Hong Kong’s data centers may not have the juice to host the Nvidia H200s Beijing is funneling their way. The territory’s installed base spans just 47 facilities and about 581 MW, and with an average PUE of 1.62 the real grid draw runs well above the IT figure, leaving little headroom for a wave of power-hungry AI servers.
ByteDance and Tencent have each received roughly 10,000 H200s, the first significant flows since Washington licensed the accelerators for China. Regulators there are steering most of each company’s allowance, up to 100,000 units per firm, away from the mainland and toward Hong Kong. Every order still requires case-by-case sign-off from the National Development and Reform Commission. The blockade had already cut Nvidia’s China share from 95 percent to zero, in Jensen Huang’s telling, and Nvidia is said to hold about 500,000 H200s built for Chinese customers.
The power math is unforgiving. Each H200 pulls up to 700W. An eight-GPU HGX server therefore draws about 10 kW, and a 100,000-unit allowance works out to roughly 12,500 servers. Frontier training in China still runs on Nvidia gear, and DeepSeek could not get its R2 model to train on Huawei Ascend chips, which shows how hard the switch is. This week’s deliveries equal about 2.5 percent of the 400,000-plus H200s the three giants were cleared to buy in January.
