DeepSeek makes Huawei training chips a strategic priority

DeepSeek chief executive Liang Wenfeng has told investors that making large-scale training work on Chinese chips is now a major priority, according to people who heard the discussions. The company expects Huawei training accelerators to become usable as early as the fourth quarter. That is a plan and a technical forecast, not evidence that Huawei already matches Nvidia for frontier training.

The hard part is the system around the silicon. Training requires thousands of accelerators to communicate reliably for weeks; networking, memory bandwidth, compiler software and failure recovery can matter as much as a chip's peak arithmetic. US controls have therefore constrained more than DeepSeek's ability to buy a particular processor: they have forced Chinese laboratories to redesign models and infrastructure around a less mature stack. DeepSeek's willingness to do that gives Huawei a demanding anchor customer and provides Beijing with a practical route for improving domestic hardware.

The broader Chinese effort is becoming less dependent on one company. At Alibaba's Apsara conference on 22 September, the group announced the Zhenwu V900 accelerator, which it says offers three times its predecessor's performance and should enter mass production in early 2027. Alibaba also described a future Qwen model with 5tn–10tn parameters and a goal of more than 20GW of data-centre capacity by 2032. Those are corporate targets rather than demonstrated capacity, and parameter count alone does not establish model quality.

Synthesis: DeepSeek is attempting to make an external supplier's chips work at frontier scale; Alibaba is developing chips, cloud capacity and models inside one corporate group. Both approaches turn US export controls into an industrial-policy accelerator. Their test is not a launch announcement but whether domestic systems can train competitive models economically and repeatedly.