The United States has announced a new AI licensing regime that places restrictions on the exports of Artificial Intelligence chips to certain countries of concern, while vastly expediting the same process for its trusted allies. Secondly, the rules also add greater restrictions on the exports for weights of models trained with 10^26 computational operations or more. Thirdly, the Bureau of Industry and Security (BIS) will also impose security conditions on the storage of advanced AI models.
A statement from the White House clarified that the policy, named Interim Final Rule on Artificial Intelligence Diffusion, was meant to prevent the smuggling of AI chips to adversaries of the United States and maintain American dominance over the industry. It stated that adversaries of the USA were using AI chips for offensive cyber operations, and aiding human rights abuses, such as mass surveillance. “To enhance U.S. national security and economic strength, it is essential that we do not offshore this critical technology and that the world’s AI runs on American rails,” said the statement.
Getting Chips Is Harder For Most But Easier For Some:
The Bureau of Industry and Security (BIS) will implement a global licensing system for the export, reexport, and transfer of advanced computing integrated circuits (ICs) and top-tier AI model weights. This system applies to all destinations and end users worldwide. The licensing decisions consider the destination’s sensitivity, computing power quantity, AI model performance, and recipient security commitments. BIS will generally deny licenses for large quantities of advanced computing ICs intended for AI model training.
The rules place countries under three categories:
- Trusted allies of the USA: Australia, Belgium, Canada, Denmark, Finland, France, Germany, Ireland, Italy, Japan, the Netherlands, New Zealand, Norway, Republic of Korea, Spain, Sweden, Taiwan and the United Kingdom.
- Adversaries of the US or nations under arms embargo: Belarus, Burma, China, Cuba, Iran, North Korea, Syria, Venezuela, Russia, Afghanistan and a number of other countries that come under “Group D:5.”
- Everybody else: This includes India.
However, the policy also offers an exception for the export, reexport, or transfer (in-country) of
eligible advanced computing ICs, related software and technologies to the first group of countries. It also provides an exception for the export of low amounts of compute, up to 26,900,000 Total Processing Performance (TPP) per year to a single consignee, as long as they are not members of adversarial nations.
However, the exporters must provide information about the ultimate consignee of the goods to the BIS, alongside their certification. They must also ensure that the equipment is not accessed by entities within adversarial nations, without the prior approval of the BIS.
Restrictions On Data Centres:
The BIS established two categories of Data Center Validated End-Users (DC VEUs) in October 2024: Universal VEUs (UVEUs) and National VEUs (NVEUs). Companies can apply for UVEU status if they or their ultimate parent have their headquarters in the first group of nations. NVEU status is available to companies worldwide, except those in Macau or Country Group D:5 destinations, for acquiring large quantities of advanced computing integrated circuits.
UVEUs may build data centres in any country in the world, except in Macau or Country Group D:5 destinations. They also face geographic restrictions on AI computing power allocation. Those headquartered in friendly countries cannot transfer more than 25% of their total computing power outside those designated countries, with a 7% limit per individual country. US-headquartered UVEUs cannot transfer more than 50% of their computing power outside the United States.
NVEUs operate under a quarterly cumulative Total Processing Power (TPP) allocation system from 2025 through 2027. The allocations begin at 633,000,000 TPP in Q1 2025 and increase progressively to 5,064,000,000 TPP by 2027.
Restrictions On Model Weights:
The policy also implements new controls on advanced AI model weights effective starting January 13, 2025. For reference, AI model weights are numerical parameters that define the internal structure and decision-making logic of a machine learning model, and allow an AI system to learn patterns from data and make predictions or decisions.
The controls focus on closed-weight models trained with more than 10^26 computational operations, as these represent the frontier of AI capabilities where security risks become significant.
Companies that want to export model weights to other countries can only do so to foreign entities that have received a license from the BIS. However, an exception exists for specified friendly nations. The regulation also excludes open-weight models from control requirements, as they currently operate below the controlled threshold and offer significant research and development benefits. Small-scale additional training operations on open-weight models remain unrestricted.
Reaction From AI Companies:
Reaction to the new export policy from the tech industry was not necessarily positive. Ned Finkle, vice president of government affairs at NVIDIA criticised the move in a blog post and accused the Biden administration of seeking to “undermine America’s leadership.”
He termed the policy to be a “sweeping overreach,” that would stifle competition and establish bureaucratic control over the industry. He also stated that the policy covered technology that was widely available in gaming PCs.
How Would This Affect India?
India is mentioned once in the policy, alongside China in a two member list of countries eligible for General VEU Authorisations. These authorisations allow for the export, reexport, and transfer of eligible items to validated end-users of any in specific eligible destinations.
Pranay Kotasthane, Deputy Director and Chairperson of the High Tech Geopolitics Program at the Takshashila Institution said that the new policies could increase friction in international trade for India. He stated that small orders made by Indian startups could be fulfilled easily, but any attempts at creating big data centres would need authorisation. “From a geopolitical point of view, it places countries into different clubs. India is not in the club with unrestricted access currently,” he said. Kotasthane drew comparisons with the geopolitics of the 1960s and 70s, when countries existed in nuclear and non-nuclear camps.
“While there is no immediate threat to Indian companies, the trend is certainly worrying. The USA could easily choose to decrease the allowed export thresholds. This creates uncertainty,” he said.
Kotasthane also weighed in on the Biden administration’s decision to establish controls on the exports of model weights. “The rationale is that essentially they think that chip controls are not enough to stop the progress of China and Russia. The reason is because once you get the model weights, you don’t need the GPUs,” he said, explaining that model weights are arrived at after developers train an extensive amount of data on certain GPUs. “But if I have access to the model weights, then I don’t even need those many GPUs,” he said. He also pointed out that even current state of the art AI models were all below the threshold above which export restrictions applied. “There’s no restriction on current models. So it is like a future-facing policy action,” he said.
Also Read:
Support our journalism:
For You
This post was originally published on here
