Nvidia plans to introduce the Groq 3 LPU inference chip to the Chinese market in May. This chip is not a downgraded version, but rather Groq's first product introduction to China. The Groq 3 LPU features high bandwidth and a unique architecture that may circumvent U.S. export controls, but it requires adaptation to other systems, and its performance remains to be seen.

Nvidia released the Groq 3 LPU chip on March 17, with shipments expected in Q3. The chip can be used in conjunction with Vera Rubin NVL72 to significantly enhance inference throughput. The company also showcased the next-generation architecture Kyber, expected to launch in 2027.