AI Model Prices Plunge to 5% of Original in Three Years, Falling Far Faster Than the PC Era
nashnova research
Large language model prices have plunged 95% in roughly three years — the same drop took PCs about fifteen years, making AI's cost curve historically unprecedented and the key variable for how fast applications reach the real world.
A 95% drop in three years — how extreme is that?
Goldman Sachs data (cited by Andreessen Horowitz) shows the LLM price index fell from a baseline of 100 to roughly 5 since the AI investment cycle began in 2022.
This means → every dollar now buys roughly 20 times the AI capacity it did three years ago.
In plain terms = a task that cost $100 to run through an AI model in 2022 now costs about $5.
How does this compare to the PC era?
During the 1980s ICT investment cycle, the PC price index took about 15 years to achieve the same 95% decline.
AI compressed that journey into three years — roughly five times faster.
This reflects a level of competitive intensity and iteration speed with no precedent in prior technology cycles.
What does the "price-intelligence ratio" signal?
Goldman also tracks a quality-adjusted metric: the price-intelligence ratio — how much AI capability a fixed dollar buys.
That ratio has fallen to near zero in the same three-year span, meaning per-dollar AI capability is scaling almost without limit.
In plain terms = prices aren't just falling — models at the same price point are many times smarter, so cost cuts and capability gains are compounding simultaneously.
How is the price measured?
The LLM price index uses March 2023 as its base period and tracks the average price per million tokens across models.
A token — the smallest unit an AI model processes, roughly one English word or half a Chinese word — is the standard billing unit for AI services.
This means → the collapse is not one vendor's promotion; it is the industry-wide average price in freefall.
What does this mean for investors?
Whether the cost curve keeps falling — and when it bottoms out — is the key variable for gauging AI application penetration speed.
If price cuts continue, application-layer companies gain margin room; if costs stabilize, the payback period on compute investment becomes calculable.
This reflects a shift in the core question: the market should no longer ask "can AI work?" but "how much cheaper can AI get, and for how long?"
市场有风险，内容仅供研究参考，不构成投资建议。