DeepSeek is secretly developing chips, focusing on inference. The project was launched a year ago, and the entire recruitment process has been kept confidential.

DeepSeek is developing its own chips!

The goal: reduce reliance on NVIDIA.

According to Reuters, this Chinese AI firm that has shaken Silicon Valley with its model algorithms is secretly developing a self-developed AI chip.

The chip is positioned for inference rather than training, and the project was launched about a year ago.

Currently, the project is still in its early stages. Sources reveal that DeepSeek has already engaged in discussions with chip design firms, wafer foundries, and memory suppliers.

From an Algorithm Company to a Hardware Player

DeepSeek has long been recognized for its extreme optimization of algorithms.

From the R1 inference model that took the world by storm in January 2025 to the later V4 series adapted for Huawei Ascend, this company has consistently demonstrated its ability to achieve more with less computing power.

But now, it is stepping directly into hardware development.

Insiders reveal that DeepSeek has been recruiting chip design engineers in recent months, but the hiring process is extremely low-profile, with no job postings published on any public recruitment platforms.

Self-developing chips by model companies is not an isolated case.

Around the globe, it has become a trend for AI model companies to develop their own chips. OpenAI just unveiled its first custom inference chip Jalapeño co-developed with Broadcom last month, and Anthropic was reported in April this year to be considering developing its own AI chips.

Liang Wenfeng, founder of DeepSeek, once revealed in a rare media interview in 2024 that insufficient chips posed a challenge to the company.

The base of the R1 model was trained on NVIDIA H800. After that, DeepSeek shifted to Huawei Ascend. The V4 model released in April this year has been adapted for Ascend chips, and Huawei also confirmed that its processors participated in part of the training of V4-Flash.

But DeepSeek feels that relying on two suppliers is still not enough.

This self-developed chip from DeepSeek is specifically designed for inference.

With the large-scale deployment of AI applications, the focus of the industry's computing power demand is rapidly shifting from training to inference. Training is a one-time investment, while inference is a continuous consumption process: every query from every user requires inference computing power to respond.

Specialized chips customized for inference scenarios can achieve lower power consumption and lower unit cost than general-purpose GPUs, which is crucial for an AI company moving toward commercialization.

DeepSeek has already laid the groundwork for hardware collaboration at the model design level.

The UE8M0 FP8 data format introduced in DeepSeek-V3.1 is considered to be specifically designed for the hardware features of next-generation domestic chips. The algorithm team was already thinking about chip design while developing the model.

However, the challenges are equally significant: designing a competitive AI chip usually requires several years and huge capital investment, and success cannot be guaranteed at this stage.

51 Billion Yuan in Financing Secured, Computing Power Infrastructure Deployed Simultaneously

This plan is backed by substantial real funding.

In June 2026, DeepSeek completed its first round of external financing since its establishment, raising about 51 billion yuan (approximately 7.4 billion US dollars), with a post-investment valuation ranging from 52 billion to 59 billion US dollars. The company had rejected external investment for many years before, and opening up to financing this time itself represents a major strategic shift.

The use of funds has been clearly defined: expanding computing power centers dominated by domestic chips, developing self-developed AI chips, and expanding its team of top global talents.

Actions at the infrastructure level are also advancing. DeepSeek has posted job openings for IDC design and planning engineers, with plans to participate in the planning and construction of data centers ranging from megawatt to gigawatt scale. The job postings explicitly mention construction locations including Ulanqab, Inner Mongolia.

DeepSeek itself maintains its usual low profile. Three insiders all requested anonymity, and the company has not made any public statements on the matter.

Nevertheless, this company that has reshaped the pace of global AI competition is now set to transform the underlying hardware foundation it relies on.

References:

[1]https://www.reuters.com/world/china/chinas-deepseek-developing-its-own-ai-chip-sources-say-2026-07-07/

This article is from the WeChat Official Account "QbitAI", written by Meng Chen, and published by 36Kr with authorization.