Training and inference compute costs have followed a multiplicative growth pattern: GPT-3 cost $100 million with 175 billion parameters, while GPT-4 cost $400 million (4x) but with nearly 10x the parameters, suggesting costs continue to grow with each subsequent model generation.