A recent analysis by venture capital firm a16z, utilizing data from Goldman Sachs, highlights a rapid decline in the pricing of large language models. The report indicates that AI model costs have dropped by approximately 95% over a three-year period, a reduction that took personal computers roughly 15 years to achieve during the ICT investment wave that began in the 1980s.
The a16z study tracked pricing against a baseline index of 100 in March 2023, which fell to approximately 5 by September 2026. This trend is attributed to three primary factors: intense competition among developers, increased efficiency in proprietary models, and the emergence of more affordable open-source alternatives.
Additional data from the Silicon Data LLM Token Expenditure Index confirms this downward trajectory. In August 2026, the index reached a record low of $0.97 per million tokens, marking a 29% decrease from July and a drop of more than 50% from the peak of approximately $2.05 recorded in May 2026.
Major industry players, including Meta, Microsoft, Amazon, and Alphabet, are currently adjusting their strategies as these economic shifts unfold. Goldman Sachs warns that if the supply of compute resources continues to outpace demand, infrastructure investments may face sustainability challenges due to potential pressure on profit margins.
