DeepSeek is said to be developing its own AI chip to circumvent US restrictions.

      DeepSeek is reportedly developing its own artificial intelligence chip, according to a Reuters article published on Monday that referenced sources familiar with the project. If this development comes to fruition, it would mark a significant shift for China's highly scrutinized AI lab, moving from software that operates on other companies' silicon to creating its own silicon.

      The startup based in Hangzhou has spent the previous year optimizing its models for Huawei's Ascend processors and other Chinese silicon alternatives, moving away from its previous dependence on Nvidia hardware. Designing its own chip would represent a further transition from adapting existing Chinese hardware to defining its specifications.

      The reported chip would focus on inference, the process where a trained model responds to user queries, rather than training, which requires significantly more computational power to build the model initially. DeepSeek launched its V4-Pro and V4-Flash models in April, boasting a trillion parameters, but the company claims they lag behind Western advancements by approximately three to six months.

      According to Reuters, this chip would be produced by Semiconductor Manufacturing International Corporation (SMIC), which is the largest foundry in China, rather than Taiwan's TSMC. SMIC has faced restrictions on accessing advanced chipmaking technologies due to US and Dutch export controls and is generally believed to be operating with a 7-nanometre process that is several generations behind the cutting edge.

      DeepSeek has not verified any of this information, having not responded to Reuters' request for comments and typically maintaining a low profile in public discussions. The lab, established by Liang Wenfeng, has remained particularly private even as its models have attracted global attention, making the details dependent on anonymous sources instead of official statements.

      The strategic reasoning is straightforward. The US limits the sale of Nvidia's top chips to China and has considered including DeepSeek on its Entity List, so developing an indigenous inference chip could reduce the lab's reliance on hardware that may be difficult to acquire.

      Currently, about 70 percent of AI computing demand is anticipated to arise from inference rather than training, which is where a specialized chip could be beneficial and where Chinese silicon is already nearly competitive.

      There is a precedent for this ambition; as early as 2025, industry publications such as Digitimes reported that DeepSeek was looking to recruit chip-design experts, and analysts suggest that any domestically produced processor would likely have to collaborate with SMIC.

      However, significant challenges exist. Creating a competitive AI accelerator is a lengthy process, and China's foundries have not yet reached the yields or performance levels of TSMC's most advanced nodes, a discrepancy that export controls aim to maintain.

      According to the most reliable estimates, SMIC's yields for AI chips are still low, which limits the number of usable components that could realistically be produced from such a design.

      DeepSeek's own challenges illustrate these limitations. Its R2 model faced multiple delays when training runs on Huawei hardware failed, leading the lab to revert to Nvidia GPUs for training while using Chinese chips solely for inference, which aligns with the intended purpose of the new chip.

      At least financial resources are no longer a significant constraint as they once were. Chinese chipmakers quickly supported the launch of V4, and state investment is flowing into the lab, with the state-backed Big Fund reportedly leading a $45 billion funding round.

      The next steps will be a matter of proving capabilities rather than intentions. A design does not equate to a functioning chip, and a working chip does not guarantee a marketable product, as even finalized silicon can fail during testing.

      Until DeepSeek provides confirmation of these developments, the most cautious interpretation is that a lab with the available engineers, motivation, and funding seems to be working to bridge the last gap in a fully Chinese AI infrastructure, from models to chips. Whether they will achieve this remains uncertain and is something that cannot yet be substantiated.