Free Intensive Nov 21–22Free 2-day hands-on Vibecoding Intensive on Nov 21–22 at 16:00 MSK (Sat–Sun) · Live Cohort 3 starts Nov 23Join Free →Cohort 3 (Nov 23)
ReleaseInsight · On RadarThe Decoder→ original

US Requires Licenses for Nvidia Chips in 5 Countries, While China Trains Models on GPT Answers

Chinese companies rented Nvidia chips through the cloud from data centers in Singapore, Malaysia, Thailand, Indonesia and Brazil, and the law did not prohibit it. Now operators of such centers will need a US government license to buy the chips. But Reuters uncovered a second route: distillation, where a cheaper model learns from GPT's answers the way a student learns from a tutor's solutions. So bans on hardware alone are not enough: access to the models themselves also needs protection.

AI-processed from The Decoder; edited by Hamidun News
US Requires Licenses for Nvidia Chips in 5 Countries, While China Trains Models on GPT Answers
Source: The Decoder. Collage: Hamidun News.
Listen to article
Article Fact Sheet & Key Figures
ParameterDetails
Primary sourceThe Decoder ↗
Editor-in-Chief
Key metric #1Reuters analyzed more than 300 Chinese papers, patents and job postings
Key metric #2Nvidia shares fell as much as 2.9%, and the semiconductor index as much as 5.5%
Key metric #3US companies' subsidiaries and recommendation systems are exempt

Chinese companies did not import the banned Nvidia chips. They rented them remotely, through the cloud, from data centers in Singapore, Malaysia, Thailand, Indonesia and Brazil. Formally, this was legal, and that is the loophole the US is now closing: operators of such centers will have to obtain a government license to buy the chips. But there is a second workaround, which Reuters reported. Chinese developers are training their compact models on the answers of American ones, like a student preparing from the solutions of a strong tutor. The technique is called distillation. Hardware bans cannot stop it, and for everyone building products on APIs, it is already a practical question.

What the new Commerce Department rule changes

The Bureau of Industry and Security (BIS) of the US Department of Commerce is preparing an interim rule. It equates remote access to computing power with the physical export of chips. According to Bloomberg, operators of data centers in five countries — Singapore, Malaysia, Thailand, Indonesia and Brazil — will need a license. This applies to purchases of modern AI chips from American manufacturers such as Nvidia. Applications will be reviewed one by one.

There are exceptions too. No license is required for subsidiaries of companies headquartered in the US. Data centers that serve only commercial recommendation systems for online stores and social networks don't need one either. The rule takes effect after publication in the Federal Register.

The market reacted immediately: Nvidia shares fell by as much as 2.9%, and the Philadelphia Semiconductor Index, which tracks US chipmakers, fell by as much as 5.5%. The political trigger is suspicion of smuggling physical chips. The rule came a few days after the sudden resignation of Nvidia CEO Jensen Huang. US authorities are investigating Singapore's Megaspeed International: banned Nvidia chips were allegedly found in a Chinese data center linked to its former employee. Megaspeed denies the violations. The scandal sparked a dispute within Washington: Treasury Secretary Scott Bessent publicly accused Commerce Secretary Howard Lutnick of inaction.

How distillation works and why it is hard to stop

Reuters examined more than 300 Chinese scientific papers, patents, job postings and government documents. The conclusion: distillation, which began as an academic technique, has become a separate industry with a division of labor.

Imagine that, instead of years of independent training on expensive hardware, a student takes ready-made solutions from a strong teacher. One costly query to an American "teacher" model replaces a large amount of the student's own computation. Chinese companies write programs that automatically collect answers, code and problem solutions from leading American models. Specialists then process this data and train cheaper local "student" models on it.

The publications themselves provide examples. In November 2024, researchers at Shanghai Jiao Tong University assembled 10,750 training examples for Alibaba's Qwen2.5-72B. They used about 5,000 answers from OpenAI's o1 and translations produced with GPT-4o mini, and they themselves warned about the risk of dependence on others' developments. In April 2025, Alibaba described systematic use of GPT-4 and GPT-4o to generate data in foreign languages. In February 2026, Jiao Tong and the food delivery service Meituan used GPT-4o to train small models that coordinate coding agents.

The state nudges this indirectly. The "AI+Manufacturing" package from eight ministries, published in January, and initiatives in Shanghai and Shenzhen support distillation and other ways to reduce computing consumption. There is no direct order to copy American models. But the desire to deploy cheap compact systems in industry in practice pushes developers to widely use the strongest American models as a data source.

Limitations and what practitioners should do

Distillation does not eliminate the need for powerful chips: models still have to be trained and run on hardware. It only greatly reduces the load. Chinese specialists also point to their own progress in algorithms and model architecture, so explaining the rapid growth by copying alone is too simple.

The legal picture is also ambiguous. Distillation is not banned by law, but it violates the terms of use of OpenAI and Anthropic. Both companies have published reports on the volumes of extracted data. Anthropic said it now blocks most such attacks. Bessent threatens sanctions and addition to the Entity List if covert distillation becomes intellectual property theft. Chinese developers, unsure where the line is, asked China's Ministry of Industry and Information Technology to issue clear rules on using foreign models.

What follows from this. Teams that sell API access to their models should treat protection against mass extraction of responses as part of security: rate limits, monitoring of anomalous queries, and clear terms of use. Companies with infrastructure in Southeast Asia and Brazil need to watch for the rule's publication and plan hardware procurement licensing in advance. Control over chips and control over access to models will now have to be discussed together.

Story

Key Questions Answered

What the new Commerce Department rule changes?

The Bureau of Industry and Security (BIS) of the US Department of Commerce is preparing an interim rule. It equates remote access to computing power with the physical export of chips. According to Bloomberg, operators of data centers in five countries — Singapore, Malaysia, Thailand, Indonesia and Brazil — will need a license. This applies to purchases of modern AI chips from American manufacturers such as Nvidia. Applications will be reviewed one by one. There are exceptions too.

How distillation works and why it is hard to stop?

Reuters examined more than 300 Chinese scientific papers, patents, job postings and government documents. The conclusion: distillation, which began as an academic technique, has become a separate industry with a division of labor. Imagine that, instead of years of independent training on expensive hardware, a student takes ready-made solutions from a strong teacher. One costly query to an American "teacher" model replaces a large amount of the student's own computation.

What are the key details on Limitations and what practitioners should do?

Distillation does not eliminate the need for powerful chips: models still have to be trained and run on hardware. It only greatly reduces the load. Chinese specialists also point to their own progress in algorithms and model architecture, so explaining the rapid growth by copying alone is too simple. The legal picture is also ambiguous. Distillation is not banned by law, but it violates the terms of use of OpenAI and Anthropic. Both companies have published reports on the volumes of extracted data.

ZK
Zhemal Khamidun — CPO at Alpina Digital, product manager of AlpinaGPT. Building AI products since 2019. Hamidun News is a pet project that grew into a production aggregator with 6,000+ readers.
✦ Private AI Builders Community · Vibe Club

Ship AI Faster with Founders, Engineers & Weekly Architecture Reviews

Inside Vibe Club (hamidun.com/club) — weekly live masterminds with Zhemal Khamidun, Tech Radar, full LMS access, and 370+ production Claude Code skills.

Join Vibe Club →Explore Hamidun Academy
What do you think?
Loading comments…