Token Prices Drop 40% Yet Enterprise AI Spending Rises 50%: Jevons Paradox Emerges
Nashnova编辑部
Usage-weighted token prices have fallen roughly 40% since late June, yet top-tier enterprises lifted per-capita AI spending by nearly 50% month-on-month in July — cheaper intelligence is not shrinking demand but fueling it, a textbook Jevons Paradox.
Prices fell 40% — so why are companies spending more?
Frank Flight, head of macro strategy at Citadel Securities, notes that usage-weighted token prices have dropped about 40% since late June.
Yet Ramp data shows the top 1% of enterprises by AI spend increased per-capita AI expenditure by nearly 50% month-on-month in July.
This means → lower unit prices did not cut budgets. They made AI feel like a bargain — so companies bought more, and total spending rose.
What is the Jevons Paradox, and why does it matter here?
The Jevons Paradox — an economic phenomenon where making a resource cheaper causes people to use so much more of it that total consumption rises — was first observed with 19th-century coal.
In plain terms = steam engines got more efficient → coal cost less per unit of work → factories scaled up → total coal use grew, not shrank. AI inference costs now play the role coal prices played then.
Flight argues this dataset shows cheaper "intelligence" is triggering a usage explosion, not compressing the AI ecosystem's overall economic value.
GPU rental prices are rebounding — what signal does that send?
In mid-June, GPU rental prices and token prices fell in tandem, raising fears that end-user demand was weakening across the board.
But Ornn data shows GPU rental prices have rebounded from their June lows, even as token prices keep falling — the two have diverged.
This means → cheaper inference is generating enough incremental usage to support demand for the underlying compute. Flight sees this divergence as early evidence that the Jevons dynamic is becoming dominant.
What does cloud-revenue data say?
Hyperscaler cloud revenue for Q2 shows newly added compute capacity is finding paying demand quickly.
This reflects a key signal: the capacity being built is not sitting idle — someone is buying it.
In plain terms = supply is expanding, and demand is keeping pace. At least for now, there is no "built it but nobody came" problem.
Has the risk of a compute glut disappeared?
Flight is explicit: concerns about overbuilding compute have not gone away. Today's high utilization rates offer limited guidance for supply-demand balance years out.
The critical uncertainty: if future model architectures achieve a step-change in efficiency, the compute needed per unit of intelligence could fall sharply — forcing a reassessment of returns on current capex.
He also cautions that GPU rental and token-pricing markets are still early-stage — more fragmented and less transparent than financial markets — and their signals should be read carefully.
What does this mean for investors?
Flight's core conclusion: whether elastic demand is strong enough will determine if the "grand expectations" embedded in current AI valuations can be met.
This means → if the Jevons Paradox holds — prices fall, usage surges, total revenue grows — then today's high valuations have a revenue foundation.
But if lower prices simply mean lower prices, and usage growth cannot keep up, those valuations lack a path to delivery. That is the single most important judgment call in AI investing right now.
Content is for reference only, not financial advice.