Nvidia's H200 accelerators are finally moving into mainland China, though in volumes small enough to show how tightly Beijing still controls the tap. ByteDance and Tencent have each taken delivery of roughly 10,000 of the chips in recent weeks, the Financial Times reported Wednesday, citing two people with knowledge of the matter. It is the first meaningful flow of the processors across the border since President Trump cleared their export in December, and several other Chinese technology groups are expected to win approval for shipments of similar size.
Measured against what both governments have already authorized, the deliveries are a rounding error. US licenses allow individual Chinese buyers to purchase as many as 100,000 H200s each, and ByteDance, Alibaba and Tencent were collectively approved in January to buy more than 400,000 units. The 10,000-unit deliveries amount to roughly 2.5 percent of that figure. Every order still requires case-by-case sign-off from China's National Development and Reform Commission. Lenovo and other Nvidia partners told Chinese customers last week that they could begin placing orders again for AI servers built around the H200, according to the FT, but the NDRC gate stays in place for each one.
The larger constraint is geographic. Beijing wants most of each company's licensed allowance kept off the mainland entirely, and regulators have directed buyers toward Hong Kong, which sits outside mainland China's customs border and is covered by the same US export licenses. The problem is that Hong Kong cannot power the hardware. An H200 draws up to 700 watts, and a fully loaded eight-GPU HGX server pulls around 10 kilowatts, meaning a single 100,000-unit allowance works out to roughly 12,500 servers and 125 megawatts of IT load. Hong Kong's entire installed base is 47 data centers totaling about 581 megawatts, and the territory's average power usage effectiveness of 1.62 pushes actual grid draw well above the IT figure. The Northern Metropolis data center cluster meant to close that gap, awarded to Range Intelligent Computing in March, is not due online until 2029.
"It's a dilemma. Everyone needs the chips but struggles to find a way to use them in Hong Kong," a person familiar with the situation told the Financial Times. "The hope is for the control to loosen up gradually."
For Nvidia, a trickle beats nothing, but it is far from the reopening the company has spent the year arguing for. Nvidia is sitting on roughly 500,000 H200s built largely for Chinese customers, a stockpile that would need about 625 megawatts to run. Trump approved H200 exports last December in exchange for a 25 percent cut of every sale going to the US Treasury, and Washington had licensed around 10 Chinese firms, including Alibaba, ByteDance, Tencent and JD.com, by May. Beijing simply never let the orders flow, a blockade that left chief executive Jensen Huang telling investors the company's share of the Chinese market had gone from 95 percent to zero. After the US licenses came through in March, Huang said: "We have received purchase orders, and we're in the process of restarting our manufacturing." As recently as mid-July, a US trade official told a congressional hearing that only a very small quantity of chips had actually shipped against those licenses.
Why It Matters
The H200 has become a bargaining chip that two governments are moving in opposite directions at once. Washington converted an export ban into a revenue stream, taking a quarter of each sale while keeping Blackwell-class silicon off the table entirely. Beijing turned import approval into an industrial policy lever, releasing just enough compute to keep its frontier labs competitive while steering the mainland market toward Huawei and other domestic chipmakers.
The timing reflects real pressure on Beijing. Moonshot released its K3 model last month with performance approaching the best US systems, and Alibaba, DeepSeek and Z.ai followed with models of comparable capability. Chinese developers increasingly run inference on domestic accelerators, but training frontier models still leans on Nvidia hardware. DeepSeek's failed attempt to train R2 on Huawei Ascend chips is the cautionary example the industry keeps citing. The H200 is at least two generations behind Nvidia's current parts, which makes it exactly the kind of concession Beijing can live with: useful enough to keep its labs moving, not good enough to undercut the case for building a domestic supply chain.
Routing volume through Hong Kong is that compromise rendered in physical form. Chinese labs get access to compute, and the mainland market stays reserved for homegrown silicon. That the arrangement does not currently work, for lack of power and floor space, may be less a bug than a feature.
What To Watch
Nvidia reports second-quarter results on August 26, and its $91 billion revenue guidance assumes zero data center compute revenue from China. Whether finance chief Colette Kress now attaches a number to Chinese demand, or keeps excluding it, will say more about how durable this opening is than shipment totals do. Watch the pace of NDRC approvals and whether Alibaba, JD.com or ZTE confirm deliveries of their own. Watch Hong Kong's power and data center pipeline, because nothing at scale can land there before the Northern Metropolis site arrives at the end of the decade. And watch whether Beijing quietly permits more licensed volume to stay on the mainland. That, rather than the first 20,000 chips, would mark an actual shift.
“We have received purchase orders, and we're in the process of restarting our manufacturing.”— Jensen Huang, Chief Executive, Nvidia