Stanford Research Tracks Rapid Escalation of AI Model Training Costs

Stanford HAI's 2025 AI Index research chapter identified three scaling rates for AI model training. The research estimates that training compute doubles about every five months, datasets double about every eight months, and power required for frontier training doubles about once a year. The report notes that estimated costs for later frontier models are in the tens or hundreds of millions of dollars, compared to Google's original 2017 Transformer, which cost about $930 to train.

The reported scaling rates are considerably faster than the cadence associated with Moore’s Law. The research establishes the rapidly escalating operational costs of frontier AI models, which directly impacts Google's business.