DeepSeek Plans 160,000 Huawei AI Chips for Inference, Still Training on NVIDIA

A huge Huawei order, but not a full Nvidia break

DeepSeek is reportedly planning to deploy at least 160,000 Huawei Ascend 950DT accelerators at a data center in Ulanqab, Inner Mongolia, a buildout that would rank among the largest known clusters based on Huawei AI chips if it is completed as described [1]. The plan is significant because it would move a major portion of DeepSeek’s production serving capacity onto domestic Chinese silicon, but it does not mean the company has abandoned Nvidia for the most demanding part of AI development [1].

The key distinction is workload. DeepSeek is expected to use the Huawei chips for inference: running already trained models to answer user requests, power APIs, and serve production traffic [1]. Training, the process of building and updating frontier models, is still expected to remain on Nvidia hardware, according to multiple reports citing people familiar with the plan [1][3]. That makes the order less a declaration of full technological independence than a pragmatic division of labor between available Chinese accelerators and the Nvidia ecosystem that remains stronger for large-scale training [3].

The reported order also remains a plan, not a completed deployment. TechNode described the timing and scale as reported intentions rather than an already finished installation, and noted that the deployment schedule depends on Huawei’s ability to produce the chips [5]. That caveat matters: an order of this size tests not just chip design, but packaging, high-bandwidth memory availability, networking, power delivery, software maturity, and the ability to operate a very large cluster reliably.

Why inference is the first target

Inference has become the commercial battlefield for AI labs because it is where models turn into recurring service costs. Every chatbot answer, coding completion, summarization request, or enterprise API call consumes compute. If a provider can run those workloads cheaply and predictably, it can cut prices, support more users, and reduce dependence on scarce imported accelerators.

That is why DeepSeek’s reported Huawei plan focuses on inference rather than training [1]. YFarmX summarized the split plainly: DeepSeek wants the Ascend 950DT chips to run models at Ulanqab while keeping Nvidia for training runs [1]. Mint also reported that DeepSeek is not currently planning to use the Ascend 950DT for model training, despite Huawei positioning the chip for more demanding workloads [3].

The division reflects a broader reality in AI hardware. Training a frontier model stresses interconnect, software stability, cluster scheduling, data pipelines, and numerical performance over long, failure-sensitive runs. Inference can still be technically demanding, especially for large language models with long context windows, but it can be optimized around throughput, memory bandwidth, batching, quantization, and workload management. In practical terms, a chip that is not yet trusted for the hardest training runs may still be valuable if it can serve many production requests efficiently.

The Ascend 950DT’s reported specifications fit that inference logic. Fresh coverage of Huawei’s roadmap describes the chip as built around 144 GB of Huawei high-bandwidth memory, roughly 4 TB/s of memory bandwidth, 2 TB/s of interconnect, and support for low-precision formats used to improve model-serving efficiency [1][4]. TechPowerUp’s report, as carried by HeadlinesBriefing, calculated that 160,000 such chips would represent about 160 exaFLOPS of FP8 peak compute and roughly 23 PB of total HBM capacity if multiplied across the cluster [6]. Those are theoretical figures, but they illustrate why memory-rich accelerators are attractive for serving large models.

The scale of the Ulanqab buildout

The reported location, Ulanqab in Inner Mongolia, is part of the story. The city sits northwest of Beijing and has become associated with large-scale computing because of available land, power resources, and cooler operating conditions [1]. Several reports describe DeepSeek’s broader Inner Mongolia infrastructure ambition as gigawatt-scale, with the 160,000 Huawei chips representing only part of the total planned capacity [1][5].

A gigawatt-class site would place DeepSeek’s infrastructure ambitions in the same conversation as hyperscale AI facilities built by the largest global technology companies. Mint noted that a data center containing more than 100,000 AI chips would broadly match the scale of major Western AI infrastructure projects, though Chinese facilities would still be working with processors generally described as less powerful than Nvidia’s leading accelerators [3]. That comparison is important because AI competition is increasingly measured not only in model quality, but in the ability to operate large, power-hungry fleets continuously.

The reported 160,000-chip quantity is also large relative to known Chinese domestic accelerator deployments. Control Plane said the cluster would be unusual by domestic Chinese silicon standards and compared it with a publicly reported 10,000-chip Huawei cluster that came online earlier in 2026 [2]. YFarmX framed the same scale in Huawei’s own system units: 160,000 Ascend 950DT chips would equal a little under twenty Atlas 950 SuperPoDs, or roughly 30 percent of an Atlas 950 SuperCluster configuration as described in Huawei’s roadmap [1].

That comparison should be read carefully. Huawei’s SuperPoD and SuperCluster figures are product architecture references, not proof that such a DeepSeek system is already running. The useful point is scale: DeepSeek’s reported order is big enough to be strategically important, yet still small enough to show that supply, not ambition, may be the limiting factor.

