Nvidia CEO: No plans to sell AI chips to China

Nvidia CEO Jensen Huang said the company is not in talks to sell its Blackwell AI chips to Chinese firms, dismissing rumors of a planned return to China.

Huang, who spoke in Tainan, Taiwan, said that Nvidia currently has no plans to ship products to China and is waiting for policy changes that would allow such sales.

Nvidia, which leads the global AI chip market, remains barred from selling its most advanced AI chips to China under current US regulations.

A recent US-China trade agreement did not alter these restrictions, and officials have said no changes are expected soon.

Huang is in Taiwan for meetings with Taiwan Semiconductor Manufacturing Co., a key Nvidia partner, as part of a global tour that included stops in Washington and South Korea.

Nvidia has recently surpassed a US$5 trillion market capitalization, making it the world’s most valuable company.

🔗 Source: Bloomberg

🧠 Food for thought

Implications, context, and why it matters.

Nvidia’s China sales plunged after export bans

Nvidia booked $17 billion from China in fiscal 2025 (13% of $130.5 billion) 1. The latest quarter hit $2.8 billion after US export rules and Chinese curbs blocked H100, A100, and H20 GPUs 2 1.
It took $4.5 billion in charges in Q1 fiscal 2026 for the H20 ban, then posted zero H20 sales in Q2 1 3. A 15% China revenue deal brought no shipments, and Huang said talks with Chinese firms are inactive, so presence stays small despite a $50 billion AI market 2 4.

Hyperscalers (large cloud providers) and enterprises get first access as Blackwell Ultra ramps

Orders for AI chips top $500 billion through 2026 2. GB300 NVL72 links 72 Blackwell Ultra GPUs with 36 Grace CPUs (Nvidia’s Arm-based server processors), delivers 1.5x more AI performance than GB200 NVL72, and lifts AI factory (large-scale clusters used to build plus run AI models) revenue potential by 50x versus Hopper-based systems 5.
CoreWeave (a specialized GPU cloud provider) is rolling out GB200 NVL72 for IBM, Mistral AI, and Cohere 6. AWS, Google Cloud, Microsoft Azure, and Oracle Cloud Infrastructure plan Blackwell Ultra instances in H2 2025 5. That timeline creates work for liquid cooling, high-speed links up to 800 Gb/s per GPU, and integration firms 5. Early builds often use BlueField-3 Data Processing Units (DPUs, smart networking accelerators that offload networking plus storage tasks) for multi-tenant networking 7 5.