600,000 NVIDIA GPUs Are Stranded in Hong Kong — and Beijing Now Holds a $51 Billion Pricing Option

On Nov 5, 2026, Reuters reported that roughly 600,000 Nvidia AI chips — reported as H200 units — are stored in Hong Kong, awaiting mainland import approval that Beijing has not granted. These are not unshipped orders: the parts were built, licensed for export by Washington and physically delivered to Hong Kong. What they are missing is a Chinese customs signature. Some chips have cleared in recent months, with major Chinese internet firms each receiving about 10,000 H200s, and total approvals reportedly may fall below 200,000 units.

Washington opened the door in January 2026 and priced it. The Commerce Department moved the H200 and AMD's MI325X from presumption of denial to case-by-case licence review, conditioned on independent third-party testing in the United States, purchaser export-compliance procedures, and no loss of capacity for U.S. customers. Beijing then closed the door on the other side: customs agents were told the H200 was not permitted to enter, and officials told domestic firms not to buy unless necessary. Both conditions must hold, so the binding constraint is now the one Beijing controls.

The clearest read on how tight Chinese compute has become is the gap between what the chips officially cost and what buyers in China are being asked to pay. Reuters put the official China price of an H200 at about $27,000 in January 2026. By the end of October, an eight-card H200 server was on offer in mainland China for more than 6 million yuan, about $900,000, and servers running Blackwell — which Chinese buyers cannot legally be sold at all — were quoted above 20 million yuan each.

Multiply that spread across 600,000 chips and the prize Beijing now holds is about $51 billion — the difference between what the official price implies and what the street is quoting. This is an arithmetic illustration of scarcity value, not a realizable gain: the resale market for 600,000 cards at those prices does not exist, and Reuters describes offers, not completed transactions. What the arithmetic does show is scale. A decision worth tens of billions of dollars to Chinese buyers is now being made by ministries that declined to comment.

The striking detail in Nvidia's own disclosures is that the stranded queue barely registers. In the Form 10-Q for the quarter ended July 26, 2026, the company booked a $0.4 billion charge tied to H200 excess inventory and purchase obligations as demand diminished, and recorded $985 million of inventory and excess purchase-obligation provisions in the quarter ($2.1 billion across the first half). Its CFO commentary for the same quarter states that Hopper shipments to China were under 1% of Data Center revenue, and that the Q3 FY2027 outlook of $108.0 billion assumes no China data-center compute revenue at all.

The clearest test of whether this is a demand event or a policy event is what happened to the companies that make the parts. Nothing. Nvidia halted China-bound H200 output in March and refocused the TSMC capacity elsewhere; TSMC's fourth-quarter guidance went up, not down. And the HBM sitting inside the stranded cards was bought and paid for before the queue existed — a customs pause delays a delivery, it does not cancel a wafer or a memory stack.

Demand is the part that is clearly intact. ByteDance, Alibaba and Tencent together ordered more than 3 million RTX PRO 5500 workstation GPUs at a combined sticker price above $35 billion, and Chinese orders in total exceed 6 million cards. What changed is supply and price. Nvidia told partners they might receive only 5% to 10% of the quantities requested, and said it could produce no more than roughly 3 million of those cards globally in 2027. The approved alternative is rationed, and the unapproved original is quoting multiples of list.

IDC's first-half 2026 shipment data show where the blocked demand is going. Nvidia shipped 678,000 AI accelerators into China, down 44% year over year, while Huawei shipped 572,000, up from more than 300,000. GPU shipments across China fell 26% to 991,000 units, but non-GPU accelerators rose nearly 80% to 1.016 million. Beijing is not banning compute; it is re-routing it. Bernstein, cited by TrendForce, sees Chinese firms lifting domestic silicon from 30% to 46% of accelerator budgets over the next twelve months and Nvidia's share of China's high-end AI chip market falling to about 8% in 2026 from roughly 40% in 2025.

The next quarter is about a single number — how many chips Beijing releases — and the next three years are about who supplies Chinese inference once the decision is embedded in procurement plans. Nvidia reports third-quarter results on Nov 17, 2026, and Tencent's board meets Nov 17 with results due Nov 18. Both prints matter more than usual, because they show whether the compute constraint is showing up in capex discipline or in allocation.

For most of the AI cycle the question was whether Washington would let Nvidia sell to China. It is now whether Beijing will let China buy, and at what price. The first question has an answer — licences exist, shipments are small, and the queue in Hong Kong is the residue. The second question is worth roughly $51 billion of spread on 600,000 chips, and the premium accrues to the holder of the import right rather than to the seller of the silicon. That is why the trade here is not a US chip-multiple trade: upstream is paid either way, the listed beneficiaries of substitution sit in Shanghai, and the risk in global AI multiples is a slow fade of Chinese order flow into domestic silicon — plus a narrow, dated and binary catalyst on the other side of the gate.