NVIDIA CEO Huang Jen-Hsun’s recent low-key visit to Shanghai, where he was seen bargaining at a local wet market in an ordinary shirt, is far more than a casual public relations stunt. Behind this display of approachability lies a critical mission: navigating the extremely complex and high-stakes battle to sell NVIDIA’s advanced H200 AI chips in the Chinese market.
The core issue is no longer just U.S. export restrictions. While the previous Trump administration has conditionally approved H200 sales to China with a hefty 25% tariff, the new hurdle lies on the Chinese side. According to reports from Bloomberg and others, Chinese customs authorities have currently halted the clearance of H200 chips. Major Chinese tech companies, desperate for top-tier computing power to train large AI models and compete with giants like OpenAI, reportedly want to place massive orders worth billions. However, they have been advised to hold off.
China’s reported new strategy is a “bundling” or “matching” policy. The logic is strategic management, not an outright ban. Regulatory bodies might allow some critical AI companies to import a limited number of H200 chips, but with a strict condition: these companies must simultaneously purchase a significant proportion of domestic Chinese GPUs. Market rumors suggest a possible 1:1 ratio. This means for every H200 chip bought, a company must invest in a Chinese alternative.
This approach serves multiple purposes. It acknowledges the current reality that domestic chips, from companies like Huawei, Biren, and Moore Threads, still lag behind NVIDIA’s offerings in raw performance and, crucially, in software ecosystem maturity (like NVIDIA’s CUDA). An unrestricted flood of H200 chips could stifle the domestic industry by denying it the user feedback and iterative development necessary to improve. The bundling policy forces the massive Chinese AI application market to also “feed” and test the domestic hardware ecosystem, using market demand to buy time for technological catch-up.
For NVIDIA and Huang, this creates immense uncertainty. Reports indicate that production of key PCB components for the H200 in China has been paused due to this market ambiguity. NVIDIA’s real “moat” is its CUDA ecosystem. By potentially locking Chinese developers deeper into CUDA through H200 sales, NVIDIA could preemptively neutralize future competition from rising Chinese GPU ecosystems. However, China’s conditional access policy directly counters this by mandating support for domestic alternatives.
The situation reveals a new, more transactional phase in the U.S.-China tech rivalry. It’s moving past simple decoupling to a complex, quid-pro-quo game. The U.S. uses tariffs and conditional approvals; China responds with its own conditional market access. For China, the goal is clear: use its vast market to nurture its own tech sovereignty and industrial chain. For Chinese AI companies, it’s a difficult balance between securing immediate, vital computing “oxygen” and ensuring long-term supply chain security and supporting national technological goals.