First Nvidia H200 shipments reach China, ByteDance and Tencent take deliveries as Beijing loosens its import block — most licensed chips must stay in Hong Kong, which can’t power them

ByteDance and Tencent have each taken delivery of roughly 10,000 Nvidia H200 accelerators in recent weeks, the Financial Times has reported, citing two people with knowledge of the matter, a development that would mark the first meaningful movement of the chips into mainland China since President Trump approved their export last December.

Beijing wants most of each company’s U.S.-licensed allowance, which the FT puts at up to 100,000 units apiece, kept off the mainland, with regulators directing buyers toward Hong Kong instead. Nvidia is reportedly holding around 500,000 H200s built largely for Chinese customers. Trump cleared H200 exports to approved Chinese customers last December, in exchange for a 25% cut of every sale to the U.S. Treasury, and Washington had licensed roughly 10 firms, including Alibaba, ByteDance, Tencent, and JD.com, by May.

However, Beijing never let orders flow: every purchase requires case-by-case approval from the National Development and Reform Commission, and the resulting blockade left Jensen Huang telling investors that Nvidia’s China market share had fallen from 95% to zero. As recently as mid-July, a U.S. trade official told a congressional hearing that only a very small quantity of chips had shipped against the licenses. Lenovo and other Nvidia partners told Chinese customers last week that they could resume orders for H200-based AI servers, per the FT, though each purchase still needs separate NDRC sign-off.

And just because chips are moving doesn’t mean the infrastructure is there to support them. The H200 draws up to 700W, with a fully loaded eight-GPU HGX H200 server pulling roughly 10 kW. At that rate, a single company’s 100,000-unit allowance works out to about 12,500 servers and 125 MW of IT load, and Nvidia’s reported 500,000-unit stockpile translates to roughly 625 MW.

Hong Kong’s entire installed base runs to 47 data centers totaling about 581 MW, per Hong Kong Free Press, and the city’s average PUE of 1.62 pushes actual grid draw well above the IT figure. “It’s a dilemma,” a person familiar with the situation told the FT, with companies unable to find anywhere in the territory to deploy the chips. The Northern Metropolis data center cluster meant to fix the shortfall, awarded to Range Intelligent Computing in March, isn’t slated to come online until 2029.

Chinese labs increasingly run inference on domestic accelerators but are still thought to train frontier models on Nvidia hardware, and DeepSeek’s failed attempts to train its R2 model on Huawei Ascend chips show why.

The 10,000-unit deliveries amount to roughly 2.5% of the 400,000-plus H200s that ByteDance, Alibaba, and Tencent were collectively approved to buy in January, but the NDRC still gates every order, and steering the bulk of licensed volumes to a territory that can’t house them keeps the mainland market captive for domestic chipmakers.