Chinese artificial intelligence startup DeepSeek is working on its own AI chip designed for inference, according to people with knowledge of the matter, a strategic step intended to lessen the company's reliance on Nvidia and Huawei processors that it has used to train and operate its widely used models.
The chip under development is targeted at inference - the phase of AI computing when a trained model produces outputs for end users - rather than at the training process required to build new models, the sources said. If successful, creating an in-house inference processor would represent a significant shift for DeepSeek, which in China has been broadly celebrated as a leading AI contender.
Early-stage program and partnerships
Sources described DeepSeek's initiative as nascent. The company began outreach about a year ago, holding talks with external partners including firms that specialize in chip design, semiconductor foundries and memory components. Recruitment of chip-design engineers has increased in recent months, the sources said, but hiring has been conducted discreetly without public listings on mainstream job platforms.
All three people who provided information declined to be named because the plans have not been publicly announced. DeepSeek has kept a relatively low public profile despite its prominence in Chinese AI circles, and the company did not respond to a request for comment.
Hardware context and supplier dynamics
DeepSeek has relied on both Nvidia and Huawei processors while building and operating its models. The foundation model that underlies R1 - a reasoning model whose cost-efficient performance sparked significant market moves in January 2025 - was trained on Nvidia's H800, a chip tailored for the Chinese market that Washington restricted in late 2023. Since then, the company has increased its use of Huawei's chips.
In April, DeepSeek released V4 in a form adapted for Huawei's Ascend family, and Huawei has said its processors were used in part of the training for V4-Flash, a lighter variant of the model. After the launch, Reuters reported that orders for Huawei's Ascend 950 chips from Chinese technology conglomerates surged.
Why an inference chip?
The startup's plan to build an inference-focused chip responds to the shifting economics of AI workloads. As AI applications proliferate, much of the industry's computing demand is moving from the expensive, power-hungry process of training models to the recurring work of running them. Inference chips can be optimized to be less costly and more energy-efficient than general-purpose GPUs, addressing the segment of demand that is growing fastest.
That market dynamic has encouraged other AI developers to pursue greater hardware control. OpenAI recently introduced Jalapeno, its first custom inference chip developed with Broadcom, and Anthropic has been reported to be considering its own chip designs. DeepSeek's reported effort would place it alongside those global developers attempting to tailor hardware to their model requirements.
Strategic and regulatory pressures
The project also has a geopolitical and strategic dimension. U.S. export controls prevent Chinese companies from purchasing Nvidia's most advanced chips, and Beijing has encouraged leading technology firms to build domestic alternatives. DeepSeek's founder, Liang Wenfeng, acknowledged in a rare 2024 interview with Chinese media that export controls have posed a challenge for the company.
At the same time, Huawei has taken on a significant portion of China's domestic AI chip market following U.S. restrictions, supplying DeepSeek and other major players. But Huawei's dominance is not unchallenged: local rivals including Alibaba and Baidu are developing their own chips and gaining share, according to the sources.
Technical and manufacturing hurdles
Designing a commercially competitive AI chip typically requires years of development and substantial capital. Manufacturing the devices introduces another layer of difficulty: U.S. rules bar Chinese designers from accessing the most advanced overseas foundries, and separate U.S. curbs have reduced Chinese access to high-bandwidth memory (HBM), a component critical to many inference designs. These constraints create significant headwinds for any Chinese firm seeking to bring a new AI processor to market.
Given those barriers, success is far from certain. The sources stressed the early stage of DeepSeek's effort and the likely length and cost of bringing a chip from concept to production.
Capital and corporate shifts
The chip initiative coincides with a notable change in DeepSeek's financing approach. After years of declining external capital, the company has begun to embrace outside investment: it was slated to raise $7 billion in an initial funding round that would value it between $52 billion and $59 billion, according to reporting in June. That planned fundraise marks a reversal of the company's prior strategy of rejecting external investment and could provide resources relevant to an ambitious hardware program.
Implications for markets and industry segments
Should DeepSeek move forward with a proprietary inference chip, the effort would touch multiple sectors: semiconductor design and manufacturing, memory suppliers, and the cloud and AI services market that depends on cost-effective inference. The project would also influence competitive dynamics among Chinese technology providers and could affect purchasing patterns for domestic AI infrastructure.
For now, the effort remains confidential and in progress. DeepSeek's planned contacts with designers, foundries and memory firms, together with private hiring of chip engineers, suggest an intent to pursue hardware autonomy. But significant technical, manufacturing and regulatory obstacles lie ahead, and the timeline and outcome remain uncertain.