The U.S. government announced a new regulation on Monday aimed at controlling the access of U.S.-designed artificial intelligence chips and technology by countries worldwide. The regulation specifically targets the export of graphics processing units (GPUs), which are crucial for advanced AI applications.
The rule restricts the export of GPUs, which are specialized processors originally developed for accelerating graphics rendering but have since become essential for training and running AI models. The restriction is based on compute power, with countries limited to a total of 790 million Total Processing Performance (TPP) through 2027. This limitation is equivalent to around 50,000 H100 Nvidia GPUs.
Companies meeting specific requirements, like Amazon Web Services or Microsoft’s Azure cloud unit with “Universal Verified End User” status, are exempt from these caps. National authorizations are also available to companies not located in “countries of concern,” with caps of approximately 320,000 advanced GPUs over the next two years.
Orders for small quantities of GPUs, up to about 1,700 H100 chips, are not subject to caps and only require government notification. This exception aims to facilitate low-risk shipments of U.S. chips globally, particularly for universities, medical institutions, and research organizations. Exceptions also exist for GPUs used for gaming.
Overall, the regulation is intended to regulate the global access to AI technology, encourage companies to secure verified user status, and prevent unauthorized distribution of GPUs, particularly to countries like China.
Destinations exempt from country caps on advanced GPUs include Australia, Belgium, Britain, Canada, Denmark, Finland, France, Germany, Ireland, Italy, Japan, the Netherlands, New Zealand, Norway, South Korea, Spain, Sweden, and Taiwan, in addition to the United States.
The U.S. government is also closely monitoring the handling of “model weights,” a critical component in the training of artificial intelligence (AI) models. These models rely on vast amounts of data to generate meaningful output. Algorithms play a key role in refining these models by evaluating their performance and adjusting numerical parameters known as model weights.
Model weights are crucial as they determine the significance of various operations within the AI model. By tweaking these parameters, the algorithms aim to enhance the model’s ability to complete tasks effectively. To safeguard the integrity of advanced “closed-weight” models, which are not publicly disclosed, the U.S. has established stringent security standards.
According to Kaushik, these regulations are designed to promote the development and deployment of cutting-edge AI technologies in secure and trustworthy environments. This strategic approach aims to ensure that the most advanced AI innovations are utilized responsibly and ethically.
(Reporting by Karen Freifeld; Editing by Chris Sanders and Jamie Freed)
