Nvidia H200 Access to China Reopens — Under Tight Allocation Controls

Small batches of H200 GPUs are entering mainland China, but U.S. licenses, Chinese approvals, and domestic-chip policy still determine where AI compute can actually be deployed.

Semiconductor Supply Chain Update | August 19, 2026

The H200 has become one of the clearest examples of how semiconductor availability is no longer defined by manufacturing capacity alone. A chip can exist in inventory, receive an export license and still remain commercially difficult to deploy because regulatory approval is required at several stages of the supply chain.

Earlier in 2026, Nvidia disclosed that the U.S. government had granted licenses allowing small amounts of H200 products to be shipped to specific China-based customers. At the end of Nvidia’s first fiscal quarter, however, the company said it had not generated revenue under the H200 licensing program and did not yet know whether imports would be allowed into China.

By July, U.S. officials said a very small number of H200 shipments had begun. The latest reported approvals suggest that the flow is now becoming more practical for selected Chinese AI companies, although access remains controlled rather than open.

01 | From Regulatory Limbo to Controlled Entry

The latest development is a change in degree rather than a complete change in policy. Chinese companies have shown strong demand for H200-class compute, while both Washington and Beijing have imposed conditions on how the chips may be purchased and used.

Financial Times reported on August 19 that small batches are now being allowed into mainland China. ByteDance and Tencent have each received about 10,000 H200 processors in recent weeks, while several other technology groups could receive permission for similar shipments.

02 | Why H200 Still Matters in the AI Supply Chain

H200 is based on Nvidia’s Hopper architecture and is no longer the company’s newest AI accelerator. However, it remains highly relevant because it combines mature software support, established server platforms and large HBM3e memory capacity.

Nvidia lists the H200 with 141GB of HBM3e memory and 4.8TB/s of memory bandwidth. For large-language-model training, inference and high-performance computing, memory capacity and bandwidth remain critical constraints.

03 | The Real Bottleneck Is Now Permission, Not Only Production

The H200 case demonstrates a new layer of semiconductor supply-chain risk. Traditional procurement analysis focuses on wafer capacity, packaging, lead time and inventory. Advanced AI chips now require an additional layer: regulatory usability.

A U.S. export license does not automatically mean a customer can deploy the product. Chinese buyers may still require domestic approval, while distributors and system integrators must ensure that destination, end user and system configuration remain compliant with applicable rules.

04 | Data-Center Location Is Becoming Part of Component Availability

Financial Times also reported that Chinese regulators are allowing companies to place H200 processors in Hong Kong. This creates an alternative deployment route, but not a simple substitute for mainland capacity.

AI servers require substantial power, cooling, networking and data-center infrastructure. A chip may be legally deployable in one location while the supporting infrastructure is insufficient to absorb demand at scale.

05 | Domestic AI Chips Remain a Strategic Counterweight

Beijing’s controlled approach reflects a second objective: maintaining access to high-performance foreign compute while continuing to support domestic semiconductor development.

The result is not a simple choice between Nvidia and domestic alternatives. Chinese AI companies may use imported accelerators where performance and software compatibility are critical while increasing adoption of domestic products in other workloads.