DeepSeek Ltd., a prominent player in China’s artificial intelligence sector, is working on designing its own custom silicon chips specifically for AI inference tasks, according to industry sources. Although still in the early stages, the company aims to lessen its dependence on external suppliers like Nvidia and Huawei by bringing critical chip technology in-house.

The focus on inference chips is especially significant because these processors handle the stage where AI models generate responses to users in real time. Controlling this segment of the AI workflow is considered crucial for companies to manage costs and sustainably commercialize their products. While DeepSeek initially gained recognition for its AI training capabilities, inference represents the ongoing operational expense directly linked to user interaction and revenue generation.

Since its breakthrough with the R1 AI model, which initially ran on Nvidia’s H800 chips tailored to the Chinese market, DeepSeek has faced supply challenges. These Nvidia chips were subject to export restrictions imposed by the United States government, prompting the company to shift toward Huawei’s Ascend GPU lineup for more recent models like its V4. This shift coincided with a spike in demand for Huawei’s Ascend 950 chips among Chinese technology firms.

Despite Huawei’s current dominance as a chip provider in China, DeepSeek is reportedly seeking to avoid becoming overly dependent on any single supplier. To that end, it has engaged in extensive talks with chip design firms, foundries, and memory manufacturers, while recruiting experienced chip engineers to build internal expertise. Its strategy aligns with trends in the U.S. AI industry, where firms such as OpenAI and Anthropic have begun developing their own inference chips to gain greater control over technology stacks.

OpenAI recently introduced an inference processor named Jalapeño, created in partnership with Broadcom, with similar objectives: reducing reliance on Nvidia hardware and enhancing proprietary control over the computing infrastructure behind AI services. This approach resembles the model used by tech companies like Apple, which designs custom chips to optimize performance and integration across products.

Experts note that this shift toward custom silicon development reflects broader industry pressures, as data center computing resources become more constrained worldwide. Having tailored chips for inference workloads could provide companies like DeepSeek with competitive advantages in cost efficiency and performance consistency.

Overall, DeepSeek’s move to develop in-house inference chips underscores a growing trend for AI companies to consolidate critical hardware capabilities amid geopolitical tensions and supply chain uncertainties.