DeepSeek plans to deploy at least 160,000 of Huawei's next-generation Ascend 950DT chips at a gigawatt-scale data center in Inner Mongolia, according to people familiar with the matter cited by Bloomberg and syndicated Friday. [1] The chips would run inference, not training. Training still uses Nvidia. [1] [2] DeepSeek did not respond to requests for comment. A Huawei spokesperson had no immediate comment. [1] Do not write that 160,000 chips are installed.
A plan is not a cluster. The Inner Mongolia hub, if completed as envisioned, would be among the largest known gatherings of Huawei AI processors and a step in Beijing's effort to replace Nvidia parts. [1] It would not be a training floor. Huawei designed and marketed the 950DT for the heavier job. DeepSeek does not currently plan to use it that way. The startup tried to train earlier models on Huawei silicon and has so far stayed on Nvidia accelerators for that step. [1] Order, yield, and training-versus-inference are still three different sentences. Friday's print is the largest known planned Huawei cluster. It is still Nvidia for training. Production is the gate.
Huawei cannot fill the order on a press-cycle clock. Shortages of components such as high-end memory will cap 950DT output at the low hundreds of thousands this year, one of the people said. Huawei is balancing other clients and trying to export small volumes, so fulfilling DeepSeek's order could take more than a year. [1] Bloomberg reported last year that Huawei planned to make around 1.6 million Ascend dies in 2026, enough for about 600,000 of the older 910C plus a smattering of other models. The 950DT is an upgraded part, widely regarded as comparable to Nvidia's previous Hopper generation. [1] The Decoder notes the memory bottleneck could ease only slowly: China's CXMT is turning out small batches of HBM3E for the first time and remains three to five years behind Samsung, SK Hynix, and Micron, which are already mass-producing HBM4. [2] DeepSeek wants more of Huawei's best chips than the champion can make. The firm has asked Beijing to lean on Huawei for faster allocation. [1] China's Ministry of Industry and Information Technology did not respond to a faxed request for comment. [1]
The site is gigawatt-scale. At full utilization that would be enough energy for roughly 750,000 homes. The 160,000 Ascend 950DTs would be only one chunk of that capacity. It is unclear what would fill the rest. [1] Trump administration officials have alleged DeepSeek installed Nvidia's top Blackwell chips in Inner Mongolia. Bloomberg has not independently verified that claim. [1] Do not write it as confirmed.
CNBC-TV18's headline is the wean-off-Nvidia sale. The Decoder called the same plan the largest known Huawei chip cluster. [2] Both are running ahead of the loading dock. Hangzhou-based DeepSeek set a template for disruptive model releases later followed by firms such as Moonshot. [1] Six months ago, an official Shenzhen newspaper said the country had begun operating its first 10,000-chip Huawei Ascend intelligent-computing cluster. Meituan said in June it trained a model on 50,000 domestic chips without further detail. Z.AI completed a 1-gigawatt hall it plans to fill with Chinese-made chips; its installed base is still in the ballpark of 10,000 cards. [1] Elon Musk's SpaceXAI brought a cluster ten times that 10,000-chip size online in 2024. [1] 160,000 on a slide is not 160,000 in a rack.
Nvidia remains the global standard and is largely barred from sale to China without U.S. permission. Private Chinese buyers still prefer it. Alibaba and Tencent placed large orders after Washington eased some restrictions. Chinese firms can still rent Nvidia compute located overseas. [1] Beijing has green-lit only a fraction of a proposed million-chip H200 import. [1] DeepSeek's ask is that the state make Huawei ship faster. Huawei's constraint is that it cannot. MSM will sell China weaning off Nvidia. The Friday record is inference chips, not training, not seated, gated by output. A purchase plan is not a powered data hall.
-- THEO KAPLAN, San Francisco