DeepSeek is reportedly designing its own AI chip to sidestep US curbs

DeepSeek is designing its own artificial-intelligence chip, according to a Reuters report published on Monday that cited people familiar with the matter. The move, if it holds up, would push China’s most closely watched AI lab from writing software that runs on other companies’ silicon to specifying the silicon itself.

The Hangzhou startup has spent the past year tuning its models for Huawei’s Ascend processors and other domestic Chinese silicon, a departure from its earlier reliance on Nvidia hardware. A chip of its own design would extend that shift a further step, from adapting to Chinese hardware to defining it.

The reported chip is meant for inference, the stage where a trained model answers user queries, rather than for training, the far more compute-hungry work of building the model in the first place. DeepSeek released its V4-Pro and V4-Flash models in April, a trillion-parameter line the company itself says trails the Western frontier by roughly three to six months.

According to the Reuters account, the chip would be fabricated by Semiconductor Manufacturing International Corporation, China’s largest foundry, rather than Taiwan’s TSMC.

SMIC has been cut off from the most advanced chipmaking tools by US and Dutch export controls, and is widely reported to be stuck on a 7-nanometre process several generations behind the leading edge.

None of this has been confirmed by DeepSeek, which did not respond to Reuters’ request for comment and has a long record of saying little in public. The lab, founded by Liang Wenfeng, has stayed unusually opaque even as its models have drawn global scrutiny, so the specifics here rest on anonymous sourcing rather than on the record.

The strategic logic is plain enough. Washington restricts the sale of Nvidia’s best chips to China and has weighed adding DeepSeek to its Entity List, so a home-grown inference part would trim the lab’s exposure to hardware it may struggle to buy.

Roughly 70 per cent of AI compute demand is now expected to come from inference rather than training, which is where a purpose-built chip would earn its keep and where Chinese silicon is already closest to competitive.

There is precedent for the ambition. As far back as early 2025, trade press including Digitimes reported that DeepSeek was recruiting chip-design talent, and analysts have argued that any domestic processor would have little choice but to lean on SMIC.

The obstacles are just as plain. Designing a competitive AI accelerator is a multi-year undertaking, and China’s foundries cannot yet match the yields or the performance of TSMC’s most advanced nodes, a gap the export controls are meant to preserve.

On the most credible estimates, SMIC’s AI-chip yields remain low, which limits how many usable parts a design of this kind could realistically produce.

DeepSeek’s own experience shows the limits. Its R2 model was repeatedly delayed after training runs failed on Huawei hardware, forcing the lab back onto Nvidia GPUs for training while reserving Chinese chips for inference, which is precisely the division of labour the new chip would reportedly serve.

Money, at least, is no longer the constraint it once was. Chinese chipmakers rushed to support V4 on launch day, and state capital is flowing toward the lab, with the state-backed Big Fund reported to be leading a $45bn funding round.

What comes next is a question of proof rather than intent. A design is not a working chip, a working chip is not a shipping product, and taped-out silicon can still fail on the test bench.

Until DeepSeek confirms anything, the safest reading is a modest one: a lab with the engineers, the motive, and now the money appears to be trying to close the last remaining gap in a fully Chinese AI stack, from model down to chip. Whether it succeeds is a separate story, and one nobody can yet source.