Everything is a Compute Trade
Last month we argued that semiconductors had become the market’s center of gravity. This month goes one level up, to what the chips are actually a claim on. The AI trade now drives equity returns and a disproportionate share of American growth, and it reduces to a single underlying variable: demand for compute. Whether or not you chose that exposure, you own it. So the discipline is to measure it honestly. Compute demand has two useful indicators, volume and price, and a third that tests whether the demand is economically real rather than subsidised.
The arithmetic is straightforward. Since the ex-AI index began trading in February, the S&P 500 has returned 11.8% while the same index stripped of its AI enablers has returned 4.7%. Both are positive. One is seven points better, and the separation has been open continuously since April.
ING puts AI-related investment at between a third and a half of US GDP growth over the past year, depending on how narrowly the category is drawn. Second-quarter growth itself slowed to 1.5%, so that share is a large piece of a shrinking number.
Concentration on this scale is consistent with both genuine booms and bubbles, and we resist treating it as proof of either. The month-to-month path is also far noisier than the trend, as the exhibit shows. What the gap does mean is practical: an investor holding a broad index fund holds a concentrated bet on compute demand.
Tokens are the unit of account here, the kilowatt-hours of the intelligence economy. Google processed 9.7 trillion of them per month in May 2024, roughly 480 trillion a year later, and more than 3.2 quadrillion by May 2026. That is a 330-fold increase in two years and roughly sevenfold in the last twelve months.
Reasoning models and autonomous agents consume between fifteen and a thousand times the tokens of a simple chat exchange, which means consumption per user is compounding on top of growth in users. The demand curve is being bent by what the models are asked to do, not just by how many people ask.
The figures are self-reported and unaudited, and volume is not the same as value, so treat the series as directional rather than precise.
If volume tells you demand exists, price tells you whether it exceeds supply. One-year contract rates for renting an H100, the workhorse training and inference chip, fell from about $3.05 per GPU-hour in April 2023 to a trough of $1.70 last October. That decline was unremarkable. Datacenter chips shed rental value as faster silicon arrives and buyers migrate up the stack, and the H100 fell at roughly the rate prior generations did.
Rates now sit near $2.80, a 65% recovery from the trough and well above where that aging curve would have carried them. Rates turned up while Blackwell, the newer and faster generation, was shipping in volume. A genuine glut shows up in the previous generation first. Instead the A100, two generations back and six years old, rents for more today than it did a year ago. Buyers paying up on one-year terms while spot has halved since early 2024 is the harder signal, because they are paying to lock capacity rather than take their chances.
A third meter tests whether that demand is economically durable. Since the end of June, usage-weighted token prices have fallen about 40% while H100 spot rates rose roughly 20%, and AI spending per employee at the most aggressive firms rose 49% in a single month. Falling price of intelligence, rising price of the machinery that produces it, and rising total outlay is consistent with elastic demand, where cheaper units drive more than proportionally more consumption. The three series come from different samples, so read it as corroborating rather than conclusive.
The forward market is already publishing a view. Compute forward curves show modest backwardation, with H100 spot about 13% above the 36-month term rate and B200 about 8%, which says the market expects scarcity to ease but only modestly. On October 5 that view is scheduled to become exchange-traded, when the first regulated compute futures are set to list on CME Group’s NYMEX.
| Our three questions | Confidence | What would change it |
|---|---|---|
| Is compute demand real, and in excess of supply? | High | How much of the price signal reflects scarce electricity rather than scarce silicon. |
| Is that demand economically durable? | Moderate to high | Backlog concentration in two cash-burning labs; a step change in model efficiency; edge substitution at the margin. |
| Is it priced attractively from here? | Tentative | Index concentration at 45%; a forward curve that already embeds easing scarcity rather than perpetual boom. |
Demand exceeding supply is close to observable: volume, price, and elasticity point the same way, and second-quarter cloud revenue grew 37% at AWS, 43% at Azure and 82% at Google Cloud.
Durability is harder, though we lean toward the upper end of it. Transaction-level data shows paying enterprises buying more as intelligence gets cheaper, which is the signature of utility rather than subsidy. Our own consumption points the same way. The tokens this firm burns on research and portfolio work have compounded well past anything we would have forecast a year ago, and the driver has been capability rather than headcount: the same people asking the same questions now generate far more computation per question.
