Key points
- DeepSeek plans at least 160,000 Huawei Ascend 950DT chips for a new Inner Mongolia site
- The chips are for running models, not training them. Training stays on Nvidia
- Huawei can make only a few hundred thousand 950DTs this year, Bloomberg reports
- Partial operation targeted for late 2027 or early 2028
DeepSeek wants to fill a new data center in Inner Mongolia with at least 160,000 of Huawei's Ascend 950DT chips, Bloomberg News reported Friday. The number is what got the headlines. The chips would run DeepSeek's models for users, rather than train them. DeepSeek still trains on Nvidia (NVDA), and Huawei's production capacity will decide how quickly any of this gets built.
So it's a real story about China building AI infrastructure without Nvidia, and it's also a story about a chip Huawei said in June it would begin deploying in August. Bloomberg's Mackenzie Hawkins and Haze Fan reported the plan, citing people familiar with it. DeepSeek did not respond to the news agency's request for comment, and a Huawei spokesperson had no immediate comment.
What DeepSeek is building
The site is in Ulanqab, about 350 kilometers northwest of Beijing, and it's part of a project of about one gigawatt, according to Bloomberg. At full capacity that's enough power for about 750,000 homes. DeepSeek wants at least part of it running by late 2027 or early 2028.
Bloomberg said this would be one of the largest known clusters of Huawei AI chips anywhere. It also reported something I find more interesting than the chip count. DeepSeek has asked Beijing to help persuade Huawei to send it more chips, and sooner. You don't ask the government to lean on your supplier when the supplier has plenty.
Why the chips run models instead of training them
Huawei built the 950DT for both jobs. At Huawei Connect last September, rotating chairman Eric Xu said the chip was "optimized for both the decode stage of inference and for model training," with 144 GB of memory and 4 TB/s of memory bandwidth from Huawei's own HBM, which it calls HiZQ 2.0. HBM is the stacked memory that sits next to an AI chip and feeds it data, and it's the part of the supply chain that's short everywhere right now.
DeepSeek currently plans to use the chips for inference only. Bloomberg reported the company "doesn't currently plan to use the 950DT chips for training," citing one of the people. Training builds and updates the model; inference runs it for users. Inference is the part Chinese labs have been moving onto domestic chips first. Zhipu said in its August 31 interim results that its models on domestic chips now match Nvidia hardware on cost per token. Zhipu is claiming cost parity and DeepSeek is reportedly planning a deployment, which are different claims, but both labs are moving inference toward domestic hardware.
Can Huawei actually deliver?
Bloomberg said component shortages, including high-end memory, "will cap Huawei's output of the 950DT at the low hundreds-of-thousands this year." Huawei also has other customers at home and abroad seeking the chips. DeepSeek's planned deployment alone would require a substantial share of Huawei's projected 950DT output this year.
And the chip is new. Xu's roadmap had the 950DT arriving in the fourth quarter of this year, after the 950PR in the first quarter. In June a Huawei vice president, Chen Lin, said it would reach Huawei Cloud in August, ahead of schedule, TrendForce reported, citing Chinese outlets. The same roadmap puts an Ascend 960 in the fourth quarter of 2027 and an Ascend 970 a year after that. If Huawei meets its roadmap, a newer generation could be available by the time the site starts operating.
Huawei's favorite comparison is at the system level, not chip against chip. Its Atlas 950 SuperPoD strings together 8,192 Ascend chips, and Huawei has said the system could beat Nvidia's NVL576 planned for 2027, per TrendForce. That's Huawei grading Huawei. I'd wait for someone else's benchmark.
What this means for Nvidia (NVDA)
The report doesn't establish an immediate revenue impact for Nvidia. The planned deployment would use chips Huawei hasn't produced in volume, at a site targeting partial operation in late 2027 or early 2028, and DeepSeek currently continues to train on Nvidia. Nvidia closed at $230.35 on Friday, up 0.8 percent, according to Robinhood market data. Bloomberg framed the plan as a step in "China's efforts to replace Nvidia Corp." That's fair, as long as you read "efforts."
When Moonshot's Kimi K3 launched in July, we looked back at the first DeepSeek moment in January 2025 and found chip demand went up afterward, not down. This one's a different animal, because nobody built a cheaper model on Nvidia chips this time. Nvidia's best-known Chinese customer is planning to run its models on a Huawei chip, and the central question is whether Huawei can make enough of them. Bloomberg said it plainly: "the timeline for installation will depend on Huawei's production capabilities." If you own NVDA, that's the sentence to watch, not the 160,000.
