DeepSeek to Deploy 160,000 Huawei Ascend Chips at Inner Mongolia Data Center for AI Inference

China's most prominent open-model lab is placing the largest known bet on domestic silicon, and the move signals how far the country has come in its effort to build an AI stack that does not depend on Nvidia. Bloomberg reported on Sep 4 that DeepSeek plans to deploy at least 160,000 Huawei Ascend 950DT accelerators at a new data center it is building in Ulanqab, Inner Mongolia, a cluster that would be among the largest concentrations of Huawei AI chips ever assembled and that the company intends to use primarily for running its models rather than training them. The report, which neither DeepSeek nor Huawei commented on as of Sep 4, landed as American and Chinese AI ecosystems were consolidating in opposite directions, with Beijing pushing domestic silicon into frontier workloads at the same time Washington was tightening the export regime that made such substitution necessary.

The scale of the planned deployment is the story. Just six months ago, China's first 10,000-chip Huawei Ascend cluster began operations, and industry trackers had treated that as a milestone in the country's drive to substitute domestic accelerators for the Nvidia hardware that US export controls restrict. DeepSeek's reported plan is an order of magnitude larger, roughly 16 times the size of that first flagship cluster, and Bloomberg reported on Sep 4 that the associated data center is about 1 gigawatt in scale, with parts of it possibly beginning operations in late 2027 or early 2028. The timeline matters as much as the chip count, because Huawei's production capacity, constrained by high-end memory supply, may take more than a year to fulfill an order of this size, and the company plans roughly 1.6 million Ascend dies for all of 2026.

Key Facts

The Ascend 950DT is Huawei's next-generation inference processor, which Bloomberg reported on Sep 4 is roughly comparable in power to Nvidia's previous Hopper generation and which was unveiled about a year before this report. DeepSeek has tested Huawei silicon before, adapting its earlier V4 model, released in April 2026, to run on Ascend 950 chips, and Chinese cloud providers have since ordered hundreds of thousands of those processors. The difference now is that DeepSeek, widely seen as one of the most efficient frontier labs in the world, is committing its own infrastructure budget to a Huawei-based cluster at national scale.

The financing behind the plan is already substantial. Bloomberg reported on Sep 4 that DeepSeek is in talks to raise billions of dollars for infrastructure and that it raised roughly 50 billion yuan, about $7 billion, in June 2026. The Inner Mongolia project would consume a meaningful share of that war chest, though the company is also reported to be working with SMIC, China's leading foundry, to develop its own custom inference chip, which analysts quoted by Bloomberg on Sep 4 described as a sign that the Huawei order may be transitional rather than permanent. For now, however, DeepSeek still relies on Nvidia accelerators to train its models, even as it shifts inference workloads to domestic silicon.

The competitive context sharpens the significance of the report. Analysts quoted by Bloomberg on Sep 4 described the deployment as a potential bridge in China's broader transition away from Nvidia's CUDA ecosystem, and the Ulanqab cluster would be the most concrete evidence yet that a credible alternative to American accelerator supply can be assembled under export controls. Bloomberg also reported on Sep 4 that DeepSeek would like to purchase even more chips than the 160,000 figure, constrained only by Huawei's inability to manufacture them faster, which frames the plan as a demand-led bet on domestic silicon rather than a token compliance exercise.

Analysis

What this really means is that the Chinese AI supply chain has crossed from substitution-by-necessity into substitution-by-choice, and the implications for Nvidia are larger than the immediate revenue at stake. When Chinese labs were forced onto Huawei silicon only for inference tasks that export controls did not fully cover, the narrative was that Nvidia still owned the strategic high ground of training. DeepSeek's reported plan breaks that framing, because it treats a 160,000-chip Huawei inference cluster as a durable part of its cost structure, not an emergency workaround. The bigger picture here is that inference, not training, is where the volume economics of AI will be decided over the next several years, as models are deployed at scale and queried billions of times; if Chinese frontier labs can serve that demand on domestic chips at competitive cost, the effective ceiling on Nvidia's addressable market shrinks even if US export policy never tightens further.

The comparison with the recent past is instructive. DeepSeek's earlier V4 model, released in April 2026, was adapted to run on Ascend 950 chips, and Chinese cloud providers have since ordered hundreds of thousands of those processors, which Bloomberg reported on Sep 4 shows the ecosystem already moving toward domestic silicon before this cluster was reported. What is different about the Inner Mongolia plan is that DeepSeek, widely seen as one of the most cost-efficient frontier labs in the world, is committing its own capital to own the inference layer at scale rather than renting it from a cloud provider that happens to use Huawei hardware. Both the lab and its cloud partners are responding to the same constraint, guaranteed access to affordable compute in a country that cannot rely on American hardware, but DeepSeek's direct ownership suggests it regards domestic inference capacity as strategically durable, not a temporary expedient.

The realistic caveats are worth stating plainly. A 1-gigawatt data center in Inner Mongolia will take years to build, and Bloomberg reported on Sep 4 that fulfilling the full chip order could take more than a year because component shortages, particularly high-end memory, are capping Ascend 950DT output at the low hundreds of thousands of units this year. Huawei must also balance DeepSeek's demand against orders from Chinese cloud providers and small overseas customers, which means the 160,000 figure may be an aspiration as much as a purchase order. And the reported work with SMIC on a custom inference chip suggests that even DeepSeek does not regard Huawei as the final answer, only as the best available option on the path to a fully domestic stack.

Why It Matters

For the global AI industry, the stakes are about whether the world settles into one compute ecosystem or several. Every large Huawei cluster that trains or serves competitive open models weakens the assumption that cutting-edge AI requires access to American chips, and that assumption underpins a large share of Nvidia's valuation and of US export-control strategy. For Huawei, the DeepSeek order is a validation of the Ascend roadmap at precisely the moment the company needs a flagship customer to demonstrate that its silicon can carry frontier workloads; the company's 2026 production plan of roughly 1.6 million Ascend dies only makes sense if demand of this scale is real. For other Chinese labs and cloud providers, DeepSeek's reported commitment lowers the perceived risk of committing their own budgets to domestic hardware, which could accelerate a broader migration away from Nvidia across the Chinese market. And for US policymakers, the report is evidence that export controls are reshaping the Chinese ecosystem rather than preventing it, a distinction that will increasingly shape the debate over whether the controls are working.

Next Up

In the coming weeks, watch for official confirmation from DeepSeek or Huawei, since a public acknowledgment of the order would signal that the companies have reached agreement on price, delivery schedule, and the allocation of Huawei's constrained 2026 output. Watch also for the pace of Huawei's Ascend 950DT ramp, because production numbers will determine whether the 160,000-chip plan is completed within a year or stretches toward 2028, and for any announcements from SMIC or DeepSeek about the custom inference chip, which would indicate whether the Huawei cluster is a bridge or a destination. The most consequential signal will come from DeepSeek's next frontier model, because if the company demonstrates that a model trained on Nvidia hardware can be served entirely on Huawei silicon without a quality penalty, the case for a parallel Chinese compute ecosystem will move from plausible to proven.