DeepSeek begins designing own AI chip to bypass U.S. export restrictions

DeepSeek, a Chinese AI laboratory based in Hangzhou, is developing its own AI chip — reported on July 7, 2026 by Reuters, citing people familiar with the situation. The company intends to move from writing software models running on others' processors to independently designing silicon accelerators.

Over the past several years, DeepSeek has become one of the most discussed players in the global AI market — thanks to its ability to create powerful models with minimal computational resources. Now the company, according to reports, is taking the next strategic step: designing hardware for its own needs.

What stands behind the decision to design a chip

For DeepSeek, this represents a fundamental pivot. Until now, the company has focused exclusively on the software layer: its models — R1, V3, and earlier versions — ran on others' accelerators, primarily NVIDIA products. Now the startup, according to Reuters sources, intends to define silicon architecture itself.

DeepSeek does not plan to develop and manufacture chips alone: in China, these tasks are traditionally divided among different companies. According to sources, SMIC — Semiconductor Manufacturing International Corporation, China's largest contract semiconductor manufacturer — is being considered as the production partner.

• Reuters publication — July 7, 2026
• DeepSeek is based in Hangzhou, China
• Chip orientation — inference tasks, not training
• Likely production partner — SMIC

Why did American restrictions become the key trigger?

Since 2022, the US has consistently cut off China's access to high-performance semiconductors. First, the ban extended to NVIDIA's flagship GPUs — A100 and H100, then to their "lightweight" versions A800 and H800, and by 2025, export restrictions covered virtually all accelerators above the established performance threshold. The goal is to prevent the use of Western hardware for training advanced AI models in military projects.

DeepSeek proved to be perhaps the most famous victim of these restrictions — and simultaneously the most inventive player in finding workarounds. In January 2025, the company released the R1 model, which closely approached the quality of the best Western counterparts at the time with significantly lower claimed training costs. This result shook investors worldwide: it turned out that algorithmic efficiency could compensate for hardware deficits.

What will change if the chip launches?

A proprietary AI accelerator would transform DeepSeek into a vertically integrated structure — following the model of Google with TPU or Amazon with Trainium. For China's AI industry, this would be another step toward technological independence from Western suppliers.

DeepSeek would not be the first here: Huawei through its structures is already creating the Ascend AI chip series and pushing it to the domestic market as an alternative to NVIDIA. However, DeepSeek's advantage lies in deep understanding of AI inference, which will allow designing an accelerator precisely tailored to its own architectural solutions, rather than copied from a generic GPU.

Then again, Reuters is only citing anonymous sources so far — there is no official confirmation from DeepSeek. Even if development is underway, the path from design to mass production will take years: SMIC operates on equipment inferior to the most advanced Western counterparts — the maximum available manufacturing process is around 7 nm.

What this means

DeepSeek is moving from algorithmic efficiency to hardware self-sufficiency. If the project succeeds, China will gain another player in its own AI chip segment — and further confirmation that American export restrictions are more likely to accelerate investment in China's semiconductor sovereignty than to contain it.