Epoch AI: AI Costs Are Falling Faster Than Any Transformative Technology in History
nashnova research
Epoch AI research shows AI costs drop 47% per quarter — roughly 12.7× per year — outpacing every transformative technology on record, from electricity to batteries to gene sequencing. The real question: can application growth keep up with collapsing costs?
How fast, and compared to what?
AI costs have fallen 47% per quarter since 2021, compounding to roughly 12.7× per year.
The study benchmarks this against four historical cost curves: electricity (1892–1973), lithium-ion batteries (1991–2024), computing power (1940–~2001), and DNA sequencing (2001–2022).
On a log scale stretching from 10× to one trillion ×, AI's slope is the steepest of all.
In plain terms = from the same starting line, AI's cost decline outpaces every major technology price collapse of the past century-plus.
Is the number credible? Any cross-check?
Appenzeller's independent 2024 estimate offers corroboration: between 2021 and 2024, per-token prices for models at equivalent capability levels fell roughly 10× per year.
This means → two research teams, using different methods and metrics, arrive at the same order of magnitude — the cost collapse is not a single-source claim.
In plain terms = whether you measure "cost to run the same task" or "price per token," the answer points one way: one order of magnitude cheaper every year.
If costs fall this fast, what is the real question?
Epoch AI flags the core tension: as usage costs compress at this pace, can application growth expand fast enough to match?
This means → if demand expansion trails cost decline, unit economics get squeezed — infrastructure sellers earn less than today's valuations assume.
The second question cuts deeper: do current high valuations on compute infrastructure already price in this downward cost curve?
In plain terms = the market prices compute companies on an assumption of "limitless demand, stable pricing." If AI costs drop an order of magnitude each year, that assumption needs re-examining.
市场有风险，内容仅供研究参考，不构成投资建议。