Chinese company DeepSeek is preparing an order that could make history in China’s chip industry. According to Bloomberg, citing sources familiar with the company’s plans, DeepSeek intends to deploy at least 160,000 Huawei Ascend 950DT accelerators in a new data center in Ulanqab, Inner Mongolia. If the plan is implemented, it will create one of the largest known clusters of the brand’s AI chips.
Chips Huawei Is Not Even Selling Yet
The Ascend 950DT is the most powerful product in Huawei’s portfolio, and the company plans to officially launch it at the end of this year. It will offer 144 gigabytes of its own high-speed memory and bandwidth of four terabytes per second. These specifications put it close to top-tier American accelerators, although it does not quite match them.
Huawei designed this chip primarily for model training, the most demanding part of artificial intelligence development. DeepSeek, however, plans to use it differently and intends to deploy the hardware exclusively to run finished models—that is, to generate responses for ordinary users and corporate customers. New systems will continue to be trained on Nvidia hardware, which the company has relied on until now.
DeepSeek previously attempted to train its earlier models on Huawei chips, but the experiment was unsuccessful. At the time, the software stack and interconnections between the chips were not ready for such a workload, prompting the company to return to American hardware. The current division of tasks, with Nvidia handling training and Huawei handling inference, therefore appears more like a practical solution than a demonstration of China’s technological self-sufficiency.
Ulanqab: Wind, Sun, and Free Cooling
The site of the new data center in Ulanqab lies approximately 350 kilometers northwest of Beijing and offers exactly what large computing centers require. This includes inexpensive electricity from wind and solar sources, plenty of available land, and a cold climate. The average annual temperature in the area is around four degrees Celsius, significantly reducing the cost of cooling servers.
In its final phase, the entire complex is expected to operate with a capacity of around one gigawatt. At full load, it would therefore consume the same amount of electricity as 750,000 households. DeepSeek plans to own some of the capacity directly and lease the rest from other operators. The aforementioned 160,000 Huawei chips will cover only part of the entire complex. No one has yet disclosed what the company will use to fill the remaining capacity.
At least part of the data center is expected to begin operating at the end of 2027 or the beginning of 2028. However, this timeline depends primarily on Huawei’s capabilities.
Huawei and Its Chips
Hardware supply represents the greatest uncertainty surrounding the entire project. Huawei currently lacks sufficient capacity to fulfill an order of this scale in a short period of time. Production of the Ascend 950DT this year is being constrained by a shortage of components, particularly the highest-performance memory modules. Total production volume is expected to be in the low hundreds of thousands of units per year. The company must also satisfy other customers that have expressed interest in these chips. Complete delivery of DeepSeek’s order may therefore take more than a year.
Nevertheless, Huawei expects its revenue from AI chip sales to rise significantly this year, specifically from $7.5 billion in 2025 to approximately $12 billion. US export restrictions intended to slow China’s technological progress have thus paradoxically cleared its domestic market of overseas competition for Huawei’s chip division.
The Price of Chinese Hardware
The fact that this is a technological compromise was acknowledged by company founder Liang Wen-feng himself at a summer meeting with investors. According to a transcript obtained by the media, he said that anything a top-tier Nvidia system can do can also be handled by a Huawei compute node, including with comparable response times. However, the cost of this solution is substantial. Four Huawei chips are required to do the work of a single American accelerator, and according to him, Huawei is two years behind technologically.
DeepSeek has chosen slower, less efficient, and less readily available hardware, even though older Nvidia models could still be obtained in China. In recent months, however, the government has repeatedly signaled to Chinese companies that they should not rely on American chips. State-funded data centers are also required to use exclusively domestic technology. Nvidia itself already describes the Chinese market as lost and assumes zero revenue from the region in its long-term forecasts.
The city of Ulanqab hosts dozens of similar projects, and more than 90 percent of its local computing capacity is used specifically for artificial intelligence. Data centers are relocating there from the densely populated east of the country as part of a government plan promoted by Beijing under the name East Data, West Computing.
DeepSeek has secured funding for the construction. In June, the company raised external capital for the first time in its history, obtaining more than $7 billion, with investors including Tencent and battery manufacturer CATL. Until then, the company had built its models primarily around maximum efficiency and low costs, but it is now investing enormous sums directly in infrastructure, cabling, and cooling units.
Sources: economictimes.indiatimes.com, thetechnologyexpress.com and thenextweb.com
