The United States has approved Nvidia H200 exports to selected Chinese companies, yet not one authorised chip has reportedly shipped. The obstacle is no longer Washington’s export licence. Beijing is preventing domestic companies from accepting hardware that US rules require them to use inside China.
Roughly ten companies, including Alibaba, Tencent, ByteDance and JD.com, hold licences for as many as 75,000 units each. Lenovo and Foxconn are approved distributors. The apparent opening has therefore produced a deadlock rather than renewed trade.
The situation shows how semiconductor policy can reshape a market even when both sides formally allow a transaction.
Conflicting rules make approved trade impossible
US conditions require Chinese buyers to keep authorised H200 chips in China. Beijing has instructed technology companies to restrict Nvidia equipment to overseas operations while prioritising domestic suppliers at home.
The two requirements cannot be satisfied simultaneously. Hardware cleared by Washington cannot legally be deployed where Beijing permits its companies to use it.
This contradiction explains why a presidential summit delivered no breakthrough. US Trade Representative Jamieson Greer said semiconductor controls were not on the bilateral agenda, despite President Trump travelling to Beijing with Nvidia chief executive Jensen Huang added to the delegation.
The blockage is structural rather than a missing form or unresolved quota. Beijing’s State Council is conducting a supply-chain security review intended to reduce reliance on US semiconductors, so approving the H200 domestically would weaken its own industrial objective.
Huawei gains while Nvidia’s licences remain idle
Chinese AI companies are increasingly optimising models for Huawei processors. DeepSeek adapted its V4 model for Ascend chips in training as well as inference, while Tencent expects domestic GPU supply to expand through 2026.
Alibaba says its T-Head accelerators have reached scaled mass production. These developments suggest that local hardware is moving from an emergency substitute toward a planned foundation for China’s AI ecosystem.
Nvidia’s share of revenue from China has fallen from more than 20% before tighter restrictions to roughly 5% in recent quarters. Its current guidance assumes no Chinese revenue.
The longer the H200 remains unavailable, the more workloads, software and engineering expertise move to domestic architectures. Even if trade later resumes, switching costs and policy priorities may prevent Nvidia from recovering its earlier position.
Huang’s presence at the summit demonstrated commercial urgency, but executive relationships cannot easily reverse a state-directed supply-chain strategy.
Hardware architecture becomes an instrument of sovereignty
The stalemate matters beyond lost sales. China is effectively choosing which computing architecture will support its models, cloud platforms and developer tools through government direction rather than benchmark performance alone.
Beijing is betting that the performance gap between domestic accelerators and Nvidia’s products will close quickly enough to justify short-term constraints. DeepSeek’s optimisation work suggests that software can compensate for at least some hardware disadvantages, particularly in inference.
The US also faces a tradeoff. Allowing controlled H200 sales could generate revenue and keep Chinese developers attached to Nvidia’s ecosystem. Strict deployment conditions aim to limit strategic risk but may accelerate the independent stack Washington sought to contain.
For enterprises, the episode is a reminder that availability cannot be assessed from an export licence alone. Procurement plans need to account for rules in both jurisdictions, domestic industrial policy and the portability of models between architectures.
The H200 deal is approved on paper and frozen in practice. Its failure reveals a semiconductor market divided not only by product capability but by incompatible definitions of acceptable dependence.
If Huawei and other Chinese suppliers use the opening to establish a complete domestic ecosystem, the lasting consequence will be larger than one lost Nvidia order. The world’s second-largest AI market may settle on a separate hardware foundation, making geopolitical policy a permanent layer in AI architecture.
