Nvidia’s H200 has finally crossed into mainland China. ByteDance and Tencent have each taken delivery of roughly 10,000 of the accelerators in recent weeks, the Financial Times reported Wednesday, citing two people with knowledge of the matter. It is the first meaningful movement of the chips into the mainland since Washington approved their export last December.
That sounds like the chip war thawing. It isn’t. Washington licensed roughly ten Chinese firms — Alibaba, ByteDance, Tencent and JD.com among them — for up to 100,000 units apiece. Beijing is directing most of that allowance somewhere else: Hong Kong, which operates under a separate customs regime and where mainland teams can still reach the hardware remotely over cross-border links. Every individual purchase still needs case-by-case sign-off from the National Development and Reform Commission.
The 10,000-unit deliveries amount to roughly 2.5% of the 400,000-plus H200s that ByteDance, Alibaba and Tencent were collectively cleared to buy in January. Nvidia is reported to be sitting on around 500,000 H200s built largely for Chinese customers.
The constraint moved from paperwork to power
Here is the number that actually decides this story. An H200 draws up to 700 watts, and a fully loaded eight-GPU HGX H200 server pulls roughly 10 kilowatts. Run the arithmetic on a single company’s 100,000-unit allowance and you get about 12,500 servers and roughly 125 megawatts of IT load. Nvidia’s reported 500,000-unit stockpile works out to something near 625 MW.
Hong Kong’s entire installed data-centre base runs to 47 facilities totalling about 581 MW, per Hong Kong Free Press. The territory’s average power usage effectiveness of 1.62 means grid draw sits well above that IT figure. So the destination Beijing has chosen for these chips does not have the headroom to run one licensed allocation, let alone several — and that is before anyone tries to evict the banks, telcos and cloud tenants already occupying those halls.
“It’s a dilemma,” a person familiar with the situation told the FT, describing companies unable to find anywhere in the territory to deploy the chips. The Northern Metropolis data-centre cluster meant to fix the shortfall — awarded to Range Intelligent Computing in March — is not slated to come online until 2029.
Our take: The export licence stopped being the binding constraint. Trump cleared these chips last December in exchange for a 25% cut of every sale to the US Treasury, and Jensen Huang still told investors Nvidia’s China share had gone from 95% to zero — because Beijing, not Washington, was holding the gate. Now Beijing has cracked it open and pointed the volume at a city with 581 MW of total capacity and a fix arriving in 2029. Rationing by geography is cheaper than rationing by decree and much harder to complain about. The mainland build-out stays pointed at domestic silicon, the licences stay technically honoured, and nobody has to say no.
What this means for the compute map
Chinese labs increasingly run inference on domestic accelerators while still training frontier models on Nvidia hardware. That split is why the destination matters: DeepSeek is already designing its own inference chip to cut reliance on Nvidia and Huawei alike, but training is the workload that needs dense, contiguous, power-hungry clusters. Pushing the H200 fleet offshore doesn’t remove access to it — it converts owned capacity into remote capacity, with latency, cost and a cross-border network dependency attached.
There is a second lesson here that has nothing to do with China. Chips are no longer the scarce input in AI — energised floor space is. Microsoft spent an earnings call explaining why its 2.2 million chips outpace the shells ready to hold them. Hong Kong is the same problem with a border drawn through it. Anyone modelling accelerator shipments as a proxy for deployed compute is measuring the wrong end of the pipe.
What to watch
- NDRC approval cadence. If the next tranche is another 10,000 rather than a step change, the throttle is deliberate policy, not a processing backlog.
- Hong Kong power announcements. Any emergency grid allocation, colocation expansion or pull-forward of the Northern Metropolis timeline signals Beijing wants these chips genuinely usable.
- Nvidia’s inventory language. A reported 500,000-unit stockpile is capital sitting idle. How the company characterises it at the next earnings print tells you what it thinks realistic sell-through is.
- Whether Alibaba and JD.com get shipments. Licensed but so far undelivered. Their turn is the test of whether this is a policy shift or a gesture.
Ten thousand chips is the headline. Five hundred and eighty-one megawatts is the ceiling. Watch the second number.