Nvidia will ship Groq AI chips to China starting May 2026. The company unveiled the plan March 17 after securing regulatory approval for H200 sales. Jensen Huang bets that inference processors will bypass training restrictions that killed H20 and B30A shipments. The $20 billion Groq licensing deal from December 2025 suddenly makes sense. China gets language processing units instead of GPUs. Beijing approved the imports despite banning training chips.

The timing follows a painful pattern. The Biden administration blocked H20 sales in January 2026. The Trump administration maintained the wall. Alibaba and Tencent scrambled for alternatives while Nvidia lost $15 billion in quarterly revenue. Groq chips slip through because they accelerate inference rather than train large language models. Huang told investors to exclude China from forecasts. Then he found the loophole.

Huang Announced Groq 3 at GTC 2026 for May Deployment

The GTC 2026 conference in San Jose revealed the strategy. Huang showed Groq 3 chips built for inference speed not model training. Chinese customers will receive the first export versions. Reuters reported the timeline citing two sources familiar with the matter. The chips handle token generation for deployed AI models. Not training. That distinction keeps them inside legal boundaries while satisfying demand from Alibaba ByteDance and Tencent.

The technical specs focus on language processing units. Groq technology processes AI inference tasks through a compiler that accelerates token generation. Nvidia redesigned the architecture specifically for Chinese export compliance. No tensor cores for training. Pure inference throughput. Sources say the chips will slot into existing H200 servers as companion processors. This creates a hybrid system that technically complies with US restrictions while delivering performance Chinese tech giants desperately need.

20 Billion Dollar Bet on Inference Avoids Trump Restrictions

Nvidia paid $20 billion for Groq technology in December. Now that purchase becomes the backdoor into China. The Groq deal included licensing agreements that transfer compiler technology and chip designs to Nvidia's control. Huang immediately redirected the engineering team toward export compliant versions. Two million H200 orders already sit in the queue. Groq adds a new revenue line without triggering national security flags.

The political calculation is risky. Trump administration officials could close the inference loophole overnight. But Huang gambles that inference chips fall outside the national security focus on training capabilities. China needs inference to run existing models. They need training to build new ones. Washington banned the latter. Nvidia now sells the former. If the strategy holds, Groq generates billions from Alibaba's cloud services and ByteDance's TikTok algorithms. If regulators object, Huang loses his last China bridge.

Source: https://www.reuters.com/world/china/nvidia-preparing-groq-chips-that-can-be-sold-chinese-market-sources-say-2026-03-17/