Two people with knowledge of the matter told Reuters that Nvidia is preparing to launch its Groq 3 LPU inference chip in China as early as May. The sources stressed that the chip is "not a downgraded version or one specifically customized for the Chinese market" — a notable twist for a company that usually ships reduced-spec silicon to stay in line with US export controls.
The Groq 3 LPU is a dedicated inference co-processor packing 500MB of on-chip SRAM and high inference bandwidth. It would be the first Groq product to enter China since Nvidia acquired the AI inference startup for roughly $17 billion in late 2025. The move is separate from the H200 GPU, which was previously approved for production restart in China.
If the timing holds, the Groq 3 LPU could hit Chinese data centers as soon as May — no downgrade necessary.