DeepSeek starts designing its own AI chip to get around US export bans

DeepSeek, the Chinese AI lab based in Hangzhou, is developing its own AI chip, Reuters reported on July 7, 2026, citing people familiar with the matter. The company wants to go beyond writing software models that run on other vendors' processors and start designing silicon accelerators itself.

Over the past few years, DeepSeek has become one of the most talked-about players in the global AI market, thanks to its ability to build powerful models with minimal compute. Now the company appears to be taking the next strategic step: designing hardware for its own needs.

What's behind the decision to design a chip

For DeepSeek, this is a fundamental shift. Until now, the company has worked exclusively on the software layer: its models — R1, V3, and earlier versions — ran on third-party accelerators, mostly from NVIDIA. According to Reuters, the startup now intends to define its own silicon architecture.

DeepSeek does not plan to design and manufacture the chips on its own: in China, those jobs are traditionally split between different companies. According to the sources, SMIC — Semiconductor Manufacturing International Corporation, China's largest contract chipmaker — is under consideration as the manufacturing partner.

• Reuters report — July 7, 2026
• DeepSeek is based in Hangzhou, China
• The chip targets inference, not training
• Likely manufacturing partner — SMIC

Why did US restrictions become the key trigger?

Since 2022, the US has steadily cut off China's access to high-performance semiconductors. The ban first covered NVIDIA's flagship A100 and H100 GPUs, then their cut-down A800 and H800 versions, and by 2025 export restrictions applied to virtually every accelerator above a set performance threshold. The goal is to keep Western hardware from being used to train frontier AI models for military projects.

DeepSeek is arguably the best-known casualty of these restrictions — and also the most inventive at finding ways around them. In January 2025, the company released R1, a model that came close in quality to the best Western models of the time at a significantly lower claimed training cost. The result rattled investors worldwide: it turned out that algorithmic efficiency can make up for a hardware shortage.

What changes if the chip ships?

An in-house AI accelerator would turn DeepSeek into a vertically integrated company, following the model of Google with TPU or Amazon with Trainium. For China's AI industry, it would be another step toward technological independence from Western suppliers.

DeepSeek wouldn't be the first here: Huawei, through its affiliates, already builds the Ascend line of AI chips and promotes them domestically as an alternative to NVIDIA. DeepSeek's advantage, however, is its deep understanding of AI inference, which would let it design an accelerator tailored precisely to its own architectural choices rather than one copied from a general-purpose GPU.

That said, Reuters so far cites only anonymous sources, and DeepSeek has not officially confirmed anything. Even if development is underway, getting from design to volume shipments will take years: SMIC runs on equipment that lags behind leading Western tools, and its most advanced available process node is around 7 nm.

What it means

DeepSeek is moving from algorithmic efficiency toward hardware self-sufficiency. If the project pans out, China will gain another player in its domestic AI chip segment — and more evidence that US export restrictions are accelerating investment in Chinese semiconductor sovereignty rather than holding it back.