The Price of Compute Collapsed. The Price of Labor Didn’t.
This chart compares the running frontier price of observed FP32 computing hardware with the BLS Employment Cost Index for civilian-worker compensation. Both are indexed to 2010 on a logarithmic scale, but the units are not task-equivalent and the gap does not measure whether machines can substitute for workers.
The compute line is a sparse observed hardware frontier carried forward between releases. FP32 compute and labor compensation are not task-equivalent units.
What does it show?
The price of the best observed unit of FP32 compute collapsed while the cost of compensating labor continued to rise. The two units are not task-equivalent, so the gap does not measure substitutability.
Methodology
For each release year, take the highest positive Price-performance value in Epoch AI's Machine Learning Hardware dataset, carry that frontier forward, and invert it to represent the price of one unit of FP32 compute. Compare it with the annual average of FRED ECIALLCIV, the BLS Employment Cost Index for total compensation of civilian workers. Both are indexed to 2010 = 100 and intended for a logarithmic axis. Release prices can be sparse and do not reflect bulk discounts. FP32 is a consistent hardware benchmark but is not a complete measure of modern AI workloads. Compute and labor are not task-equivalent units, and this descriptive comparison does not imply that one can substitute for the other.