Less Than 1% Against $5.6bn (September 2026): The Licensed Route Into China Is a Rounding Error — and Nvidia's Own Geography Table Disclaims the Rest

Two numbers published within a fortnight of each other describe the same subject and refuse to sit in the same sentence. Nvidia's 10-Q for the quarter ended 26 July 2026 says that shipments made under the US government's H200 licensing programme — the legal, permitted, politically negotiated route into China — account for less than 1% of Data Center revenue, against $89.0 billion of Data Center revenue in the quarter. On 6 September the New York Times reported that one California server maker, the renamed US operation of a blacklisted Chinese parent, exported at least $5.6 billion of advanced technology to Southeast Asia over twenty-two months, more than $3 billion of it Blackwell-based systems ultimately serving Chinese customers. With Xi Jinping due at the White House on 24 September and a first bilateral AI dialogue tentatively pencilled in, the question worth understanding is not who is guilty. It is why a control regime that works on boxes keeps missing a product that is sold by the hour.

The two numbers, and why they belong to different systems

Start with the filing, because it is the least contested document in the story. Nvidia disclosed that the US government granted licences allowing it to ship small amounts of H200 products to specific China-based customers, that those sales were then restricted by the PRC government, and that the company has been unable to sell all the product for which it holds licences. It took a $0.4 billion charge in the first half of fiscal 2027 for excess H200 inventory and purchase obligations as demand for the part diminished. The licences require each unit to pass a US inspection before shipment, which attracts a 25% tariff on importation into the United States, a cost the company says it has been unable to pass to customers.

That is what a fully legal, bilaterally sanctioned trade channel looks like after two years of negotiation: under 1% of Data Center revenue, a write-down attached, a tariff on the way through, and a buyer's own government discouraging the purchase.

Now set the other number beside it. Twenty-two months, $5.6 billion, more than half of it Blackwell-based systems. Divide it out and the run rate is roughly $255 million a month of hardware following a path the rules did not contemplate — not smuggled in a suitcase, but exported through ordinary channels to ordinary destinations and then, per the reporting, ending up serving customers the controls were written to exclude.

What the rule actually controls: an owner, a box, a border

Export control is a property regime. It attaches to an item, asks who owns or controls it, and asks which border it crosses. Everything in the architecture follows from that, and so does everything that leaks out of it.

Inspur Group was added to the Entity List in March 2023; six of its subsidiaries were added in March 2025. In September 2025 BIS introduced its affiliates rule, extending listings down through ownership chains rather than name by name — an admission in policy design that the entity-by-entity approach was being outrun by corporate restructuring. Then on 31 May 2026 Commerce issued guidance confirming that licence requirements for advanced AI chips apply to all businesses with headquarters or a parent company in China, wherever the business is physically located. Asked whether it was enforcing pre-existing licence requirements after the previous administration's diffusion framework was scrapped, BIS replied: "The answer is yes." Nvidia's response at the time was that the guidance reaffirmed its existing approach — licences are required to ship controlled products to PRC-headquartered companies.

Read those three steps as one motion. Each closes the previous gap by extending the ownership test: from the parent, to the subsidiaries, to the headquarters. Every extension is a tighter answer to the question "whose box is this?" None of them touches the question of who is using it.

The second gap: renting the chip instead of owning it

The diversion story is the visible one because it leaves shipping records. The larger one leaves almost nothing.

CNBC reported in August 2026 that Chinese hyperscalers including ByteDance, Alibaba and Tencent have accessed Nvidia compute remotely via facilities in Thailand, Malaysia and Japan. Michael Kratsios, the White House science and technology adviser, accused Moonshot AI of using GB300 chips through a Thai facility less than a week after the company released a new model in July. ByteDance's reported arrangement runs through Aolani, a Singapore-headquartered cloud provider with Nvidia hardware in Malaysia; Aolani told CNBC that the companies it serves have no ownership, future claim or physical access to the chips powering its solutions, and that permitted access is fully compliant with applicable regulations.

That statement is worth reading twice, because it is almost certainly accurate and it is also the entire problem. Non-ownership is the compliance argument. Non-ownership is also irrelevant to model capability, which depends on accelerator-hours consumed, not on title held.

The infrastructure making this possible is being built for reasons that have nothing to do with export control. Real-estate firm JLL estimates global data-centre capacity could roughly double to 200GW by 2030, and DC Byte counts 31 planned 100MW-plus facilities across Malaysia, Indonesia and Thailand against just two today. That build-out would be happening if no export control existed. It simply creates, as a by-product, a large pool of advanced compute inside jurisdictions that are neither the exporter nor the destination the rules are aimed at.

Why the geography table cannot settle the argument

Here is where the market's favourite shortcut breaks. When people want to know where the chips went, they reach for Nvidia's revenue-by-geography disclosure. That disclosure explicitly tells you it cannot answer the question.

