On 20 August, Nvidia denied a Reuters report that it planned to launch a customised “Language Processing Unit” (LPU) for Chinese customers before the end of the year. “The report on Nvidia’s LPU is wrong. We do not sell an LPU in the China market, nor do we have an LPU product roadmap for China,” a company spokesperson told Reuters.
The denial came hours after The Information, citing two Nvidia employees, reported that the company planned a small-batch release of a China-designed AI chip before year end, with several customers already placing orders. The chip is said to be a revised version of Nvidia’s LPU, a processor licensed from Groq that works alongside graphics processors to speed up AI chatbot responses.
Nvidia bought Groq’s core team and technology for 20 billion dollars at the end of last year and shipped its first LPU in March. As AI demand shifts from training to inference, traditional GPUs have grown expensive and power-hungry, pushing Nvidia to develop cheaper, more efficient inference silicon.
US export controls block Nvidia’s next-generation Vera Rubin system from reaching China. Nvidia adjusted its software so the LPU complies with the rules and runs with processors already inside China. Even so, the chip may do little to recover share.
Under the twin pressure of US controls and domestic substitution, Nvidia’s slice of the China AI chip market keeps shrinking. Baidu and others are building their own inference silicon, and model labs such as DeepSeek are reported to be building independent software-hardware ecosystems with Huawei. As Huang Renxun said in May, Nvidia has “basically ceded the China AI chip market to Chinese competitors.”
Editor’s note: This is an adapted translation of the original Sohu report. It has been trimmed and restructured for readability for an international business audience.