DeepSeek reportedly plans to train AI on Huawei chips

DeepSeek plans to begin pretraining on Huawei chips, according to The Information, while also continuing to pursue Nvidia chips in parallel. The outlet reported this in a report dated October 19, 2026. Neither Huawei nor DeepSeek has confirmed the information.

What is the plan?

According to the report, DeepSeek will begin what is known as pretraining on Huawei's Ascend 950DT chip as early as the end of 2026. Pretraining is the first and most demanding phase, in which an AI model learns from enormous amounts of data. According to the report, the actual start date depends on stability tests that DeepSeek is still running.

According to the report, Huawei plans to launch the Ascend 950DT in the fourth quarter of 2026. The chip is said to have 144 GB of its own memory of the HiZQ 2.0 HBM type and a memory bandwidth of 4 TB/s. Bandwidth indicates how quickly data can be moved in and out of memory.

The report also says Huawei plans to give DeepSeek first priority on most of the next generation, the Ascend 960DT, which is expected in the first quarter of 2027.

Closer collaboration

According to the report, Huawei engineers are stationed at DeepSeek and are feeding DeepSeek's feedback into the design of upcoming Ascend chips.

Major caveats

Nvidia hardware remains the primary tool for heavy training, and according to the report, DeepSeek is also continuing to pursue Nvidia chips in parallel.

The report points out that training a leading model may require more than 10,000 Huawei chips computing together nonstop for several weeks. Huawei has not yet publicly documented that the chips are reliable enough for this.

An earlier attempt to train the DeepSeek model R2 on Huawei chips stalled because of technical problems, according to the Financial Times via Dr. Web. According to earlier reporting relayed by Ecosistema Startup, DeepSeek went back to Nvidia for training, while Huawei chips are used for inference, meaning when a finished model responds to users.

A Spanish source, Ecosistema Startup, also states that Nvidia's B300 has roughly 7 times the raw computing performance of Huawei's Ascend 950. The figure comes only from the Spanish article and not from The Information's report.

The three retellings of the story (AI Weekly, Dr. Web and Ecosistema Startup) are all based on The Information's report and are not independent confirmations. For now, it is therefore a plan, not something that has happened.