Does making intelligence cheaper expand demand for it? That is the question Frank Flight puts at the centre of his latest Global Macro Strategy note, Elastic Expectations1 (Citadel Securities, August 18, 2026), and it is the right one. Flight opens with Dickens and Jensen Huang in the same breath: the AI economy, like Pip, is valued on anticipated fortune, on the durability of rents to be extracted from models, compute and the application layer. Whether those expectations are reasonable, he argues, increasingly comes down to a single economic variable: elasticity. His tentative verdict, based on the evidence of recent weeks, is yes.
The Numbers Behind the Thesis
The core empirical claim is striking. Usage-weighted token prices, as measured by Silicon Data, have fallen roughly 40% since the end of June. Yet July saw AI spend per employee rise approximately 49% month over month among the top 1% of enterprises in Ramp's transaction data, 25% among the top 10%, and 9% at the median. Falling prices are expanding the quantity of intelligence consumed rather than shrinking the ecosystem's aggregate value. Upstream, GPU rental markets have held firm relative to June lows, and Q2 hyperscaler cloud revenues reached $106.3 billion, up roughly 43% year over year and 15% quarter over quarter, suggesting the enormous capital committed to compute is finding paying demand quickly.
From Tokenomics to Jevons
The note is, in part, an honest update of a prior view, and Flight says so plainly. In June's Tokenomics, the team flagged declining token-price baskets as a challenge to the AI complex, with market pricing assuming compute scarcity far into the future. They flagged Jevons Paradox, the idea that cheaper inference could stimulate enough incremental consumption to sustain or increase compute demand, as a plausible offset, but treated it as second order. The simultaneous decline in GPU rental and token prices into mid-June deepened the concern that end demand itself might be weakening.
That relationship, Flight argues, has since broken down. GPU rental prices measured by Ornn have rebounded even as token prices continued falling. This emerging divergence between token and compute prices is the pivot of the whole piece: tentative evidence that Jevons dynamics are becoming dominant. Flight is careful to note these markets remain young, fragmented and opaque, so the signal warrants caution, but the change in the relationship is encouraging.
Compute as a Commodity
A second insight: compute is acquiring commodity characteristics. Capacity is finite in the short run, spot prices clear supply against demand, and the term structure carries information. Mid-July backwardation looked benign: H100 spot roughly 13% above 36-month term rates, B200s around 8%, with the shallower curve for newer Blackwell chips notable given obsolescence risk. Kalshi's prediction-market forward curves add price discovery: H100 spot fell 41% from its May peak but recovered 20% from June lows, while B200 pricing near $6.17/hr now trades above its May peak. Firm physical compute pricing coexisting with falling token prices reads as elasticity further up the stack.
The Quality Adjustment
Tokens are not homogeneous, and Flight insists the relevant price is the cost of completing a useful unit of work, not the price of a token. Capable open-weight models such as Kimi K3 let users substitute toward cheaper models without commensurate loss of output, so the cost of intelligence per unit of effective output is falling. But this cuts two ways: cheaper intelligence supports broader consumption while pressuring the rents available to closed-source frontier labs. That tension, Flight concedes, remains unresolved.
Markets Are Not Euphoric
Three market observations close the note. Q2 S&P 500 EPS growth is tracking roughly 33%, the strongest outside post-recession recoveries, driving the steepest earnings revision path since at least 2000. Hyperscaler revenues confirm demand. Yet Citadel's cross-asset PC1 growth factor sits at just +0.84σ, the 65th percentile of its five-year history. Restrained, given the fundamentals. As Flight puts it, "Reconciling AI's great expectations with rapidly declining token prices requires a highly elastic demand curve." July's evidence suggests we may have one.
Five Key Takeaways for Advisors and Investors
- Elasticity is the load-bearing variable. The AI trade's valuations depend less on today's utilisation than on whether cheaper intelligence keeps expanding consumption.
- Watch the token-compute divergence. Falling token prices alongside firm GPU rental prices is the healthiest configuration for the ecosystem; renewed co-movement lower would be a warning.
- Adoption is concentrating. The 90th-to-median spend ratio has doubled from 27x to 54x since October 2023. The winners of the AI dividend may be narrow, at the firm level and in portfolios.
- Compute term structures are a new dashboard. Backwardation, forward curves and prediction-market pricing now offer investors commodity-style signals on future scarcity.
- Sentiment has room. With cross-asset growth pricing at only the 65th percentile, strong fundamentals are not yet fully embedded, an argument against reflexively fading AI-linked assets.
Footnote:
1 Flight, Frank. "Elastic Expectations." Citadel Securities, 18 Aug. 2026, https://www.citadelsecurities.com/news-and-insights/global-macro-strategy/elastic-expectations/.
