Chinese startup DeepSeek is reportedly developing its own custom chip to run its viral AI models. The project focuses explicitly on inference processing—the computing phase responsible for generating user responses—in an effort to reduce the company’s operational reliance on Nvidia and Huawei hardware. The company has quietly expanded its internal team of chip-design engineers over the past year while initiating discussions with foundry and memory partners.
DeepSeek is ready to transition into a hardware player. The firm aims to break its heavy reliance on processors from tech giants Nvidia and Huawei with its next move. More specifically, reports indicate the startup has embarked on a strategic shift to develop its own custom AI chip.
According to Reuters, the early-stage project has been underway for about a year. To support the secret initiative, the company has quietly ramped up its recruitment of specialized chip-design engineers, bypassing public job boards to keep its hiring pipeline strictly confidential. The startup is also following in the footsteps of other major AI names like Google, OpenAI, Anthropic, and even Meta.
Related: Building an On-Demand Medicine Delivery Service Like CVS and Capsule
The organization is not building general-purpose processors to train massive LLMs from scratch. Instead, the engineering team is focusing specifically on “inference” chips. Inference is the operational computing phase where an already-trained model processes user queries and instantly generates answers, translations, or code.
As AI applications gain massive global adoption, running these everyday models has become the fastest-growing and most expensive segment of AI infrastructure. Developing specialized, highly focused chips allows companies to process user queries at a fraction of the cost. They consume significantly less power than general-purpose graphics processing units (GPUs).
To bring the project to life, DeepSeek is currently negotiating with external chip-design firms, foundries, and memory manufacturers. The massive undertaking coincides with the startup recently reversing its long-standing policy against outside funding. The startup is now moving toward a massive $7 billion investment round that values the company at over $52 billion.
Building independent hardware puts the organization in line with a broader global movement among top-tier software labs. For example, OpenAI recently introduced Jalapeño, a custom inference chip built alongside Broadcom. Rival Anthropic is reportedly in talks with Samsung for its own AI hardware.
Related: Lab Diamond a good option?
However, DeepSeek faces a unique set of geopolitical hurdles. Strict export controls prevent Chinese organizations from buying Nvidia’s cutting-edge graphics cards. This scenario forces the startup to rely on older, modified chips or domestic hardware like Huawei’s Ascend processors. While Huawei currently commands roughly half of China’s domestic AI infrastructure market, its grip is loosening as other internet powerhouses like Alibaba and Baidu deploy independent in-house silicon. The entry of a highly aggressive competitor adds pressure to the local supply chain.
DeepSeek became a household name by proving it could deliver world-class reasoning models for a fraction of what its Western competitors spend on computing power. Taking that same efficiency-first philosophy and applying it to physical silicon seems like a logical evolution.
Designing competitive AI silicon is an incredibly complex, multi-year gamble. This is especially true when trade restrictions limit access to advanced overseas factories and high-bandwidth memory. However, if the project succeeds, they will build solid foundations to protect their profit margins from corporate bottlenecks and international trade disputes alike.
