160,000 Huawei Chips for DeepSeek: Inside the Order That Could Reshape China's Compute Endgame

DeepSeek — the Hangzhou lab whose lean model releases have repeatedly reset expectations for what frontier-adjacent AI costs — is about to make its biggest infrastructure bet yet, and it is betting entirely on domestic silicon. According to a Bloomberg report published on September 4, 2026, the company plans to deploy at least 160,000 of Huawei’s Ascend 950DT accelerators at the gigawatt-scale data center it is building in Ulanqab, Inner Mongolia.

If the installation is completed as envisioned, it would become one of the largest known clusters of Huawei AI chips anywhere — and arguably the single most concrete datapoint yet in China’s effort to replace Nvidia at home. The people familiar with the matter, who spoke on condition of anonymity to discuss a confidential deal, cautioned that the timeline depends heavily on Huawei’s ability to actually produce and deliver the chips.

What the report actually says

The details matter here, because they sketch the shape of a deal that is enormous in ambition but deliberately constrained in scope:

• Scale: At least 160,000 Ascend 950DT chips, described as Huawei’s top next-generation accelerator, deployed at DeepSeek’s 1GW Ulanqab facility roughly 350 kilometers northwest of Beijing.
• Inference, not training: DeepSeek intends to use the 950DT fleet to operate its models — serving users, running agentic workloads — not to train them. That is notable because Huawei designed and marketed the 950DT for that more demanding training function. DeepSeek tried to train previous models on Huawei silicon and has so far relied on Nvidia accelerators for that crucial step.
• Supply is the bottleneck, not intent: DeepSeek reportedly wants to buy even more of Huawei’s best chips, but Huawei does not currently have the production capacity to satisfy the demand. Shortages of components such as top-end memory are expected to cap 950DT output at the low hundreds-of-thousands of units this year.
• Delivery could take more than a year: Huawei is balancing demand from other clients and attempting small export volumes overseas, so fulfilling DeepSeek’s order may stretch well past twelve months.
• Beijing is involved: DeepSeek has asked for the government’s help in convincing Huawei to allocate more chips to the startup — and faster. China’s Ministry of Industry and Information Technology did not respond to a request for comment.

Both DeepSeek and Huawei declined or did not respond to requests for comment, which is standard for deals of this nature at this stage.

Why 160,000 chips is a meaningful number

To understand why this order is significant, it helps to look at the trajectory of publicly known Huawei clusters.

Just six months ago, an official newspaper in tech hub Shenzhen reported the launch of the country’s first-ever 10,000-chip intelligent computing cluster built on Huawei’s Ascend series. Compared to data centers in the US, that was a tiny facility. For perspective, Elon Musk’s SpaceXAI brought a cluster ten times that size online back in 2024, and Western hyperscalers now routinely build campuses with six-figure GPU counts.

A 160,000-chip installation would put DeepSeek’s ambitions squarely in the ballpark of those Western builds — albeit, as Bloomberg notes, filled with processors that are less capable than Nvidia’s current flagship silicon. The Ascend 950DT, unveiled roughly a year ago, is an upgraded version of Huawei’s marquee accelerator and is widely regarded as comparable in power to Nvidia’s previous Hopper generation — not Blackwell, and not whatever comes after.

One person with knowledge of Huawei’s production capabilities told Bloomberg that by the time the Ulanqab hub comes online, China could have several clusters of comparable size. That is the quietly ambitious part of the story: the first one is a milestone, but the expectation of more suggests an industrial plan rather than a one-off procurement.

Context: supply chains, sanctions, and the 1.6 million dies

This order does not exist in a vacuum — it sits at the intersection of US export controls, Huawei’s constrained manufacturing, and Beijing’s sovereignty push.

Washington first imposed export controls on advanced AI chips to China in 2022, over concerns the hardware could boost China’s military. Nvidia has argued that the transition to homemade silicon now underway in China is the natural consequence of those years of restrictions. Huawei, sanctioned and cut off from leading-edge foundries, plans to make around 1.6 million Ascend dies in 2026, per earlier Bloomberg reporting — enough base components for roughly 600,000 of its marquee 910C chips plus a smattering of other models. Memory shortages add another ceiling on 950DT volumes.

Meanwhile, the politics on both sides are complicated. While President Trump’s team was prepared to allow the sale of as many as a million Nvidia H200 chips to China, Beijing has so far approved only a fraction of that total for import — effectively forcing labs toward domestic hardware even where US rules would permit the alternative. Chinese private sector giants like Alibaba and Tencent still prefer Nvidia and have placed large orders when restrictions eased. And Washington has accused both Moonshot and DeepSeek of using smuggled Nvidia hardware; Trump administration officials have further alleged that DeepSeek procured Nvidia’s top Blackwell chips and installed them in Inner Mongolia, a claim Bloomberg says it has not independently verified.

Against that backdrop, DeepSeek’s move reads as a hedge: a deliberate, state-negotiated commitment to domestic inference capacity at scale, even while training continues on whatever Nvidia silicon it can access.

What it means

For DeepSeek, the deal secures the operational floor for its next phase. The company is reportedly in talks to raise billions of dollars for infrastructure buildout, and the Ulanqab hub is described as key to driving its AI development. A 1GW facility at full utilization would consume enough energy to power roughly 750,000 homes — and the 160,000 950DT chips would comprise only one chunk of that capacity, leaving the remainder’s silicon mix an open question.

For Huawei, it is the clearest proof-of-demand yet for its flagship accelerator, and a test of whether its supply chain can convert intent into installed capacity.

And for the broader US–China AI race, the order marks a shift from rhetoric to procurement. China’s compute sovereignty push has long been a policy slogan; a six-figure Ascend cluster from the country’s most influential AI lab — negotiated with Beijing’s encouragement and constrained only by manufacturing, not demand — is the strongest signal yet that the slogan is being taken seriously. The gap between Chinese and US chips remains real, and memory shortages will keep Huawei’s output capped for now. But if several clusters of this size really do come online over the next few years, the global map of AI compute will look meaningfully different.