- Strategic. DeepSeek — whose V3 and R1 models shook global chip markets earlier this year — has begun developing an in-house AI chip focused on inference, seeking to reduce operating costs and cut dependence on Nvidia hardware.
- Early stage. The project is roughly a year old and remains far from production: the company is in early talks with chip design, contract manufacturing, and memory partners, and has been quietly hiring semiconductor engineers.
- Obstacles. US export controls block Chinese designers from accessing the most advanced overseas foundries, and separate curbs have restricted China’s access to high-bandwidth memory — a component critical for inference chips.
DeepSeek, the Chinese AI startup behind the DeepSeek-V3 and R1 models, has launched an internal chip development initiative targeting the inference segment of the AI compute stack, according to TechNode, citing people familiar with the effort. The project began approximately one year ago and is being pursued with considerable discretion: the company has relied on targeted outreach rather than public job listings to recruit chip architects and verification engineers.
What DeepSeek is building — and why
The chip in development is designed for inference — the stage at which a trained model generates responses for users — rather than for training new models from scratch. That distinction matters economically. Unlike training, which demands concentrated bursts of computing power, inference must handle continuous, high-volume user requests. Custom inference silicon can, in principle, deliver meaningfully lower cost-per-query and power consumption than general-purpose GPUs.
For DeepSeek, the business case is direct: compute costs represent more than half of operating expenditure for many foundation model companies, and the company currently runs its infrastructure on Nvidia and Huawei hardware subject to supply constraints and geopolitical risk. The company is simultaneously seeking $7 billion in external funding at a valuation between $52 billion and $59 billion; sources told TechNode that chip investment is expected to become a priority if that raise completes.
The obstacles ahead
Competitive inference chips take years and substantial capital to bring to production. DeepSeek’s project faces two compounding constraints specific to China. First, US export controls prohibit Chinese chip designers from engaging the most advanced overseas foundries — specifically the leading-edge nodes at TSMC and Samsung at which Nvidia manufactures its GPUs. Second, separate American restrictions have curtailed China’s access to high-bandwidth memory, which inference processors require in volume to move model weights efficiently between storage and processing cores.
The result is that any chip DeepSeek ultimately produces would likely rely on domestically available process nodes and memory, limiting the performance ceiling achievable. That is a well-understood constraint across China’s AI sector: Meituan’s LongCat-2.0 became the first frontier model trained entirely on domestic Chinese chips — a significant symbolic milestone, but one that required extensive engineering workarounds to close the gap with foreign hardware.
Significance in the broader chip race
DeepSeek’s move into silicon reflects a broader pattern among Chinese AI companies: as access to imported chips becomes less certain, developing proprietary hardware is increasingly treated as a strategic necessity rather than a distant aspiration. Whether DeepSeek’s effort ultimately produces a competitive product or primarily serves as an internal hedge against supply disruption, its entry into chip design adds another node to a rapidly expanding Chinese semiconductor ecosystem operating under sustained external pressure.