The First H200s Reached China. Beijing Told the Buyers to Leave Them in Hong Kong.

ByteDance and Tencent have each taken delivery of roughly 10,000 Nvidia H200 processors in recent weeks, and Chinese regulators have instructed them to keep most of the hardware in Hong Kong rather than install it on the mainland. The Financial Times reported it; Engadget carried it at 03:11 EST on 19 August, 08:11 UTC.

The instruction

The stated reason is protective rather than permissive: Beijing wants to support its domestic chip industry as Chinese designers build alternatives. Hong Kong sits outside the mainland customs border, so hardware parked there has been imported without displacing domestic accelerators in mainland data centres.

What the common telling gets wrong

Three layers. First, a licence ceiling is not a delivery. The United States cleared each company to buy up to 100,000 chips; about 10,000 arrived. The gap is 90%, and nothing in the reporting commits anyone to closing it.

Second, delivered is not deployed. Hong Kong does not have the data-centre shells or the power to absorb tens of thousands of H200s. Chips in a bonded facility are inventory, not training runs, and no reported figure describes any of this silicon as racked and running.

Third, the disclosure is the event, not the shipment. The deliveries happened "in recent weeks." Anyone writing this as "Nvidia just shipped to China" has the tense wrong.

Where the constraint now sits

This is the first concrete evidence of volume moving under the licensing regime that followed the lifting of the H200 ban, and it relocates the binding constraint. Export-control analysis that treats Washington's licence caps as the ceiling on Chinese AI compute is now modelling the wrong variable: the limiting decision in this case was made in Beijing, and it cut the effective number by an order of magnitude below what Washington permitted.

How firm is this

Worth stating plainly: the chain originates in a single paywalled report resting on unnamed sources. Reuters said it could not independently verify it, and Nvidia has not commented. The figures are consistent across the outlets carrying it, but nothing here is confirmed by a named party.