Two H200s Reached a PLA-Linked University — and the Rule That Missed Them Still Takes 25% From NVIDIA

Northwestern Polytechnical University, one of China's \"Seven Sons of National Defense\" and on the U.S. Entity List since 2001, bought two NVIDIA H200 accelerators in October for ¥874,700 — about $123,000 — according to a procurement record dated Oct. 21. The chips came from Shaanxi Bosuan Network Technology, a domestic reseller, rather than from NVIDIA itself, and went to the university's management school. Two chips are noise on any revenue model. The transaction matters because it documents a route the licensing regime never closed: a purchase or a rental completed inside China leaves no trace in any license or import record.

In January 2026 the Commerce Department moved the H200 and AMD's MI325X to case-by-case licensing and attached four conditions: the Chinese buyer must run export compliance procedures including customer screening, the chip must clear independent third-party testing in the United States, the shipment must not delay orders for U.S. customers, and global capacity must not be diverted from U.S. end users. NVIDIA's July-quarter filing adds the commercial mechanics — licensed H200s are inspected on U.S. soil before shipment and carry a 25% duty on entry into the United States, which the company says it cannot pass on to customers.

Reuters reported in July that shipments had begun but the number was \"very few\": the chips enter the United States, take the 25% duty and are then re-exported only to approved Chinese buyers. That routing is the whole enforcement design in miniature. Every control in the rule attaches to a crossing of the border — a test certificate, a screening list, a customs entry — so a transaction completed on Chinese soil triggers none of them.

The volume ceiling is generous. The Council on Foreign Relations reads the rule as allowing exports to China of up to half the volume shipped to U.S. customers, roughly a million H200s on current run rates, and argues the certifications behind it are \"difficult or impossible to make credibly\" when high-bandwidth memory supply is inelastic and Commerce has few means to prove a certification false. IAPS' March 2026 assessment of H200 exports puts the implied allowance at about 890,000 GPUs and concludes that verifying end users and end use is nearly impossible once chips are exported, and that monitoring GPU end use is not feasible even where know-your-customer rules exist. The regime was never designed to survive contact with a rental contract.

The China line is the wrong place to look. In fiscal 2026 China including Hong Kong supplied NVIDIA $19.68 billion of $215.94 billion, or 9.1%, and once the licensing regime began the H200 channel itself fell below 1% of Data Center revenue. It has already cost $400 million in excess-inventory and purchase-obligation charges in the first half of fiscal 2027, inside $2.1 billion of total inventory and excess purchase-obligation provisions, and the 25% duty could not be passed through. The channel now carrying enforcement risk is the $42.35 billion Taiwan-billed line, more than twice the size of the China line.

The clearest read on what enforcement costs comes from Super Micro Computer, whose co-founder was charged in March over at least $2.5 billion of diverted servers. Its September-quarter filing, made Nov. 6, shows net sales of $19.95 billion against $5.02 billion a year earlier, with Asia falling from 46.2% of sales to 5.5% and the United States rising from 36.6% to 80.1%. One customer accounted for 44.0% of quarterly net sales and finished goods reached $11.59 billion, while trailing free cash flow ran negative $6.5 billion even as trailing net income was $3.66 billion.

The market has already made its judgment on the integration layer. Super Micro Computer trades at $27.79, 46% below its 52-week high of $51.40, on 0.37 times trailing sales with trailing revenue of $54.0 billion. A record order book, a de-rated multiple and negative trailing free cash flow is what an enforcement overhang looks like in a price. The same spread shows up inside NVIDIA, whose trailing free cash flow of $127.0 billion funds the very channel Washington is trying to close, and whose market value fell from $5.26 trillion on Apr. 26 to $4.07 trillion on Nov. 11 even as trailing revenue grew 19.5% over the same window.

Nothing in this story changes how many H200s get made. TSMC posted September-quarter revenue of NT$1,493.4 billion (US$46.7 billion) at a 66.4% gross margin, and the accelerator buildout driving it is indifferent to where the end customer sits. SK Hynix reported third-quarter revenue of ₩104.6 trillion and operating profit of ₩83.3 trillion, and expects HBM4 to overtake HBM3E in both volume and revenue in the fourth quarter. The constraint on supply is memory and packaging, not paperwork. Redirecting an H200 from a diversion channel to a U.S. buyer leaves every upstream margin untouched.

Demand is now blocked on both sides. Roughly 600,000 H200s are sitting in Hong Kong awaiting Beijing's import approval, while the major Chinese internet companies that hold licenses took about 10,000 chips each. Mainland buyers are paying for the scarcity: an eight-card H200 server was listed above ¥6 million by late October, roughly three times the price a year earlier, and approvals so far could total fewer than 200,000 chips. Withholding import approval is, in effect, a subsidy to domestic silicon.

Cambricon is the clearest listed beneficiary, with third-quarter revenue of ¥3.53 billion, up 104.5% year over year, and net profit of ¥1.40 billion. ZTE shows the other side: it reported record first-half 2026 revenue of RMB 78.0 billion with computing above 35% of the mix, and Reuters has documented it as one of the few Chinese firms holding a license to buy H200s. Holding a license is now a compliance liability rather than a commercial asset.

The October purchase is a $123,000 event attached to a $4.07 trillion company, and that ratio is the point. The licensing regime controls the border; the leak happened inland, through a rental and a domestic reseller, where no certificate, no screening list and no inspection reaches. Because Washington collects 25% of the volume it allows and the exporter cannot pass the cost on, the regulator has more reason to widen the channel than to police it. For investors, the consequence is not a revenue forecast — the China H200 line is under 1% of Data Center revenue and already carries $400 million of charges — but a multiple decision: NVIDIA's trailing sales multiple fell from 20.8 times in late April to 13.5 times on Nov. 11 while trailing revenue grew 19.5%, and the question for the next two quarters is whether the market stops paying for China optionality or starts paying for channel risk.