Nvidia's Groq 3 LPU Inference Chip to Launch in China in May, Non-Downgraded Version May Bypass Export Controls

For Chinese enterprises needing efficient AI inference capabilities but restricted by export controls, a non-downgraded high-performance inference chip is coming. According to Reuters, Nvidia plans to launch the Groq 3 LPU inference chip in the Chinese market in May this year—its first introduction of this product line to China since acquiring Groq in late 2025.

The Groq3 LPU is an inference-specific coprocessor with 500MB on-chip SRAM and an inference bandwidth of 150TB/s, but it has low floating-point computing power and is not suitable for model training. Its performance parameters may be below the current U.S. export control thresholds, allowing it to bypass GPU export restrictions. However, the chip originally required pairing with the Vera Rubin GPU (which cannot be exported to China), so the Chinese version needs to adapt to other systems to run independently—its actual performance remains to be seen.