking orders, putting DeepSeek in competition with cloud and internet companies for the same supply.

As a result, government support or favorable allocation could improve DeepSeek’s position in the queue, but it would not remove the underlying component and manufacturing bottlenecks. The practical test will be how many accelerators are delivered, integrated and kept operating—not the size of the reported target.

How the project fits China’s push beyond Nvidia

The proposed Ulanqab cluster aligns with China’s effort to build an AI stack less dependent on U.S. technology. DeepSeek’s V4 preview was adapted to run on Huawei’s advanced Ascend chips, and Huawei said V4 was supported on its Ascend 950-based supernode clusters. Huawei also said its chips were used for part of V4-Flash’s training. 23

That software work is strategically significant. AI hardware adoption depends on more than accelerator availability: models, frameworks, tooling and deployment practices must work well together. A prominent model optimized for Ascend gives cloud providers, enterprises and developers a reason to invest in Huawei’s platform, potentially reinforcing demand for both the chips and the surrounding software ecosystem.

Still, the reported Ulanqab plan should not be treated as proof of complete Nvidia independence. The project is focused on inference, while DeepSeek’s training workloads reportedly continue to rely on Nvidia systems. 8

Why scale is both the opportunity and the problem

A 160,000-chip cluster would be a high-profile demonstration that a Chinese AI developer can aim for infrastructure on a hyperscale footing. It could provide DeepSeek with substantial capacity to operate models internally rather than relying entirely on outside cloud capacity.

But its size also concentrates supply-chain risk. A smaller deployment might be fulfilled through staged production and allocation; a 160,000-chip target must compete for a substantial share of a constrained product line while the wider Chinese market is trying to secure the same hardware. 8

The milestone to watch

The headline number signals ambition, not completion. The meaningful milestones are confirmed orders, the pace of Ascend 950DT deliveries, usable installed capacity, and the economics and reliability of running DeepSeek models on the system.

If the cluster reaches operation at anything close to its reported scale, it would be an important step for Huawei’s Ascend ecosystem and China’s domestic AI infrastructure effort. If deliveries slip, it will illustrate the gap between designing a domestic AI stack and manufacturing every scarce component required to deploy it at scale. 58