DeepSeek, the Chinese AI lab whose R1 model triggered a rout in US technology stocks in January 2025, is now developing its own artificial intelligence chip designed for inference, the process of running trained models rather than building them. The effort, confirmed to Reuters by three people familiar with the matter in an exclusive report published on July 7, began approximately a year ago and remains at an early stage. DeepSeek has increased its hiring of chip-design engineers in recent months, conducting that recruitment privately without public job postings, and has been reaching out to chip-design firms, foundries, and memory companies as potential external partners.
The news sent Nvidia’s shares down approximately 1.6 percent in premarket trading, a reaction that reflects the market’s sensitivity to developments that accelerate China’s domestic chip transition. DeepSeek did not respond to a request for comment.
Why Inference, Not Training
The decision to focus on inference rather than training chips is strategically significant. Training chips, the high-end GPUs used to build large language models, require cutting-edge manufacturing processes and are the primary target of US export controls. Inference chips, which run trained models at scale, have different performance requirements and are generally less dependent on the most advanced semiconductor nodes. By targeting inference, DeepSeek is pursuing a segment where domestic manufacturing capabilities are more adequate and where the commercial need is immediate: DeepSeek’s models are being run by millions of users and thousands of enterprise customers, generating inference workloads that currently depend on Nvidia and Huawei hardware.
The move also reflects a broader pattern in China’s AI hardware strategy. Alibaba and Baidu have both developed proprietary chips for internal use, and Huawei’s Ascend series has become the dominant domestic alternative to Nvidia for training workloads. DeepSeek’s entry into chip development would add a new dimension to that ecosystem, a model developer building its own inference silicon, similar to what Google has done with its Tensor Processing Units and what OpenAI has done with its “Jalapeno” inference chip, developed with Broadcom and unveiled last month. Anthropic is also reportedly weighing whether to build its own chips, according to a Reuters report from April 2026.
The Funding Context
DeepSeek’s chip development effort is unfolding alongside the company’s first external fundraising round. The lab is reportedly raising $7 billion at a valuation of $52 to $59 billion, a round that, if completed, would give it the capital to invest in chip development at a scale not possible from operating revenue alone. The fundraising was first reported in June 2026 and has not yet been formally announced by the company.
The combination of external funding and chip development signals a transition for DeepSeek from a research-focused lab to a vertically integrated AI infrastructure company. The lab trained its R1 model on Nvidia H800 chips, a downgraded version of the H100 designed to comply with US export control thresholds, and in April 2026 adapted its V4 model to run on Huawei Ascend chips. Building its own inference chip would give DeepSeek a third option entirely under its control, reducing its exposure to both US export controls and Huawei’s commercial terms.
(Related: Huawei and China Mobile Validate China’s First Carrier-Grade AI Inference System on Domestic Chips)
The Analyst View: Nvidia at Zero in China
Richard Windsor, an analyst at Radio Free Mobile, offered a blunt assessment of Nvidia’s position in China following the Reuters report: “Nvidia is at zero in China and staying there.” Windsor also noted the limits of DeepSeek’s chip ambitions: “DeepSeek has almost no chance of selling silicon outside of China unless it gets access to leading edge manufacturing”, a constraint that reflects the ongoing impact of US export controls on China’s access to advanced semiconductor fabrication equipment. TSMC, the world’s leading chip foundry, is barred from manufacturing chips for Chinese AI companies above certain performance thresholds, and SMIC, China’s domestic foundry, remains one to two generations behind TSMC’s leading-edge processes.
Those manufacturing constraints mean that DeepSeek’s inference chip, even if successfully designed, will likely be built on a process node less efficient than Nvidia’s for its own products. But chip-level efficiency is only one variable in the total cost of inference. DeepSeek’s algorithmic innovations, including the mixture-of-experts architecture that allowed R1 to achieve frontier-level performance at a fraction of the training cost, have demonstrated that software-level efficiency can compensate for hardware limitations. A DeepSeek inference chip optimized for its own model architectures could deliver competitive performance even on older process nodes, particularly for the specific workloads that DeepSeek’s models are designed to handle.