DeepSeek is designing its own inference chip, Reuters says

Every 'Chinese lab designs its own accelerator' story has to be read at two speeds, and this one is no exception. Reuters is reporting, citing unnamed sources, that DeepSeek has begun work on its own inference-focused AI chip aimed at reducing its dependence on Nvidia and on domestic alternatives from Huawei. The sourcing is anonymous, so take the specifics as reported, not settled.

The move is not out of nowhere. Since early 2025 DeepSeek has been openly recruiting semiconductor design talent, in what industry outlets described at the time as a major hiring drive signalling potential plans to develop proprietary processors. The company's public chip posture has been split for a while: it trains on Nvidia's China-specific H800 and higher-end H100 GPUs, and runs inferencing workloads on Huawei's Ascend accelerators. Around the release of a recent model DeepSeek itself said it had used a data format 'well suited for home-grown chips soon to be released', which corroborating coverage read as a nod toward a domestic silicon roadmap. An in-house accelerator is the logical next step from that trail.

Why inference first, and not training, is the interesting part. Training is where Nvidia's software and interconnect lead is widest, and where China's fab constraints hit hardest under US export controls on advanced lithography. Inference is more forgiving on process node, more sensitive to per-query serving cost, and it is the workload DeepSeek runs at scale for real users. If you were going to bet on one silicon project inside a Chinese AI lab, that is the one that pays back soonest, and it is the segment where a homegrown chip could most credibly eat share from both Nvidia and Ascend rather than only from one of them.

The honest caveats are the ones the reporting itself does not resolve. In the corroborating coverage available, the sources do not name a foundry, a process node, a timeline to first silicon, or whether the chip is captive to DeepSeek's own serving stack or destined for wider Chinese use. Any one of those details would move the story from signal of intent to actual pressure on Nvidia's China inference business. Until they surface, the sensible read is that DeepSeek is joining, not yet leading, China's homegrown accelerator push, and that the next data point worth watching is which fab quietly picks up the tapeout.