SAN JOSE, California, March 19, 2026, 02:09 PDT
Nvidia has won Beijing’s approval to resume sales of its H200 artificial intelligence chips in China and is preparing a version of its new Groq chip for that market, sources told Reuters. The step reopens a country that once generated 13% of Nvidia’s total revenue and gives the chipmaker a new route into inference, the stage when trained AI systems answer prompts or carry out tasks. 1
The move lands as Nvidia tries to turn the product burst from its GTC developer conference into sales beyond model training and into a broader data-center stack. Huang said this week the revenue opportunity for Blackwell and Rubin could top $1 trillion through 2027, but Reuters reported that target does not include H200 sales to China, so any recovery there would sit outside the forecast. 2
The H200 is a Hopper-based part and Nvidia’s second-most powerful AI chip, one generation behind its Blackwell and Rubin lineup. Huang told reporters the company had already taken orders after U.S. export approvals and added, “Our supply chain is getting fired up.” 1
On the Groq side, one source told Reuters the China-bound chip is not a stripped-down model designed only for China but a variant that can work with other systems. Nvidia licensed Groq technology in a $17 billion deal in December and plans to use it for inference work such as writing code or answering questions; the source said the China-ready version could be available in May. 3
At GTC in San Jose, Nvidia said its Vera Rubin platform is in full production, pairing Rubin GPUs with Groq 3 LPX accelerators, Vera CPUs and networking gear in rack-scale systems for large AI data centers. It also launched Dynamo 1.0, open-source software that it described as an operating system for inference workloads across clusters of chips. 4
Nvidia leaned on customer names to validate that push. In the company’s March 16 release, OpenAI CEO Sam Altman called Nvidia infrastructure “the foundation” for pushing the frontier of AI, while Anthropic CEO Dario Amodei said more complex reasoning needs systems that “can keep pace.” 4
Competition is thicker here than in training. Reuters reported Google and Amazon are bringing rival AI chips to market, AMD is pushing software to loosen Nvidia’s grip, and Baidu already makes inference chips in China; analyst Richard Windsor wrote that Nvidia’s lock-in is “not nearly as strong in inference.” 5
But the route back is still narrow. Rubin systems cannot be sold in China, the Groq version for that market is not expected until May, and Nvidia’s own March 16 release said the timing of some new products and features can still change. 3
Nvidia said Rubin-based products should start reaching partners in the second half of this year, including Amazon Web Services, Google Cloud, Microsoft Azure and Oracle Cloud Infrastructure. That gives the company a broad rollout outside China while it tries to rebuild there with an older Hopper chip and a not-yet-shipping Groq variant. 4
