The Digital Press

Governments are setting rules for AI models based on the amount of computing power used in their training, aiming to regulate the most resource-intensive systems. This study forecasts how many AI models will exceed these compute thresholds in coming years.

Why it matters: Identifying models by training compute helps regulate potentially high-risk AI systems effectively.

The big picture: The number of large compute AI models will grow rapidly, surpassing current regulatory thresholds increasingly each year.

Stunning stat: By 2028, an estimated 103-306 models will exceed the EU’s $10^{25}$ FLOP training compute threshold.

Quick takeaway: Compute thresholds tied to the largest training run show stable annual model captures, unlike absolute thresholds that rise superlinearly.