Goldman Sachs Data Shows AI Model Costs Fell as Fast in Three Years as PC Prices Did in 15
Key Takeaways
- •An a16z analysis of Goldman Sachs data shows AI model prices dropped roughly 95% between March 2023 and September 2026, a decline that took personal computer prices about 15 years during the ICT wave that began in the 1980s.
- •Intense competition among model developers, growing efficiency in proprietary models, and cheaper open-source alternatives are credited with driving the rapid price collapse.
- •The Silicon Data LLM Token Expenditure Index, which tracks realized buyer spending, hit a record low of $0.97 per million tokens in August 2026, down 29% from July and more than 50% below its May 2026 peak of about $2.05.
- •Goldman Sachs warns that if compute supply outruns demand, margins and the sustainability of infrastructure investments could come under pressure, with the risk falling most heavily on Meta, Microsoft, Amazon, and Alphabet.
- •Buyers of AI services have benefited as prices that would have seemed implausibly low in March 2023 are now standard, and the Silicon Data index suggests the price slide has not yet stopped.
AI model prices have fallen about as much in three years as personal computer prices did over 15, according to Goldman Sachs data highlighted in a recent analysis by venture capital firm a16z. The comparison suggests the AI industry is compressing a generation of technology-driven deflation into a single product cycle, a pace of price decline that few corners of the technology sector have matched.
The numbers behind the collapse
The a16z analysis tracks large language model pricing against a baseline index set at 100 in March 2023. By September 2026, the index had dropped to approximately 5, a decline of roughly 95% in three and a half years. Put differently, a service that cost a dollar at the start of the period now costs about a nickel.
The benchmark used for the comparison is the personal-consumption-expenditures (PCE) price index for computers. According to the analysis, computer prices needed roughly 15 years to fall by a similar amount, a decline that played out during the information and communications technology (ICT) investment wave that began in the 1980s. That reference period matters: it spans the stretch in which computing moved from specialized business equipment into mainstream offices and homes.
The analysis credits three forces for the speed of the AI price drop: intense competition among model developers, growing efficiency in proprietary models, and the arrival of cheaper open-source alternatives.
Token prices tell the same story, only louder
A separate gauge points in the same direction. The Silicon Data LLM Token Expenditure Index measures how much buyers spend per million tokens. In August 2026, the index hit a record low of $0.97 per million tokens, a 29% fall from July. The reading also left the index more than 50% below its peak of roughly $2.05, set in May 2026. Because the gauge tracks realized spending rather than quoted rates, it reflects what buyers actually pay in practice.
What cheaper AI means for the companies building it
The overcapacity risk lands most heavily on the biggest spenders. Meta, Microsoft, Amazon, and Alphabet are racing to adjust as the economics of AI shift under their feet.
Goldman's warning boils down to this: if compute supply outruns demand, margins could come under pressure. The sustainability of those infrastructure investments could face real questions unless demand catches up with supply. That supply-versus-demand gap is the variable to watch as capacity plans roll out and monthly pricing readings continue to arrive.
For buyers of AI services, the trend has been almost entirely favorable. Costs that would have looked absurdly low in March 2023 are now standard, and the Silicon Data index suggests the slide has not yet stopped.