Huawei’s bottleneck: memory and production capacity

Several recent reports point to the same constraint: Huawei may not be able to fill an order of this size quickly [1][3][5]. Mint reported that Huawei is expected to produce only a few hundred thousand Ascend 950DT chips this year because of shortages of key components, including advanced memory [3]. TechNode also said the installation schedule depends on Huawei’s production capacity [5].

That supply issue changes the meaning of the deal. If DeepSeek wants 160,000 chips, and Huawei’s annual production is only in the low hundreds of thousands while other customers are also waiting, then allocation becomes a strategic decision. YFarmX reported that DeepSeek has asked Beijing to help persuade Huawei to allocate more chips to it, and faster [1]. In that scenario, the bottleneck is not simply whether Chinese AI chips can be designed, but whether they can be manufactured, packaged, and delivered at the volume required by frontier AI labs.

High-bandwidth memory is especially important. Large model inference is often constrained by how quickly model weights and context data can move through memory. That makes HBM capacity and bandwidth central to serving economics. Tom’s Hardware Italia reported that each Ascend 950DT integrates 144 GB of HBM and reaches 4.0 TB/s of bandwidth, while also noting that annual production constraints could stretch delivery timelines [4]. For DeepSeek, the result is a race between model demand and hardware availability.

Nvidia remains inside the training stack

The most important nuance is that Nvidia remains essential to DeepSeek’s training workflow. Mint reported that DeepSeek previously explored training with Huawei processors but has continued to rely on Nvidia accelerators for that critical process [3]. TechTimes likewise framed the order as “inference moves to sovereign hardware” while training still depends on Nvidia [2].

That split undercuts any simplistic reading that Huawei has displaced Nvidia across DeepSeek’s stack. What appears to be happening is more targeted: DeepSeek is moving the cost-heavy serving layer toward Chinese chips while preserving Nvidia for model creation and improvement. This lets DeepSeek reduce exposure where Chinese hardware is most usable today without risking the stability of major training runs.

It also reflects the strategic impact of export controls. Nvidia’s most advanced chips have faced severe restrictions in China, while selected China-bound products have depended on licensing and policy decisions [3]. For Chinese AI companies, a domestic inference fleet offers resilience even if access to Nvidia hardware remains uncertain. But resilience is not the same as parity. The continued use of Nvidia for training suggests that software maturity, ecosystem depth, and performance still matter as much as national chip availability.

What this means for the AI hardware race

DeepSeek’s reported 160,000-chip Huawei plan shows that China’s AI hardware shift is moving from demonstrations to large procurement decisions. It also shows that the transition is uneven. Domestic accelerators are being trusted for production inference at enormous scale, while Nvidia remains the preferred or necessary option for frontier training workloads [1][3].

For Huawei, the order would be a validation of the Ascend roadmap if it can be fulfilled. For DeepSeek, it would provide more control over serving capacity and potentially lower the marginal cost of user traffic. For Nvidia, the story is mixed: it reinforces the pressure on its China business, but also demonstrates that its hardware remains deeply embedded where training performance matters most.

The clearest conclusion is that AI infrastructure is becoming bifurcated by task. Inference can move faster toward domestic Chinese platforms because it is continuous, monetizable, and easier to optimize around memory-rich accelerators. Training remains harder to dislodge because it depends on a mature hardware-software stack that can survive massive synchronized runs.

If the Ulanqab deployment proceeds, DeepSeek will not have built a post-Nvidia AI company. It will have built something more subtle: a hybrid compute architecture in which Huawei handles the visible scale of daily model serving, while Nvidia continues to support the invisible work of creating the models those users interact with.

Sources from the last 72 hours

1. [1]DeepSeek plans at least 160,000 Huawei Ascend chips in Inner MongoliaSep 5, 2026, 12:00 AM UTC
2. [2]DeepSeek’s 160,000-Chip Huawei Order Puts PRC Law Over Every API QuerySep 5, 2026, 2:05 PM UTC
3. [3]China's Nvidia alternative? DeepSeek plans massive Huawei AI chip rollout at Inner Mongolia data centreSep 4, 2026, 5:35 PM UTC
4. [4]DeepSeek sceglie 160.000 chip Huawei, ma NVIDIA resta essenzialeSep 6, 2026, 3:00 PM UTC
5. [5]DeepSeek reportedly plans 160,000-chip Huawei cluster in Inner MongoliaSep 7, 2026, 12:00 AM UTC
6. [6]Huawei Supplies 160K Ascend 950DT Chips to DeepSeekSep 6, 2026, 6:53 PM UTC

AI-generated article based on recent web research, then preserved as a dated editorial snapshot.