Nvidia Groq 3 LPU Inference Chip Expected to Enter China in May, May Circumvent GPU Export Controls

Gate News reports that on March 18, NVIDIA is preparing to launch the Groq 3 LPU inference chip in the Chinese market, with a expected release as early as May. According to sources, this chip is not a downgraded version nor specifically customized for China. This marks NVIDIA’s first introduction of its product line into China after acquiring AI inference chip company Groq for approximately $1.7 billion by the end of 2025. This move is separate from the previously approved restart of production for the H200 GPU and reflects an independent China-focused chip strategy. The Groq 3 LPU is a dedicated inference co-processor with 500MB on-chip SRAM and inference bandwidth up to 150 TB/s, but it has lower floating-point performance and is not optimized for model training. Its architecture allows it to potentially stay below the current US export control thresholds (TPP < 21,000 and DRAM bandwidth < 6,500 GB/s), thus avoiding export restrictions faced by GPUs like the H200. However, in NVIDIA’s original plan, the Groq LPU needed to be paired with the Vera Rubin GPU, which cannot be exported to China (recommended ratio approximately 25:75). The Chinese version will need to be adapted to operate independently with other systems, and its actual performance remains to be seen.