China's DeepSeek secretly develops its own AI inference accelerator

By Reuters (July 2026), DeepSeek has been secretly developing its own inference accelerator for about a year, aiming to build independence from American microchips.

Why does DeepSeek need its own chip?

The main driver is American export restrictions. Since 2022–2023, Washington has progressively banned the supply of advanced GPUs to China: first the A100 and H100 series, then H800 and A800, introduced as "restricted" versions for the Chinese market. As a result, Chinese AI companies are forced to work with lower computational power than their Western competitors.

DeepSeek gained fame for extracting competitive results from limited resources: in January 2025, the R1 model sparked widespread discussion, showing results on par with Western leaders while using significantly lower computational costs. However, the operational scale of a commercial AI service—millions of user requests per day—requires hardware optimized specifically for continuous inference. Its own accelerator would allow it to reduce the cost per request and eliminate dependence on an unstable supply chain.

What is known about development?

Details are scarce—DeepSeek intentionally maintains silence. According to Reuters sources:

• The project started about a year ago—presumably in mid-2025
• The company is in talks with potential partners on semiconductor manufacturing and packaging
• Simultaneously, it is conducting targeted hiring of chip design engineers without public job announcements
• The development focus is on an inference accelerator, not model training

The distinction of tasks is fundamental: for training models, DeepSeek uses its own computational resources—this task is already solved. The bottleneck is commercial inference, that is, continuous real-time request processing. Each token costs money: with millions of users per day, hardware savings directly determine service profitability and price competitiveness.

Developing a specialized chip is a task of fundamentally different scale than writing model code. It requires expertise in VLSI design, close partnerships with lithography foundries, and significant capital investment.

Who has already walked this path?

Development of proprietary inference accelerators has long been standard among major AI players:

• Google – TPU since 2016
• Amazon – Inferentia for AWS since 2019
• Meta – MTIA (Meta Training and Inference Accelerator)
• Microsoft – Maia 100, announced in 2023

Among Chinese companies: Baidu is promoting its own Kunlun processor, Huawei—the Ascend line. If DeepSeek brings the project to mass production, the company will enter this circle—but under incomparably stricter restrictions both in access to technologies and manufacturing capacity.

What does this mean

DeepSeek's secret chip development is yet another confirmation of a sustained trend: American sanctions are pushing the Chinese AI sector toward vertical integration—from algorithms to proprietary silicon. If development succeeds, this will not only strengthen DeepSeek's independence from Western hardware but also add a new player to the AI accelerator market, where Nvidia currently holds dominant position.

*Meta is recognized as an extremist organization and banned in the Russian Federation.