Data, compute, algorithms.
You need all three for training a machine learning model. Recently, the biggest improvements have come from increasing the amount of data and beefing up compute. The assumption is that this trend can continue, while holding the algorithm relatively constant (some kind of transformer architecture). There is an opportunity to look at this third dimension, especially if innovations can be made to keep performance constant while reducing the impact on compute and/or data. And if you can improve performance above and beyond what the current models can do, with the same or less data? Then you are really in business.