USA tighten export rules for AI chips

The US government has announced that it will implement new regulations to control global access to AI chips and technologies developed in the United States. This regulation affects the export of graphics processing units (GPUs), originally designed to accelerate graphic displays but have proven valuable for training and operating AI models. Nvidia's market-leading GPUs are a prime example of this. The aim is to restrict global access to advanced GPUs needed to build clusters for training advanced AI models. The new regulations set capacity limits for computing power, capped at a maximum of 790 million TPP by 2027. This limit equates to nearly 50,000 Nvidia H100 GPUs, as explained by AI expert Divyansh Kaushik from Beacon Global Strategies. Despite these restrictions, there are exceptions for companies like Amazon Web Services and Microsoft Azure, which receive special authorization. Further exceptions apply for the acquisition of smaller quantities of GPUs, which require only a notification to the government instead of a license. This regulation aims to expedite risk-free shipments of American chips worldwide. There are also specific exceptions for GPUs in the gaming sector. Overall, the new measures also affect the so-called "model weights," which play a crucial role in improving the performance of AI models. The US government has adopted security standards to protect the weighting of advanced, non-public models. These measures are intended to ensure that the most advanced AI is developed and deployed in safe and trusted environments.