Reuters reported July 7 that DeepSeek is working on its own AI chip. The part is reportedly for inference, not training.
That is a smaller claim than a training-chip program. The company is not challenging Nvidia's training stack here. It is trying to lower the cost of serving its own models.
Training silicon has to compete with Nvidia's H-series chips and the CUDA software moat around them. That is expensive for anyone and especially hard for a Chinese firm boxed in by export controls. Inference is the easier place to attack. It is the recurring cost behind every token DeepSeek serves. The reporting says the chip is being built with outside foundry and memory partners, not a captive fab.
Outside foundry partners make this a chip-design effort, not a fab strategy. DeepSeek is trying to shave per-token cost on models it already sells cheaply.
China's AI silicon market now splits three ways:
| Player | Silicon | Built for |
|---|---|---|
| Huawei | Ascend line | Training + inference |
| Cambricon | AI accelerators | Training + inference |
| DeepSeek | New in-house design | Inference only |
Huawei and Cambricon went vertical to reduce dependence on Nvidia at the training layer. DeepSeek appears to be aiming at high-volume inference, where demand is heavy and Nvidia has less room to defend premium margins.
Nvidia slipped about 1.6% premarket after the report. The move was small, but it fits the shape of the risk: Nvidia is more exposed first in cheap, high-volume inference than in frontier training.
The foundry name will show the ceiling. A design forced onto SMIC's mature nodes instead of TSMC will ship slower, run hotter, and show whether the chip can cut serving costs enough to matter commercially.