Nvidia states that revenue by geographic region is designated based on the location of the headquarters of direct customers, and that the end customer and shipping location may be different from that headquarters location. Direct customers are add-in-board makers, distributors, ODMs, OEMs, cloud providers, model developers and system integrators — the buyers of record, not the operators of the buildings.

For the quarter ended 26 July 2026: United States $60,074m, Taiwan $26,985m, China including Hong Kong $7,880m, other $1,282m, on total revenue of $96,221m. The Taiwan line alone is 28.0% of company revenue and has come from $8,902m a year earlier — roughly a tripling. That is not a statement about demand in Taiwan. It is a statement about where the contract manufacturers that assemble AI systems keep their headquarters, and those systems ship onward to whoever ordered them.

So the most-cited public dataset on chip destinations resolves, on inspection, into a dataset about corporate registration. Anyone building a view on how much advanced compute sits within reach of Chinese developers is working from an input whose own publisher has disclaimed that use. Keeping that straight is the same discipline that applies to reading a compute target without its precision attached, which is the trap examined in China's 9,800 EFLOPS five-year target.

What would actually close it, and what closing it would cost

The drafted answer is the Remote Access Security Act, H.R.2683, which would extend BIS authority to remote access of items subject to the Export Administration Regulations. The House passed it 369–22 on 12 January 2026 — a margin that tells you the political question is settled even though the legal one is not. The Senate companion, S.3519, has not moved.

Two things follow. First, enactment would be a beginning rather than an end: BIS would still have to write a rule specifying which compute is covered, who may not reach it and what customer-verification scheme providers must operate. King's assessment was that BIS could move quickly with White House backing, and that the difficulty lies in drafting something enforceable. Second, the cost lands on cloud providers, who would carry the know-your-customer burden — which is why industry pushback is expected, and why the timeline is a function of lobbying as much as of drafting.

This is the point where politics becomes a price input rather than a subject. Xi is due at the White House on 24 September. Reuters reported that a first bilateral AI dialogue under this administration, led on the US side by Treasury Secretary Scott Bessent, was tentatively planned for mid-September, while a White House official said there was no planned AI-related meeting in that window. Whatever the venue, chip access is the highest-value item on the table, and a negotiated outcome pushes in the opposite direction to a tightening rule. Anyone reading headlines out of that summit should keep the two tracks separate: a diplomatic loosening changes the licensed channel currently worth under 1% of Data Center revenue, while a remote-access rule would change something considerably larger that does not appear in the China line at all.

Where this reaches the instruments people trade, and where it does not

Be precise about the transmission, because the loose version of this story is wrong in both directions.

The revenue at stake in the diversion narrative is small. Twenty-two months at $5.6 billion is about 5.8% of one Nvidia quarter, and the hardware in question was sold and recognised already. Enforcement against a diverter does not claw back a past sale; it removes a future buyer. A market that already marks the licensed China channel at approximately zero cannot be surprised much further down that particular line — a point developed in the quarter preview.

The revenue at stake in a remote-access rule is potentially larger and sits somewhere nobody is currently counting. Demand from third-country cloud operators is booked to whichever entity signs the purchase order, and it is part of the same flow that has driven the Taiwan line to 28% of revenue. A rule requiring those operators to vet and exclude end users adds compliance friction to a fast-growing buyer segment. That is an index-level consideration rather than a single-name one, because the concentration of a handful of AI-complex names in US500 and NAS100 means a policy shift affecting their order books reaches the index through its largest weights.

The currency channel is the honest place to stop short. There is no clean FX trade in an export-control rulemaking, and pretending otherwise would be the same category error as reading a headquarters table as a shipping manifest. What exists is second-order: chip policy is one input into the risk-sentiment factor, and a summit outcome read as de-escalation is mildly risk-positive in the way that usually flatters the growth-linked currencies and takes a little of the haven bid out of the dollar. That is a small, conditional, easily-swamped effect, and the live read on it belongs on the USD factor page rather than in a forecast here. The reasoning behind separating factors that way is set out on the about page.

What would change the picture

Four things, in rough order of how much they would move:

Senate action on S.3519. A committee vote would convert a settled political question into a live legal one and start the BIS rulemaking clock.

A BIS rule on remote access, with or without the statute. Watch for the definition of covered compute and the customer-verification standard; the enforceability of those two clauses is the whole rule.

An outcome from the 24 September summit that touches chips explicitly. A licensing expansion and a remote-access crackdown are opposite signals and can arrive in the same week.

Any enforcement action naming an entity in the third-country chain. Not because the penalty would matter financially, but because it would signal that Commerce believes existing authority already reaches the conduct — which would remove the need for the statute and accelerate everything above.

Until one of those lands, the gap stays open because it is definitional, and the most reliable thing to do with the story is to stop treating published geography as geography. The related question of what happens to prices when policy reroutes physical supply is traced in mature-node chip pricing.

Educational macro context only — not investment advice.