U.S. Closes Loophole on Cloud-Based AI Chip Access for China
nashnova research
The U.S. Commerce Department issued an interim final rule bringing cloud-based compute access under China export controls — four GB200 racks are enough to trigger a license requirement. If the chips can't be sold, the compute can't be rented either.
What does this rule actually control?
The core logic in one line: if exporting the chip is banned, "exporting the compute" through the cloud is banned too.
This means → the most common workaround — Chinese firms renting Nvidia compute remotely via overseas data centers — is now formally shut down.
In plain terms = the old rule policed chip sales; the new rule polices compute rentals, regardless of which country the server rack sits in.
How low is the threshold?
The rule creates a new end-use control: exporting advanced chips to any data center requires a license if there is "knowledge" the center will provide more than 12 million TPP (Total Processing Performance) to entities headquartered in China, Macau, or the D:5 country group.
Roughly converted using Nvidia's published FP8 specs, that is ~760 H100s, ~330 B200s, or just over four GB200 NVL72 racks.
This means → the threshold sits far below frontier-model training scale. Selling even a small cluster of Blackwell racks to a Chinese client could trigger a license.
The rule also counts "associated data centers" — infrastructure linked digitally — as a single unit. Splitting compute across multiple buildings no longer works.
Which countries are singled out?
Indonesia, Malaysia, Singapore, Thailand, and Brazil face separate destination-license requirements.
BIS has determined these five countries carry "above-normal diversion risk" — this reflects Washington's judgment that Southeast Asian data centers have become the primary transit corridor for compute.
Even if the cloud customer is not itself a Chinese company, the rule applies whenever the customer is substantively training models for or serving models to a restricted entity.
What is the link to Jensen Huang's departure?
According to Bloomberg, the rule's release is directly tied to heightened scrutiny of semiconductor exports following Nvidia's former CEO Jensen Huang's resignation four days earlier.
The trigger was Huang's relationship with Alice Huang, whose Singapore-based firm Megaspeed had long been suspected of routing Nvidia compute to Chinese clients.
In plain terms = Megaspeed imported at least $4.6 billion in Nvidia hardware since its 2023 founding — exactly the kind of operation this rule targets.
What happened to the Biden-era AI Diffusion Rule?
The new rule simultaneously repeals the Biden administration's AI Diffusion Rule, which BIS says was "never enforced."
Everything in it is gone: the country-tier system, model-weight controls, the data-center VEU program — all scrapped.
This means → the new framework is not a patch on the old one. It is a tear-down-and-rebuild, replacing a complex tiering system with a simpler license regime.
What comes next?
Data-center operators currently providing significant compute to restricted entities must obtain a license within 90 days of the rule taking effect.
Whether that 90-day window is enough for compliance overhauls is the first real test of whether the rule can be enforced.
This reflects a broader pattern: the rule's design intent is clear, but the execution-side uncertainties are only just beginning.
市场有风险,内容仅供研究参考,不构成投资建议。
