Compute

Training an AI model takes an enormous amount of calculation, and even a fully trained model needs calculation for every question it answers. The power to do these calculations is called compute. In practice, it means the number of AI chips you have secured, together with the electricity and facilities to run them.

"Not enough compute" means there are too few chips and data centers, so a bigger model cannot be built, or so many people are using the service that it slows down. That is why AI companies sign contracts for compute they will use several years from now. Techniques that get better results from the same amount of calculation have become just as important.

Think of a factory. It is how many machines it has and how many hours a day they can run. No matter how many orders come in, if there are not enough machines, it cannot make more.

Along with data and people, it is counted as one of the three things that decide the AI race. Behind a delayed new model or a new usage limit, a lack of compute is often the cause.