What is the real reason for Nvidia's determination to hold onto the Chinese market?

To boost sales in the Chinese market, Nvidia CEO Huang Renxun has been actively campaigning, even making three visits to Beijing in 2025 and sharing Air Force One with President Trump. Despite Nvidia receiving a conditional export license for the H200 series, exclusive to the Chinese market (with a 25% revenue share paid to the U.S. government) at the end of last year, no approved domestic companies have received any substantial shipments as of mid-May this year.

The latest article published by the global semiconductor authority Global Semi Research (GSR) indicates that the proportion of Nvidia's total revenue from the China region has plummeted from 26% in fiscal year 2022 to around 5% last year. Given the extreme global demand for computing power, it is financially difficult to justify Nvidia's stubbornness in holding onto the Chinese market.

GSR notes that Nvidia's true purpose in China has never been "sales" but rather to acquire "market intelligence (Market Intelligence)." Huang Renxun's deepest fear is not losing revenue but becoming "blind".

"The intelligence window" is closing: China's AI ecosystem begins a non-CUDA-style breakthrough

GSR points out that although China's AI supply chain is still in its early stages compared to the West, a large number of cutting-edge explorations are no longer mere derivatives of the West.

The super-node (Super-node) experiments being conducted by Huawei, Alibaba, and ByteDance are highly pioneering in architecture. What alarms Nvidia even more is that in April this year, DeepSeek's V4 model released was tailored and highly optimized for Huawei's Ascend chips at the foundation, and did not grant Nvidia or AMD equal in-depth optimization rights. Currently, ByteDance, Tencent, and Alibaba are adding a large number of orders for Huawei's Ascend 950 series chips.

If Nvidia loses this window in China, it means it will be "entirely blind" in the evolution of global non-mainstream, yet explosively powerful AI architectures.

Lethal disconnection: The "scientists' arrogance" of the Chinese team

However, GSR's industry findings show that as "intelligence sentinels," Nvidia's Chinese team is seriously disconnected from the local market. Channel partners and collaborators reflect that Nvidia's Chinese employees often display a "scientist-like (Scientists) arrogance," rarely taking the initiative to contact customers, and in public or closed-door meetings, the speakers tend to deliver obscure speeches on the basic mathematics of machine learning, rather than discussing practical product implementation.

This mentality stems from the inertia of Nvidia's products being "hard currency" that doesn't need to be sold. However, by 2026, under the premise of the rise of local AI forces, this arrogance directly triggered two product and inventory disasters:

Disaster one: Almost无人问津 special supply card RTX 6000D

The RTX 6000D is a workstation GPU based on Nvidia's Blackwell architecture, specifically for the Chinese market, with specifications adhering to the U.S. limit (593 TFLOPS of FP4 computing power, 84 GB GDDR7 memory). Before the release, Nvidia strongly urged channel merchants to stock up in advance while concealing the core parameters, claiming it would "absolutely be a big seller".

However, after the release, the market found that compared to the global flagship, the 6000D was severely castrated. Fatally, it lacked NVLink interconnection and HBM memory, resulting in extremely poor cost-performance ratio. Nvidia had expected to stock a million units by the end of last year, and the reality was that several core partners' thousands of cards "none of them were sold".

Disaster two: The "resale" illusion of DGX Spark and hardware hard injuries

The official price of DGX Spark (a desktop AI supercomputer based on the GB10 chip) is $3,999, which also suffered a misjudgment. In China, channels and users, under unofficial implications, once believed they could link multiple Sparks via Ethernet at low cost to build a "reasoning server array".

However, the brutal technical limitations were soon exposed: DGX Spark has an interconnection limit of only 2 units at the hardware level, making it impossible to expand into a distributed reasoning cluster, and its 273 GB/s memory bandwidth faced intense criticism from the technical community. This caused channel merchants who blindly stocked up to suffer catastrophic losses, with some partners accumulating thousands of systems for months without moving.

More dangerous than sanctions is "being detached