Reports about H200 reaching Chinese customers now have an additional primary source: Nvidia’s quarterly filing. It confirms limited shipments, while leaving important customer-level quantities and deployment details unspecified.
What was reported in August
An August Reuters report, citing the Financial Times, described roughly 10,000 H200 chips each for ByteDance and Tencent and US permissions for up to 100,000 each. Reuters said it could not independently verify the report. These quantities should remain attributed reporting, not be rewritten as a customer inventory audited by this site.
What Nvidia subsequently confirmed
Nvidia's fiscal 2027 second-quarter filing, for the quarter ended July 26, describes US licences for limited H200 sales to specific Chinese customers, alongside Chinese restrictions. It says a fraction of allowed shipments occurred and represented less than 1% of Data Center revenue in the most recent quarter.
The filing also describes required US inspection and an associated 25% import tariff, which Nvidia says it did not pass to customers. That company disclosure is more specific than saying every GPU sold into China carries an identical customer surcharge.
The filing supports the existence of shipments. It does not independently confirm every named customer's reported allocation, rack location or utilization rate.
Five stages that should not share one number
An export permission sets an authorized scope. An order expresses a commercial purchase. Shipment and delivery move equipment. Commissioning turns installed hardware into usable capacity. Utilization describes how much of that capacity is actually working. Advancing through one stage does not establish completion of the others.
For an illustrative inventory, 100 permitted units, 60 delivered and 40 commissioned would mean 40 usable units at that point, not 100. These invented numbers explain the cover; they are not H200 customer figures. Even the 40 commissioned units would not tell us their workload occupancy.
The earlier article inferred that equipment must sit idle because named companies lacked particular Hong Kong facilities. That does not follow: rented capacity, actual destination and commissioning status would need separate evidence. Likewise, a reported shipment count cannot establish whether devices came from old inventory or new production, or predict availability for unrelated buyers. Capacity planning needs confirmed delivery and commissioning dates for the specific system, not the largest number in a licence headline.
Add NVIDIA quarter filing confirming limited H200 shipments, qualify Reuters/FT customer counts, remove unsupported Hong Kong idle-capacity and inventory assumptions.