The guidance that assumes China buys nothing

The most informative line in Nvidia's current guidance is a negative, and it is the company's own. Revenue for the third quarter of fiscal 2027 "is expected to be $108.0 billion, plus or minus 2%", the release of 26 August says, and then: "NVIDIA is not assuming any Data Center compute revenue from China in its outlook." That is the text of the outlook, not a reading of it; the accounts now circulating, which say the forecast assumes China buys zero AI chips this quarter, are for once an accurate paraphrase. On the call the same day the chief financial officer, Colette Kress, put it in her own words: "given ongoing geopolitical uncertainty, there is no China data center compute revenue in our forward outlook."

What matters is how little that assumption costs. In the quarter ended 26 July the company billed $7,880m to customers headquartered in China including Hong Kong, against $96,221m of total revenue — a shade over eight per cent, and nearly double the $3,985m of a year earlier. Read the half rather than the quarter and the direction reverses: $12,430m against $13,644m, a fall of nine per cent in a period when group revenue almost doubled. The geography is assigned by the direct customer's headquarters in any case, so it records where an invoice went, not what kind of chip was sold.

Strip out the parts export rules do not touch and almost nothing remains. Washington began granting licences in February to ship small volumes of H200 to named China-based buyers; Beijing restricted the purchases, and the company says it has "been unable to sell all the products for which we have licenses". Of the fraction actually shipped, the quarterly filing says those "shipments account for less than 1% of Data Center revenue in our most recent quarter". Data-centre revenue was $89.0bn.

The economics of that sliver are worse than its size suggests. Nvidia took a $0.4bn charge in the first half of fiscal 2027 on H200 inventory and purchase obligations, after a $4.5bn charge on H20 a year before. Licensed H200s must be inspected in the United States before shipment, which makes them imports carrying a 25 per cent tariff the company says it has "been unable to pass along any of". That is a rounding error with legal risk attached, not a forecast.

So the zero-China baseline is a floor rather than a gap. KeyBanc's John Vinh has modelled Chinese demand at roughly $30bn a year; against revenue running at $108bn a quarter that is about seven per cent. Excluding it converts China from a forecast input into an option held for nothing. If Washington and Beijing reopen the market, every dollar is upside against a published number; if they do not, the guidance still stands. That asymmetry, rather than any expectation about Thursday's Trump–Xi summit, is why the assumption is worth making.

Nvidia's own account of the cost is more candid than its guidance. At the end of the quarter, the filing says, while it could still ship uncontrolled gaming and workstation parts, "we were effectively foreclosed from competing in China's data center computing/compute market, and our effective foreclosure from the China market helped our competitors build larger developer and customer ecosystems to challenge us worldwide." That is a company telling shareholders the policy has produced a rival, not merely a lost quarter.

This week that competitor acquired a name. At Alibaba's Apsara Conference in Hangzhou, the chief executive, Eddie Wu, unveiled the Zhenwu V900, designed by the group's T-Head silicon unit and described by him as the most powerful AI chip in China today — three times the performance of the previous Zhenwu M890, NBC News reported. Alibaba also said it would train a model of five to ten trillion parameters, against 2.4 trillion for its current Qwen3.8-Max, and expects more than 20 gigawatts of computing capacity by 2032. Every performance figure there is the vendor's, untested by anyone outside. Nor is design the binding constraint; fabrication and high-bandwidth memory are, as Wu conceded: "global shortages across the AI data center supply chain are currently limiting the speed at which we can scale."

Whether the foreclosure spreads will be settled elsewhere. American policy now tries to keep controlled parts out of China by tracking them everywhere else, and everywhere else is South-East Asia, where data-centre capacity is projected to reach 9.4 gigawatts by 2035. Politico reports that the research group C4ADS documented fifty shipments of Nvidia GPUs diverted through Vietnam, Malaysia and India to China and Hong Kong between 2023 and 2025, worth $13.4m — a figure researchers call almost certainly understated. Three bills are queued for the defence authorisation, among them the Chip Security Act, which would put location verification inside advanced chips. The countries being asked to accept that surveillance are the same ones Alibaba, Moonshot and Z.ai are selling cheaper models into.

Set beside the private restriction this desk examined on Thursday morning, the shape is clear enough: a statutory layer, a licensing layer and a layer of classifiers no statute mentions, none of them written on the same page as the others. The guidance prices the whole stack at zero and moves on. The figure that follows — $108.0bn, plus or minus two per cent — is what the largest market in the world is now worth to the company that cannot sell into it.