In early November 2026, Reuters and the Financial Times reported on the same day that Nvidia had communicated a new rule to compute service providers in Singapore, Malaysia, Japan, and elsewhere: at least 51% of their service revenue must come from non-Chinese companies, with business from Chinese customers capped at 49%. No written document exists for this directive—Nvidia conveyed it orally through its sales and regional teams, enforcing it via on-site data center inspections and confirmation of contract end users.
The Physical Structure of the Loophole
To understand the strategic significance of the 51% rule, one must first understand the door it seeks to close. The core logic of U.S. chip export controls on China is physical location: chips must not be delivered to Chinese territory. But compute itself can be transmitted over networks—an H100 sitting in a Singapore data center can provide compute services to an AI lab in Beijing through cloud interfaces, while the silicon itself never crosses a border.
This is precisely the path that giants like Tencent Cloud and Alibaba Cloud have commonly used: deploying data centers in Japan, Australia, and elsewhere, acquiring restricted chips in the form of "offshore compliant deployment," and then channeling compute to domestic operations as cloud services. According to multiple media reports, Tencent has procured AI chips through data centers in Japan and Australia to circumvent direct import restrictions.
The 51% rule is designed to intervene at this level: regardless of where the chips are located, if more than half of a provider's revenue flows to Chinese companies, Nvidia will cut off its hardware supply. The rule does not care which fiber optic cable the data packets travel through—only where the money comes from.
The Calculus Behind the Oral Directive
Nvidia chose not to issue a written document.
Written rules imply legal obligations and traceable liability. An oral directive leaves a buffer zone between compliance pressure and business flexibility: Nvidia can demonstrate to the U.S. Commerce Department that it has dutifully enforced export control guidance, while avoiding exposing itself to legal risks arising from specific wording. For service providers, the oral directive similarly creates ambiguous space—no one can produce a contract clause saying "Nvidia breached the agreement."
But enforcement is anything but ambiguous. According to Crypto Briefing, Nvidia's compliance checks have escalated to a three-pronged approach: on-site data center inspections, case-by-case confirmation of contract end users, and direct interviews with the actual users of the hardware. This level of on-site enforcement far exceeds the scope of traditional export control document reviews, more closely resembling the穿透式—look-through—inspection logic of financial regulators.
An earlier move confirms that this tightening is not an isolated event: in July 2026, Nvidia had already cut its approved customer list in Asia by more than 50%. Those hit hardest were the so-called "neoclouds"—providers that specialize in renting out GPU compute and primarily handle AI training and inference workloads, with particularly dense clusters in Singapore, Malaysia, and Japan, where Chinese companies generally account for a high share of the customer base.
Who Bears the Adjustment Costs
From a market structure perspective, the cost distribution of this policy is highly asymmetric.
Nvidia's position is nearly painless. Global demand for Blackwell-series processors is robust; cutting some Asian customers merely reallocates scarce supply without affecting overall shipments. For Nvidia, the main compliance cost is the additional inspection teams, and these expenses can be viewed as the political price of maintaining market access permits in Washington.
The pressure falls on the middle layer: neocloud providers. These companies' business models are highly dependent on Chinese AI customers for a simple reason—Chinese AI labs have massive training compute needs and a strong willingness to pay, making them among the world's best GPU tenants. The 51% rule requires these providers to rapidly bring in sufficient non-Chinese business on top of their existing customer base, and finding replacement customers of matching scale in the short term is no easy task. Some neocloud operators in Singapore are already facing pressure to restructure their customer base; failure to meet the threshold will mean losing eligibility for Nvidia hardware supply.
For Chinese AI companies, the impact is gradual but persistent. In the short term, existing lease agreements are unaffected—Nvidia's new rule only governs newly signed contracts, and existing leasing agreements are temporarily exempt. But exemptions are assets with expiration dates; as contracts come up for renewal, the new rule must be confronted. The "Asia relay" model previously built by Tencent, Alibaba Cloud, and others through overseas data centers will gradually become unviable over the next one to two contract cycles.
Background Timeline of Policy Tightening
This wave of action did not emerge from thin air but is the latest node in a continuously escalating chain of controls.
In 2022, the United States first restricted exports of A100 and H100 chips to China. Subsequently, Chinese companies turned to overseas compute leasing. In May 2026, the U.S. Commerce Department issued new export control guidance explicitly bringing "cloud resale" of compute into the regulatory scope, targeting the Asian bypass channels. In July, Nvidia cut its Asian customer list, and in early November, the 51% rule surfaced.
The timeline shows that the precision of controls is accelerating: from prohibiting chip sales, to restricting where chips can be deployed, to tracking the ultimate destination of compute flows. Each round of escalation closes the technical gaps left by the previous round.
The Forced Logic of Domestic Compute Self-Sufficiency
Chinese AI companies' dependence on overseas compute stems to a considerable extent from the performance gap of domestic alternatives. Although domestic AI chips such as Huawei's Ascend 910B have become practical in inference scenarios, the efficiency gap versus the H100/B200 in large-scale pre-training remains significant, particularly in interconnect bandwidth and cluster scalability.
What the 51% rule accelerates is precisely the pressure to narrow this gap. As channels for obtaining overseas compute systematically narrow, top domestic model labs seeking to maintain training capabilities have only two paths: spend more engineering costs on adapting and optimizing for domestic chips, or accept a contraction in training scale. Both paths mean higher marginal costs.
Over a longer cycle, this pressure has bidirectional effects. It will indeed slow the iteration speed of some Chinese AI labs in the short to medium term. But it is also forcibly accelerating the reprioritization of investment in the domestic semiconductor supply chain, pushing engineering teams to find software-level efficiency compensations under constrained conditions—similar to the "resource-constrained innovation" logic of the Soviet space program.
Independent Judgment
The true significance of the 51% rule lies not in how strong its physical interception capability is, but in the fact that it marks an important upgrade in the enforcement paradigm of U.S. export controls: from controlling physical assets to controlling the direction of service revenue flows.
Once this logic takes hold, the direction of subsequent tightening has a clear extension path—the percentage cap can be pushed further below 49%, the inspection mechanism can extend from new contracts to existing ones, and the scope of controls can expand from Asian neoclouds to data center operators in Europe and the Middle East.
For China's AI industry, what truly warrants vigilance is not this single rule but the regulatory toolkit behind it—one that "traces compute flows and determines eligibility by revenue ratios." Once this toolkit is standardized, any data center anywhere in the world could become a pressure point for compliance. In this sense, the 51% rule is a starting point, not an endpoint.
© 2026 Winzheng.com 赢政天下 | 转载请注明来源并附原文链接