Against that, an estimated half of the roughly $2.1 trillion in contracted cloud backlog traces to two cash-burning research labs, though no hyperscaler discloses backlog by customer and Amazon has rejected the framing. The cleanest disclosed figure is Microsoft’s, whose commercial obligations grew 99% year over year but 26% once OpenAI commitments are stripped out. Adoption also stays narrow, with the median firm raising AI spend per employee 9% in July against 49% at the top percentile.
Valuation is hardest, and we sit barely on the constructive side of it: Citadel Securities puts cross-asset growth pricing near the 65th percentile of its five-year range even as second-quarter earnings track up 33%, which says the market is less euphoric than the rhetoric around it. The risk we take most seriously is that a step change in inference efficiency, or capable models running locally, shaves demand growth below what several hundred billion dollars of annual capital spending assumes. That question is at least empirical: substitution would show up first in older-chip rates, and A100s are signing contracts into 2029.
From here we watch GPU rental rates, which the forward curve will price daily from October 5; enterprise spending breadth, which tells us whether the median firm ever closes the distance to the leaders; and backlog concentration, which tells us whether revenue reliance on two labs widens or narrows.
The financing consequence matters beyond technology. If developers expect sufficiently high returns from compute, they can justify borrowing at rates that other projects cannot. That additional demand for capital can put upward pressure on real interest rates across the economy. Other projects may still secure funding, but at higher hurdle rates. Productive investment and more expensive money can coexist, and the benefits of the buildout and the burden of financing it need not fall on the same businesses.
The RQA Economic Forecast Model rose to 0.39 in August, its fourth gain in five months and the strongest reading since November 2024. The model has strengthened substantially since January, when it sat barely above zero. What has changed since spring is where the strength comes from: output surveys, residential permits, and credit conditions all improved in August, joining a consumer that has carried the index most of the year. Labor remains the one channel subtracting from the forecast model, and the level is still well below what a mature expansion produces.
July extended the inflation break. Headline CPI eased to 3.3%, with energy again doing most of the work, and core CPI matched its 2026 low at 2.5%. Core PCE has been stickier, holding at 3.5%, so the disinflation is not yet uniform across the measures. Those readings predate the current move in crude, which has since rallied back above $90 on renewed strikes around the Strait of Hormuz, up roughly 13% in a month and 38% from a year ago, and the Energy Information Administration does not assume Middle East production normalises until early 2027. Two months of disinflation are real, and unlikely to hold at $90 oil.
Output is the surprise. Manufacturing posted its strongest standardised reading in more than a year, services held solidly in expansion, and residential permits rebounded into positive territory at their strongest reading of the window. The consumer cooled from an unusually strong June without turning, with retail sales and real spending both moderating while staying comfortably in growth.
Labor is best described as full employment with hiring stalled. Jobless claims sit far below year-ago levels, so layoffs remain contained. Payrolls are flat to negative year over year and the employment-to-population ratio keeps slipping, so net employment growth is weak. The composite reads positive because the claims side dominates it. That combination is how the unemployment rate can sit near four percent while the flow measures look soft.
The curve stayed positively sloped and corporate spreads are tighter than a year ago, which normally reads as an all-clear. But the 30-year Treasury reached 5.34% in August, its highest since 2007, and has now closed above five percent on more days than in any year since 2006. The move extends beyond near-term Fed expectations: the two-year yield has moved five basis points since the end of June. It is a duration problem, and AI financing is one source of the pressure alongside federal borrowing, term premium, and inflation expectations, and an increasingly important one. AI-linked issuance is running near $220 billion this year, about 13% of all investment-grade supply, and Bank of America estimates that corporate and mortgage supply together have added roughly three tenths of a point to the ten-year.
