DeepSeek Develops Its Own AI Chip as Hardware Race Heats Up
Chinese AI startup DeepSeek is building its own inference chip, moving beyond model development into custom hardware. The move signals a broader shift in AI competition toward silicon and infrastructure control.
DeepSeek, the Chinese AI startup that rattled the industry earlier this year with its low-cost models, is now building its own chip.
The company is developing an inference chip, designed for the stage where trained AI models generate responses to user queries. The project marks a significant shift for DeepSeek, which built its reputation on software efficiency rather than hardware.
Reuters reported the development on July 7, 2026, citing people familiar with the matter. DeepSeek has not made a public announcement.
The move puts DeepSeek in a growing group of AI companies trying to reduce their dependence on Nvidia, which currently dominates the market for AI chips. Anthropic has entered early talks with Samsung for custom chip manufacturing. Amazon, Google, and Microsoft have all developed proprietary silicon for their data centers.
For DeepSeek, building its own chip could help the company scale its models more cheaply and avoid the export restrictions that limit Chinese companies' access to Nvidia's most advanced processors. The U.S. government has blocked sales of high-end chips to China, pushing Chinese AI firms to find alternatives.
Inference chips are different from training chips. Training requires massive parallel computation over weeks or months. Inference needs to be fast and efficient, delivering responses in milliseconds. The two tasks have different hardware requirements, and several startups have found success building chips optimized specifically for inference.
DeepSeek's models have already attracted significant attention from U.S. developers. Data from OpenRouter shows that Chinese AI models, including DeepSeek's, accounted for more than 30 percent of weekly token usage by U.S. companies in recent months, peaking at 46 percent.
If DeepSeek succeeds in building competitive inference hardware, it could further reduce costs and make its models even more accessible to developers worldwide.
