Friday, November 20, 2026

Understanding China. Seeing what’s next.

Understanding China. Seeing what’s next.

Huawei engineers have moved into DeepSeek’s office. The target is Nvidia

Huawei engineers have moved into DeepSeek's office. The target is Nvidia

Huawei engineers are working out of DeepSeek’s offices and feeding the AI lab’s requirements directly into the design of future Ascend chips, The Information reported on October 19, citing people familiar with the arrangement. The plan is for DeepSeek to begin pre-training a model on Huawei’s Ascend 950DT as early as the end of 2026. Neither company has commented.

The collaboration is unusually hands-on. According to the report, the two sides are working on how computational tasks are scheduled across chips and how data moves between them, the plumbing that decides whether a cluster of accelerators behaves like one big computer or thousands of small ones. The 950DT offers around 1 petaflop of FP8 compute, 144 GB of Huawei’s own high-bandwidth memory and 4 TB/s of memory bandwidth.

DeepSeek has tried this before and lost. Its attempt to train the R2 model on Huawei hardware ran into chip stability problems, communication failures between chips and gaps in software support, and the company finished the job on Nvidia hardware. That history is why the current timetable is described as provisional. Testing so far has involved more than 1,000 chips. A full training run could need tens of thousands, and some of the software required to manage a cluster of that size is reportedly unfinished.

Huawei, for its part, is committing supply. The report says most of the company’s Ascend 960DT output in the first quarter of 2027 is earmarked for DeepSeek. That is a bet by China’s biggest chip designer that its most important AI customer will stay on its hardware this time.

The sceptical reading is that the heavy lifting of training new models still happens on whatever Nvidia hardware DeepSeek can get, and that Huawei chips will move into inference first, where the demands are gentler. The optimistic reading is the one Beijing prefers: a frontier lab and a domestic chipmaker designing each other’s next generation together, which is roughly how Nvidia and OpenAI got where they are. Which reading is right will be visible by January, in whether a 950DT training run actually starts.

Sources

Leave a Reply

Your email address will not be published. Required fields are marked *