According to LessWrong, a new analysis suggests that token generation speed—how fast AI models produce text—is constrained by memory bandwidth: the speed at which a GPU can read model weights. This creates a fundamental ceiling on model size for any given hardware. Working through these constraints, the analysis projects model sizes could grow from ten trillion parameters in twenty twenty-six to one point four quadrillion by twenty thirty-one. But there's a critical shift: starting in twenty twenty-seven, models will have to grow larger than compute alone would justify, because training becomes bottlenecked not by processing power but by available training data. By twenty thirty-one, models might need to be four times bigger than unlimited data would predict—suggesting that before AI hits the limits of silicon, it'll hit the limits of having enough text to train on.

Listen to this story

Hear this and more stories in a personalized audio briefing.

Open The Chonkerton