That's the gap Washington just moved on. Chinese companies didn't need to own Nvidia chips. They could rent them from a data center in another country. The new rule still doesn't regulate renting as such. It regulates who can be sent chips.

Who writes and enforces the rules?

The Commerce Department's Bureau of Industry and Security (BIS) writes and enforces them, under the Export Administration Regulations. The legal authority for the latest rule is the Export Control Reform Act of 2018. Under that law, the rule skips the usual notice-and-comment process. It takes effect when it's published in the Federal Register, and comments are due within 60 days after that.

There's one unusual detail this time. BIS says the rule was written "by and at the direction of" the Treasury Department and the National Security Council. You'll also see the phrase "super intelligence," or SI, in the rule and in BIS's announcement.

How does the government decide which chips are "advanced"?

It uses a score called total processing performance, or TPP. TPP is a chip's peak multiply-accumulate rate multiplied by the bit length of the numbers it's working with. The rule's own technical note gives an example: a chip doing 300 teraFLOPS at 16 bits reaches the 4,800 TPP threshold.

For a sense of scale, SemiAnalysis works out Nvidia's H100 at about 15,831 TPP, roughly three times that line. Chips above the thresholds fall under the controlled categories, and exporters list each chip's category, TPP and quantity in the reports the new rule requires.

Can these chips be sold to China at all?

Some can, case by case. From 2022, high-end AI chips going to China faced a blanket presumption of denial, meaning applications were expected to be refused.

A BIS rule effective 15 January 2026 moved some chips to case-by-case review. It covers only chips under 21,000 TPP and 6,500 GB/s of memory bandwidth, roughly the level of Nvidia's H200 or AMD's MI325X. It also comes with conditions, such as China and Macau shipments not exceeding 50% of US shipments.

The day before, the White House put a 25% tariff on chips that meet the same thresholds. The president tied it to a policy of allowing H200 sales to China in exchange for 25% of the revenue.

So the policy runs in two directions at once. Washington has loosened direct sales of mid-tier chips while tightening indirect access to more powerful ones.

Why was renting chips overseas a gap?

The rules define "export" in a way that predates cloud computing. BIS advisory opinions from 2009 and 2011 held that a cloud provider isn't the exporter when users create and download technology on its servers. A 2014 opinion said using software in the cloud without downloading it isn't an export.

As a result, the 2022 controls kept high-end GPUs out of China, but they generally didn't stop countries of concern from remotely using GPUs in the US or elsewhere. Sen. Dave McCormick has said BIS "has no authority to require a license" when bad actors train models on chips under US jurisdiction.

Chinese companies did use that route. RAND reported that ByteDance planned at the end of 2024 to spend $7 billion on access to Nvidia chips on servers outside China. It also reported that Chinese firms had carried hard drives of data abroad to train models on foreign servers.

Was it really a loophole?

BIS says not entirely. The rule says the October 2023 end-use control "is not, and has never been limited" to shipments made directly to China-linked companies. In BIS's reading, it already covered data centers where such a company bought the chips or used more than 50% of the compute.

What's new is a much lower trigger. Exporters may not knowingly send advanced chips to a data center anywhere that will give companies headquartered in China, Macau or other Country Group D:5 countries (or owned by parents there) access to more than 12,000,000 "aggregate TPP". That applies no matter what share of the data center's capacity those customers use. The limit counts a single data center or a set of "linked" ones, including data centers that are only digitally linked.

How much computing power is 12 million TPP?

It's roughly 760 H100-class chips. That's our arithmetic, using SemiAnalysis's figure of about 15,831 TPP per H100. It isn't an official BIS figure.

Newer chips score higher, so the same threshold allows fewer of them. Nobody has published how much China-linked computing power is rented overseas, so it isn't known how many data centers sit above the line.

What else does the new rule change?

  • Existing data centers have 90 days. Those already supplying significant compute to China-linked companies must get a license within 90 days of the rule taking effect or "face future export restrictions." If a data center switches a chip it already has to a prohibited use, that counts as an in-country transfer and needs authorization.
  • Five more countries need licenses. Advanced chips going to Indonesia, Malaysia, Singapore, Thailand and Brazil now need a license because those countries carry a "higher-than-normal risk of diversion." Companies headquartered in the US are exempt if they file monthly reports listing each chip's category, model, TPP and quantity.
  • Some uses are exempt. "Commercial recommender systems," like those used by online stores and social media, aren't covered.
  • An old rule is formally cancelled. The rule rescinds the Biden-era AI Diffusion Rule, which was never enforced.

Does this ban Chinese companies from renting chips remotely?

No. Legally, the rule still regulates exports, reexports and in-country transfers of physical chips to data centers. It is triggered by what the exporter knows about how the chips will be used.

Direct power over remote access is what the Remote Access Security Act would give BIS. The bill would amend the 2018 export law to define and license "remote access." It passed the House 369-22 in January 2026. As of that report, the Senate version from McCormick and Sen. Ron Wyden was still in the Senate Banking Committee. We found nothing newer on it.

Chris McGuire of the Council on Foreign Relations called the rule "a major step forward". He also said loopholes remain, including some that cover activity inside the US.

How will it be enforced?

Mostly through licenses and through pressure on governments. Commerce says it will watch for countries that tolerate evasion. It may respond with import restrictions, new license requirements, a downgrade in a country's export-control status, Entity List designations, or temporary export bans on the people involved.

BIS expects the extra paperwork to be modest. It estimates about 100 more license applications a year, plus around 970 extra filings, which include the monthly reports.

How well this works abroad is an open question. Enforcement depends on exporters knowing who their chips will serve, and on cooperation from places like Singapore. Singapore police opened an investigation in October 2025 into Megaspeed, a company at the center of a US probe into restricted Nvidia chips reaching China.

When does it take effect?

It isn't known yet. The rule takes effect when it's published in the Federal Register, and no publication date had been announced. The 90-day deadline for existing data centers and the 60-day comment window both count from that date.