Abstract:
Nvidia’s competitive prospects in China’s artificial intelligence chip market are being impacted by news of a large-scale purchase. Wccftech quoted Bloomberg as reporting that DeepSeek is said to have ordered 160,000 Ascend 950DT chips from Huawei for its new 1GW data center in Inner Mongolia. The report believes that this order further highlights Huawei’s competitive advantage in China’s local AI market and puts Nvidia’s special products launched for the Chinese market under greater pressure.

According to the market price cited in the report, each Ascend 950DT chip is priced at about 111,000 yuan, equivalent to about 16,000 US dollars. If all 160,000 pieces are purchased at this price, the total order value may reach approximately US$2.56 billion. In comparison, Nvidia's H20 GPU for the Chinese market is said to be priced between $15,000 and $25,000. The report pointed out that DeepSeek’s choice of Huawei this time was not purely due to price factors.
This batch of chips will be used in DeepSeek’s new 1GW data center in Inner Mongolia, mainly to meet the inference computing needs. The report also pointed out that Huawei’s current Ascend 950DT production capacity is still limited, with annual output of only a few hundred thousand units, so it may take up to a year to complete this order. For AI companies that need to quickly deploy computing power, such a delivery cycle is undoubtedly an important test.
Ascend 950DT uses Huawei's self-developed third-generation DaVinci architecture DaVinci v3, and uses SMIC's N+3 process, which is an advanced 7-nanometer process. The chip can handle both training tasks and inference tasks, but DeepSeek’s main purpose of this purchase is obviously to meet inference needs. Each chip is equipped with 144GB of dedicated high-bandwidth memory, with a memory bandwidth of 4.0TB/s, and an interconnection speed between GPUs of 2.0TB/s. It also natively supports FP8, FP4 and Huawei’s own HiF8 low-precision format.
Huawei will also use Ascend 950DT for the Atlas 950 SuperPoD infrastructure, which can connect up to 8192 chips and achieve a peak AI computing power of 8 EFLOPS. Reports believe that this large-scale cluster solution will become an important basis for Huawei to compete in China's AI computing power market.
DeepSeek founder Liang Wenfeng has previously compared the computing power products of Nvidia and Huawei. He said in July that Huawei super nodes can also complete the tasks that Nvidia GB300 can complete, and there is no obvious difference in latency and other aspects. However, he also pointed out that Huawei needs about 4 GPUs to reach the performance level of 1 GPU from Nvidia, and is about two years behind in technology. This statement also shows that DeepSeek does not believe that Huawei chips have completely surpassed Nvidia, but makes a choice after comprehensively considering cost, supply chain and localization factors.
The report quoted Sphere Semi CEO Steven Glinert as saying that Nvidia now seems to have begun to regard the Chinese market as a "lost market" and no longer strives for it as actively as it did in the past. Glinert said that Nvidia used to attach great importance to the Chinese market and would even push those who took a tough stance on China to leave government or think tank positions, but now the attitude seems to have changed.
However, this statement still belongs to the judgment of industry insiders and does not mean that Nvidia has officially announced that it will abandon the Chinese market. Whether Nvidia will continue to provide special chips to the Chinese market and whether it will launch new Chinese versions of its products in the future remains to be further confirmed by the company.
From a market competition perspective, DeepSeek’s purchase of Huawei Ascend 950DT does have certain symbolic significance. The report pointed out that DeepSeek’s choice of Huawei chips means that local AI companies may still give priority to domestic solutions even if the performance, delivery cycle and price differences are not completely favorable. For Nvidia, this trend may be more noteworthy than the single order itself.
At the same time, Inner Mongolia is becoming an important area for China's AI infrastructure construction. The local cold and dry climate is conducive to data center heat dissipation and is also suitable for the development of solar power supply. The report mentioned that the electricity cost of Alibaba’s data center in Ulanqab is approximately 0.32 to 0.35 yuan per kilowatt hour, equivalent to approximately US$0.05. ByteDance is also currently said to be negotiating to add 5 to 6 GW of data center capacity in Inner Mongolia.
According to reports, the construction cost of each 1GW of data center capacity in Inner Mongolia is currently approximately US$23.8 billion. As more large-scale AI companies deploy locally, Inner Mongolia may continue to become an important center for China's AI computing power construction.
Overall, DeepSeek’s large-scale purchase of Huawei Ascend 950DT further illustrates that China’s AI industry is accelerating the construction of a local computing power system. Although Nvidia still has strong technical and ecological advantages, the competitive environment it faces in the Chinese market has changed. In the future, as domestic chip production capacity increases, the software ecosystem improves, and more AI companies adopt domestic solutions, NVIDIA's share and strategic position in the Chinese market may be continuously affected.
Comments