The US Commerce Department has issued a rule that extends export controls on advanced AI chips from the chips themselves to the computing they provide. Data centres anywhere in the world above a set size that may supply computing power to Chinese companies will need a US licence for American AI chips, and five fast-growing data-centre markets, Malaysia, Singapore, Indonesia, Thailand and Brazil, face a new licence requirement. The rule also formally scraps the Biden-era AI Diffusion Rule. It reshapes where the world’s AI infrastructure can be built and who may use it.
For three years, Washington has tried to keep the most powerful AI chips out of China. One route around those controls has been obvious: rather than buy the chips, rent the computing. A Chinese company could train models on servers in a third country without a single chip crossing its border. On Monday, the US Bureau of Industry and Security (BIS) moved to close that route.
What the rule does
The BIS described the measure as an interim final rule intended to ensure that “existing restrictions on access to sensitive American technology cannot be circumvented through remote computing arrangements involving data centers in third countries.” According to the bureau’s release and rule text, and reporting by TechCrunch and Fortune, its main provisions are:
- A global licence requirement. Exporting, re-exporting or transferring advanced AI chips and servers to a data centre anywhere now requires a licence if that data centre may supply computing power to companies headquartered in China, Macau or other countries in the US arms-embargo group known as D:5, including through brokers.
- A size threshold. The rule applies to data centres with more than 12 million units of “total processing performance” (TPP), aggregated across a site or linked sites. The Berkeley Existential Risk Initiative estimates this as roughly 750 Nvidia H100-class chips, or about 220 of the newest B300s.
- Five named countries. Advanced chips going to Malaysia, Singapore, Indonesia, Thailand and Brazil now need a licence, because Commerce says these markets “present a higher-than-normal risk of diversion”. US-headquartered customers are exempt but must file monthly shipment reports.
- A 90-day window. Existing operators already serving D:5 companies must obtain licences within 90 days of the rule taking effect.
- An exclusion for commercial recommendation systems.
- Rescission of the AI Diffusion Rule. The January 2025 rule, which sorted countries into tiers and was never enforced, is formally removed, along with its controls on AI model weights.
The rule takes effect on publication in the Federal Register, with 60 days for comment; NSH did not find a publication date. The text says it was written “by and at the direction of” the Treasury and the National Security Council, and uses the term “super intelligence” chips, an unusual label for an export regulation.
What “TPP” measures
Total processing performance is the yardstick US export rules use for AI chips. It combines a chip’s raw arithmetic throughput with the precision of the numbers it handles, so that a chip doing many low-precision operations scores comparably to one doing fewer high-precision ones. Using an aggregate TPP threshold for a whole site, rather than counting chips, means the rule captures a data centre’s total capability regardless of which generation of hardware it uses. As chips become more powerful, fewer of them will be needed to cross the line.
Why remote access matters
Training a frontier AI model requires thousands of specialised chips working together for weeks. Controls on chip shipments assume that possession equals capability. Cloud computing breaks that link: capability can be rented by the hour from anywhere with a network connection. The rule’s own rationale is that such access “can be used to train and run advanced SI models capable of supporting military- and intelligence-related activities.”
The five countries named are among the fastest-growing data-centre hubs in Asia and Latin America. Johor in Malaysia, across the strait from Singapore, has become one of the world’s busiest data-centre construction sites, partly because of demand from Chinese technology firms.
The context
The rule arrives in an unusual week for the industry. Jensen Huang resigned as Nvidia’s chief executive last week; the company’s board said it had “no reason to believe that either national security or NVIDIA secrets were compromised”. Fortune and CBS have linked the timing of the rule to that episode and to allegations, which the company concerned denies, that a Singapore firm channelled Nvidia chips to China. NSH has not verified those allegations. Nvidia was due to report quarterly results on Tuesday, 17 November, US time, its first under interim chief executive Colette Kress.
Analysts expect more to follow. Hanna Dohmen of Georgetown’s Center for Security and Emerging Technology told Fortune that “the regulatory path in Washington is going to be harder and bumpier”, with several bills on chip security pending in Congress. Joshua Kurlantzick of the Council on Foreign Relations warned that “for Malaysia, the controls could be disastrous.”
What it means for India
India is not one of the five named countries, and nothing in the rule singles it out. But it matters in three ways. First, India is courting large AI data-centre investments, including in Andhra Pradesh and Telangana; a market seen as low diversion risk could become more attractive to operators seeking licence-free certainty. Second, according to the Berkeley Existential Risk Initiative’s reading of the rule, Indian-headquartered companies running capacity in the five named countries do not get the exemption given to US firms. Third, the rule reinforces that access to advanced compute is now a matter of geopolitics, which strengthens the case for India’s own semiconductor and GPU ambitions. These implications are NSH’s analysis; no official Indian response was found.
Limitations
This is an interim rule, open to comment and amendment. How strictly licences are granted, and how the US will verify who uses a data centre’s computing, is not yet clear. The TPP-to-chip conversions are estimates by a third party.
What to watch
- Publication in the Federal Register and the start of the 90-day window.
- Responses from Malaysia, Singapore and Nvidia.
- Whether Congress passes the pending Remote Access Security Act and Chip Security Act.
– Kalyan M



