Font size:
Print
New US Regulations on AI Chips and Technology
Context:
The US government announced a new regulation aimed at controlling the flow of advanced Artificial Intelligence (AI) chips and technology to other countries.
Overview of the New Regulation:
- Objective: The US government aims to regulate the flow of advanced AI chips and technology to prevent misuse, particularly in “countries of concern.”
- Scope: Focuses on restricting Graphics Processing Units (GPUs), which are crucial for training and running advanced AI models.
- Implementation: New export controls focus on the total processing performance (TPP) of chips. Caps on TPP are set for certain countries through 2027.
Why GPUs Are Targeted?
- Purpose: GPUs, originally developed for graphics rendering, have become essential for AI due to their ability to process massive amounts of data simultaneously.
- Applications: AI models like OpenAI’s ChatGPT rely on tens of thousands of GPUs for training and improvement.
- Demand Factors: GPU sophistication, Volume of training data, Model size, and Time allocated for training.
Details of the New Restrictions:
- Caps on Compute Power: Countries under restrictions are limited to 790 million TPP through 2027. Equivalent to approximately 50,000 Nvidia H100 GPUs, a highly advanced GPU used in cutting-edge AI research.
- Exemptions: Companies with Universal Verified End User status (e.g., Amazon Web Services, Microsoft Azure) are not subject to country caps. National authorisations provide additional exemptions, allowing up to 320,000 advanced GPUs over the next two years.
- Objective of Caps: Encourage companies to secure verified user status, providing visibility to US authorities. Prevent GPUs from being smuggled into restricted countries like China.
Countries Exempt from Caps:
- Unlimited Access: 18 destinations are exempt from GPU restrictions, including:
- Australia, Belgium, Britain, Canada, Denmark, Finland, France, Germany, Ireland, Italy, Japan, Netherlands, New Zealand, Norway, South Korea, Spain, Sweden, Taiwan, and the United States.
Regulations on Model Weights:
- What Are Model Weights? Numerical parameters in AI models are adjusted during training to improve performance on tasks. Essential for AI models to produce meaningful outputs.
- Security Standards: Advanced “closed-weight” or non-public AI models are subject to security measures to protect model weights from being exploited.
Purpose of the Regulation:
- National Security: Prevent sensitive AI technologies from reaching adversarial nations.
- Accountability: Encourage transparency among companies using advanced GPUs.
- Global AI Leadership: Maintain the US as a leader in AI technology development while controlling potential risks of misuse.
This regulation reflects the US government’s proactive approach to managing advanced AI technology while ensuring compliance, security, and global leadership.
