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

      DeepSeek is reportedly developing its own artificial intelligence chip, as stated in a Reuters article published on Monday that referenced individuals familiar with the situation. If realized, this would mark a significant shift for China’s leading AI lab, allowing it to move from merely developing software that operates on other companies' chips to creating the chips themselves.

      The Hangzhou-based startup has spent the last year optimizing its models for Huawei’s Ascend processors and other domestic Chinese silicon, shifting away from its previous dependence on Nvidia hardware. This new chip design would deepen that transition from adapting to Chinese hardware to defining it.

      The chip in question is intended for inference—the phase where a trained model responds to user inquiries—rather than for training, which requires significantly more computational power to create the model in the first place. In April, DeepSeek introduced its V4-Pro and V4-Flash models, which comprise a trillion parameters and are reportedly trailing the leading Western advancements by about three to six months.

      As per the Reuters report, the chip would be produced by Semiconductor Manufacturing International Corporation (SMIC), the largest foundry in China, rather than Taiwan's TSMC.

      SMIC has been restricted from accessing the most advanced chip manufacturing technologies due to US and Dutch export controls and is believed to be limited to a 7-nanometre process, which lags several generations behind the latest technology.

      DeepSeek has not confirmed these reports, having not responded to Reuters’ inquiries and generally maintaining a low profile in public communications. Founded by Liang Wenfeng, the lab has remained unusually secretive despite global attention on its models, meaning the details come from anonymous sources rather than official statements.

      The strategic reasoning is clear: the US has restricted the sale of Nvidia’s top chips to China and has considered adding DeepSeek to its Entity List, making a domestically produced inference chip a way to reduce the lab's dependence on hardware that may be difficult to acquire.

      Currently, nearly 70 percent of AI computing demand is forecasted to originate from inference rather than training, which is where a specialized chip would be most valuable, and where Chinese silicon is nearing competitiveness.

      This ambition has precedents; as early as 2025, media outlets like Digitimes reported that DeepSeek was in search of chip-design expertise, and experts have suggested that any domestic processor would have to rely heavily on SMIC.

      However, significant challenges remain. Designing a competitive AI accelerator is a complex, multi-year effort, and China's foundries are not yet capable of achieving the yields or performance levels of TSMC’s cutting-edge technologies, a difference that the export controls aim to maintain.

      According to credible estimates, SMIC’s AI chip yields are still low, limiting the number of usable parts that such a design could realistically generate.

      DeepSeek’s previous experiences also highlight this challenge. Its R2 model faced several delays after training runs were unsuccessful on Huawei equipment, forcing the lab to revert to Nvidia GPUs for training while reserving Chinese chips for inference—the very separation of roles that the new chip would ideally address.

      Financial support, however, is less of a problem now. Chinese chip manufacturers rushed to support the launch of V4, and state funding is reportedly flowing into the lab, with a state-backed venture leading a $45 billion funding round.

      The next steps will be a matter of validation rather than mere intention. A design does not equate to a functioning chip, a functioning chip isn’t necessarily a market-ready product, and silicon that has been taped-out can still fail during testing.

      Until DeepSeek provides confirmation, the most cautious interpretation is a modest one: a lab equipped with engineers, motivation, and now funding seems to be striving to bridge the final gap in a fully domestic AI stack, from the model level down to the chip. Whether it achieves this goal is another matter and remains uncertain for now.