DeepSeek wants to build a large chip cluster with Huawei, but for what?

Chinese AI company DeepSeek plans to deploy at least 160,000 of Huawei's Ascend 950DT AI chips in a data center in Ulanqab, Inner Mongolia. Bloomberg reported it first, citing anonymous people familiar with the matter. Neither DeepSeek nor Huawei is quoted on the plans.

According to Bloomberg's story, it is one of the largest known clusters of Huawei AI chips. It is described as part of China's effort to become less dependent on the American chipmaker Nvidia. How quickly the chips can be installed depends, according to Bloomberg's sources, on how much Huawei can produce.

Sources disagree on the purpose

The picture is not clear-cut here. Bloomberg wrote that DeepSeek did not plan at the time to use the chips for training, even though Huawei has developed and marketed them for that purpose. According to Bloomberg, the chips would instead run the company's models. This is called inference: a finished model answering users' questions.

The news site Dataconomy, on the other hand, wrote on November 4, 2026, that the chips will be used to train DeepSeek's next model, V5. Dataconomy says operations at the data center are set to begin in late 2026, and that DeepSeek will first evaluate the hardware for inference and then for pretraining, the first and most demanding part of training a model. Dataconomy cites no original source.

A major strategic bet

AI Weekly relays reports from The Information. According to them, DeepSeek founder Liang Wenfeng has told investors that training on Huawei chips is one of the company's biggest strategic bets. The Information also wrote that DeepSeek could begin pretraining on the Ascend 950DT as early as late 2026, and that the chip is expected to launch in the fourth quarter of 2026.

According to the same report, Huawei will prioritize DeepSeek for most of the next chips, the Ascend 960DT, which are expected in the first quarter of 2027. Huawei engineers are to sit in DeepSeek's offices and use the company's feedback in the chip design.

Earlier problems

Dataconomy writes that DeepSeek has so far used Nvidia hardware for its most compute-intensive tasks. The company is said to have run into problems with stability and performance when it trained on earlier generations of Huawei Ascend.

According to AI Weekly, training a leading AI model can require more than 10,000 Huawei chips working together for weeks. Huawei has not publicly shown that its chips can meet that level of reliability.

For readers, this means the plans are a signal that China wants to build AI without American chips. Whether it will succeed in training large models remains unresolved.