Reponsible Diffusion of AI Technology

Read the below explainer to understand how the new US rules around AI hardware effect you.

The AI Diffusion Framework, introduced by the U.S. Department of Commerce’s Bureau of Industry and Security in January 2025, is a set of export control regulations aimed at managing the global spread of advanced AI technologies, including AI servers powered by cutting-edge chips like GPUs

It seeks to balance U.S. national security interests with technological innovation by controlling who can access these powerful systems and under what conditions.

Here’s how it impacts customers looking to buy AI servers as of February 28, 2025:

The framework divides the world into three tiers. Tier 1 includes the U.S. and 18 close allies (e.g., Japan, South Korea, NATO partners), who face minimal restrictions and can purchase AI servers with advanced chips relatively freely. Tier 2 covers most other countries, where customers can buy AI servers but are subject to strict caps on computing power-measured in Total Processing Power (TPP)-such as a limit of about 1,699 NVIDIA H100-equivalent GPUs per customer annually without a license. Tier 3 includes adversarial nations like China and Russia, where access to advanced AI servers is effectively blocked.

For customers in Tier 2 countries-think businesses, research institutions, or governments in places like Singapore or much of the Middle East-this means navigating a licensing process for anything beyond modest AI server purchases. They can buy smaller setups (up to the TPP cap) without much hassle, but scaling up requires approval from U.S. regulators, which could delay projects or increase costs. Companies relying on U.S.-based cloud providers (e.g., Microsoft, AWS) might still access AI compute remotely, but on-premises server deployments face supply chain hurdles and compliance burdens.

For now, customers in Tier 1 enjoy smoother access, while those in Tier 2 must plan around caps or pivot to cloud solutions-assuming they can meet U.S. security standards. With the Trump administration taking office, though, these rules could shift after the public comment period ends in May 2025, adding uncertainty for buyers worldwide.

The AI Diffusion Framework doesn’t provide an exhaustive, publicly detailed list of every country in each tier. Instead, the tier assignments are inferred from existing U.S. export control policies, strategic alliances, and statements from the rulemaking process.

Based on available information, here’s a breakdown of how countries are likely categorised:

These are countries with strong U.S. security alignments, often part of formal alliances or trusted trade partnerships.

They face the least restrictions on purchasing AI servers with advanced chips.

Australia
Belgium
Canada
Denmark
Finland
France
Germany
Ireland
Italy
Japan
Netherlands
New Zealand
Norway
Republic of Korea
Spain
Sweden
Taiwan
United Kingdom
United States

This is the broadest tier, covering nations that aren’t top-tier allies or adversaries.

Customers here can buy AI servers up to a Total Processing Power (TPP) cap (e.g., equivalent to ~1,699 NVIDIA H100 GPUs annually) without a license, but scaling beyond that requires U.S. approval.

Examples:

Asia: India, Malaysia, Thailand, Philippines, Indonesia.
Middle East: UAE, Saudi Arabia, Qatar, Egypt.
Latin America: Brazil, Mexico, Chile, Argentina.
Europe (non-NATO): Switzerland, Austria, Sweden (if not yet fully NATO-aligned), Ireland.
Africa: South Africa, Kenya, Nigeria.

These are countries where advanced AI server exports are effectively banned or tightly controlled due to national security risks or sanctions.

Afghanistan
Belarus
Burma
Cambodia
Central African Republic
China (including Hong Kong)
Congo (Democratic Republic of)
Cuba
Eritrea
Haiti
Iran
Iraq
Lebanon
Libya
Macau
Nicaragua
North Korea
Russia
Somalia
South Sudan
Sudan
Syria
Venezuela
Zimbabwe

Notes and Caveats

Dynamic Assignments: The tiers aren’t static. A country’s status could shift based on diplomatic relations or compliance with U.S. security goals (e.g., Sweden might move to Tier 1 post-NATO integration).
Data Gaps: The exact list isn’t fully public as of now-tiering often builds on existing Commerce Control List (CCL) designations like Country Groups A, B, D, etc., under the Export Administration Regulations (EAR). Tier 1 aligns with Group A:1 (e.g., Wassenaar Arrangement members), Tier 3 with Group D:5 (embargoed states), and Tier 2 with the rest.
Cloud Loophole: Tier 2 and even some Tier 3 customers might still access U.S.-based AI compute via cloud services, though this is under scrutiny and could tighten.

For a precise list, customers would need to consult the U.S. Department of Commerce directly or monitor updates post-May 2025, when the framework’s comment period ends and potential revisions under the Trump administration might clarify borders.

To better understand the The AI Diffusion Framework, how it effects you, and how Broadberry can meet your GPU solution needs, please contact us, or find out more here .