First Nvidia H200 shipments have reached China with ByteDance and Tencent taking deliveries after Beijing relaxed import restrictions, though most licensed units must remain in Hong Kong due to mainland power constraints.

Roughly 10,000 Nvidia H200 accelerators have been delivered to both ByteDance and Tencent in recent weeks, marking the first meaningful movement of the chips into mainland China since President Trump authorized their export last December, according to a Financial Times report citing two individuals with knowledge of the matter.

Beijing has directed that most of each company’s U.S.-licensed allowance—estimated by the FT at up to 100,000 units apiece—remain off the mainland, steering purchases toward Hong Kong instead. Nvidia is reportedly holding approximately 500,000 H200 units manufactured primarily for Chinese clients. The Trump administration cleared H200 exports to vetted Chinese buyers last December, contingent on a 25% levy on every sale remitted to the U.S. Treasury. By May, Washington had issued licenses to roughly ten firms, including Alibaba, ByteDance, Tencent, and JD.com.

Initially, Beijing prevented orders from moving forward, requiring case-by-case approval from the National Development and Reform Commission (NDRC). This regulatory blockade prompted Nvidia CEO Jensen Huang to inform investors that the company’s market share in China had plummeted from 95% to zero. As recently as mid-July, a U.S. trade official testified before Congress that only a minimal quantity of chips had actually shipped under the licenses. However, Lenovo and other Nvidia partners informed Chinese customers last week that H200-based AI server orders could resume, though each transaction remains subject to separate NDRC authorization.

The physical deployment of these accelerators faces significant infrastructure hurdles. Each H200 draws up to 700 watts, and a fully loaded eight-GPU HGX H200 server consumes approximately 10 kilowatts. At that density, a single company’s 100,000-unit allocation would require roughly 12,500 servers and generate a 125-megawatt IT load, while Nvidia’s reported 500,000-unit inventory translates to about 625 megawatts.

Hong Kong’s existing capacity cannot absorb this demand. According to the Hong Kong Free Press, the territory operates 47 data centers with a combined capacity of approximately 581 megawatts. With an average power usage effectiveness (PUE) of 1.62, actual grid consumption significantly exceeds the IT load figures. “It's a dilemma,” a person familiar with the situation told the FT, noting that companies are struggling to find adequate space within the territory to deploy the hardware. The Northern Metropolis data center cluster, intended to address this shortfall and awarded to Range Intelligent Computing in March, is not scheduled to become operational until 2029.

The recent 10,000-unit deliveries represent merely 2.5% of the more than 400,000 H200s that ByteDance, Alibaba, and Tencent were collectively permitted to purchase in January. Because the NDRC continues to gate every order and regulators are channeling the majority of licensed volumes to a jurisdiction lacking sufficient hosting capacity, the strategy effectively keeps the mainland market captive for domestic semiconductor manufacturers.