DeepSeek, the Hangzhou-based artificial intelligence startup, is designing its own chip to reduce dependence on Nvidia and Huawei, according to three people familiar with the matter. The chip targets inference, the stage where a trained model answers user queries, rather than training, which is more demanding and where Nvidia’s lead is widest due to U.S. export controls. Inference is more forgiving on manufacturing and sensitive to serving cost, a metric a specialized chip can improve. In May, DeepSeek cut the price of its V4-Pro model by 75%, dropping the top rate to under $0.85 per million tokens from $3.30. A purpose-built chip could lower costs further. DeepSeek designs both the model and chip, allowing co-tuning that general-purpose hardware cannot easily match. Around a recent model release, it used a data format suited to home-grown chips, hinting at this co-design approach. Chinese firms plan to shift 46% of their AI-accelerator budgets to domestic suppliers within a year, according to industry estimates, as Nvidia’s position in China has collapsed under export restrictions.