DeepSeek is going to expose its need to make its own chips as well: a one-year plan to move on from both NVIDIA and Huawei at the same time

DeepSeek has reportedly been secretly developing chips for nearly a year, targeting data center inference chips, and has already contacted hardware partners and recruited engineers, with the goal of reducing dependence on both NVIDIA and Huawei.

(Previous context: ByteDance's self-developed CPU to be finalized by early 2027 at the latest, mass production in the second half, bringing in Qualcomm for wafers) (Background: AI NVIDIA chip smuggling case progress - mysterious Chinese female financier appears, drug investigation accidentally reveals)

Table of Contents

Toggle

• Quietly seeking partners and engineers for nearly a year
• Why now
• China's computing power autonomy gamble

DeepSeek, a well-known large language model company in China, is quietly making moves in the chip industry. According to a Reuters report citing three informed sources, this chip self-development plan has been underway for nearly a year, targeting not training chips but data center inference chips—the chip that allows the model to "answer questions."

Huawei controls about half of China's data center chip market, while NVIDIA is barred by US export restrictions. DeepSeek doesn't want to pick a side; it wants to be independent of both.

Quietly seeking partners and engineers for nearly a year

According to Reuters, DeepSeek has been in contact with potential partners in the hardware and chip fields over the past year, simultaneously recruiting engineers. The plan remained low-key until recently being exposed. To understand why it is necessary, we must first look at the chip that made it famous.

DeepSeek rose to fame with its R1 model, achieving performance comparable to OpenAI and Anthropic at extremely low costs, shaking Silicon Valley's valuation logic. However, few have examined that R1 was trained using NVIDIA's H800 chips. The H800 is a specially designed version for the Chinese market, with deliberately reduced performance to comply with US export regulations—essentially a "neutered" high-end chip. Even so, the US banned the H800 in late 2023, cutting off DeepSeek's last fallback option. In other words, its most remarkable achievement was built on a foundation that could be cut off at any time. This insecurity is the deep motivation driving it to bring chip development in-house.

DeepSeek is now focusing on data center chips. Simply put, these are specialized chips placed in server rooms for running AI computations. More precisely, DeepSeek aims to develop inference chips—the stage where a trained model "answers questions and generates results," as opposed to training, which involves teaching the model from scratch. Focusing only on inference, not training, is a pragmatic but limited choice. The pragmatic aspect is that once a model is deployed, it faces massive daily queries, consuming significant computing power on the inference side. Solving "affordable and efficient operation" first can immediately alleviate reliance on external chips. The limitation is that training still requires the most advanced processes and chips, meaning DeepSeek is tackling the easier problem first, leaving the hardest challenge for later—a phased breakthrough approach.

Why now

US chip export controls are the most direct reason DeepSeek is rushing to develop its own chips. NVIDIA supplies chips to most AI companies in North America and Europe, but export controls—essentially government bans on selling specific chips to certain countries—prevent NVIDIA's most advanced products from entering China. This forces Chinese AI companies into two paths: buy Huawei's chips or make their own.

Huawei currently holds about 50% of China's data center chip market, seemingly an off-the-shelf solution. However, DeepSeek, Alibaba, and Baidu have all chosen the second path for an understandable reason: handing the lifeline to Huawei merely replaces reliance on NVIDIA with reliance on Huawei, still dependent on another's mercy. For companies committed to long-term frontier model development, the only truly reassuring option is to own a chip themselves.

There's a subtle contrast here: both aim to break free from NVIDIA dependency, but the motivations are poles apart. DeepSeek is cornered by export controls, driven by "must-do" survival. Meanwhile, OpenAI and Broadcom's recently announced Jalapeño chip—OpenAI's first designed for "large-scale inference"—follows a "better to do it" stability path.

OpenAI partly wants to reduce dependence on NVIDIA, and partly aims to control the entire tech stack like Apple—essentially owning everything from chips and servers to software, no longer bottlenecked by any supplier. Anthropic is also exploring self-developed chips but has no public roadmap yet.

China's computing power autonomy gamble

Self-developing chips is never easy. DeepSeek's year-long preparations are still far from actual tape-out and mass production. Moreover, data-center-grade inference chips demand extremely high standards in yield rates, supply chains, advanced process foundries, and firmware ecosystems—capabilities that China's semiconductor industry still lacks.

DeepSeek is not alone. Alibaba and Baidu are also developing their own AI chips. These tech giants have tacitly agreed: rather than betting on export controls loosening, it's better to hold fate in their own hands. This momentum aligns with China's national strategy of "AI self-sufficiency." In the past, China was most praised for software and algorithms; R1 has already proven that China can catch up with the top tier at the model level. What remains is the foundational layer of chips and computing power.