Loss as compute grows Each equal step down
Model loss falls as a power law against training compute. Plotted with both axes logarithmic the relationship is a straight line: each tenfold increase in compute buys the same fixed step down in loss, so equal gains in quality cost ten times more every time.

Both axes are logarithmic, which is exactly what makes a power law look like a straight line. The staircase is the same fact restated in money: every equal step down in loss costs ten times the compute of the one before it.