| Indicator | Aug-26 | Jul-26 | Jun-26 | May-26 | Apr-26 | Mar-26 | Feb-26 | Jan-26 | Dec-25 | Nov-25 | Oct-25 | Sep-25 |
|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Labor | ||||||||||||
| Non-Farm Payrolls (YoY%) | -0.4 | -0.5 | -0.4 | -0.5 | -0.5 | -0.5 | -0.3 | -0.0 | 0.2 | 0.4 | 0.3 | 0.5 |
| Initial Unemployment Claims (Inverse YoY%) | 11.4 | 9.6 | 8.9 | 10.4 | 14.9 | 6.2 | 12.4 | -1.0 | 9.1 | -1.4 | -0.9 | 0.0 |
| Employment-to-Population Ratio (YoY%) | -1.2 | -1.2 | -0.8 | -1.5 | -1.2 | -1.0 | -0.5 | -0.5 | -0.3 | -0.5 | -1.0 | -0.7 |
| Average Weekly Hours Worked (YoY%) | 0.3 | 0.6 | 0.3 | 0.0 | 0.0 | 0.6 | 0.9 | 0.0 | 0.0 | 0.0 | 0.0 | 0.0 |
| RQA Labor Composite (YoY%) | 2.5 | 2.1 | 2.0 | 2.1 | 3.3 | 1.3 | 3.1 | -0.4 | 2.2 | -0.4 | -0.4 | -0.0 |
| Commercial Output | ||||||||||||
| ISM Manufacturing PMI (% over Base) | 11.2 | 6.6 | 8.0 | 5.4 | 5.4 | 4.8 | 5.2 | -4.2 | -4.0 | -2.4 | -2.2 | -2.2 |
| ISM Services PMI (% over Base) | 8.2 | 8.0 | 9.0 | 7.2 | 8.0 | 12.2 | 7.6 | 8.8 | 5.2 | 4.8 | 0.0 | 4.0 |
| Industrial Production Index (YoY%) | -1.0 | -1.3 | -0.9 | -1.3 | -2.0 | -1.6 | -1.1 | -0.8 | -0.2 | 1.6 | 1.2 | 0.8 |
| Residential Real Estate Permits (YoY%) | 6.6 | -2.1 | 1.4 | 2.1 | -7.4 | -5.5 | -2.4 | -4.8 | -12.8 | -7.3 | -8.1 | -11.1 |
| Income & Consumption | ||||||||||||
| Real Personal Incomes (ex. Transfer Receipts) (YoY%) | 0.4 | 0.7 | 0.2 | -0.3 | 1.1 | 2.0 | 1.1 | 1.3 | 1.9 | 2.1 | 1.8 | 1.9 |
| Retail Sales (YoY%) | 1.8 | 3.2 | 2.5 | 0.7 | -1.0 | -0.6 | -0.6 | -1.4 | -1.3 | -0.8 | -0.3 | 0.0 |
| Real Personal Consumption Expenditures (YoY%) | 1.8 | 2.1 | 1.8 | 1.2 | 1.2 | 1.7 | 1.7 | 1.1 | 1.2 | 1.5 | 1.6 | 2.0 |
| RQA Consumer Spending Composite (YoY%) | 1.8 | 2.6 | 2.1 | 1.0 | 0.1 | 0.5 | 0.5 | -0.1 | -0.0 | 0.4 | 0.7 | 1.0 |
| Financials & Sentiment | ||||||||||||
| Treasury Yield Curve Spread - 10-Yr Less 3-Month | 0.8 | 0.9 | 0.6 | 0.8 | 0.7 | 0.6 | 0.3 | 0.6 | 0.5 | 0.1 | 0.2 | 0.1 |
| Treasury Yield Curve Spread - 10-Yr Less 2-Yr | 0.4 | 0.5 | 0.3 | 0.5 | 0.5 | 0.5 | 0.6 | 0.7 | 0.7 | 0.6 | 0.5 | 0.6 |
| Corporate Bond Spreads (Inverse YoY%) | 33.8 | 30.4 | 28.0 | 16.4 | 12.5 | 15.0 | 12.9 | 1.7 | 6.2 | 10.3 | 16.2 | 11.0 |
| U.S. Monetary Base (YoY%) | -3.8 | -4.5 | -1.9 | -4.6 | -5.5 | -4.0 | -3.8 | -4.1 | -5.6 | -3.7 | -2.0 | 0.3 |
| S&P 500 Return (YoY%) | 19.0 | 18.1 | 20.9 | 28.2 | 29.4 | 16.3 | 15.5 | 14.9 | 16.4 | 13.5 | 19.9 | 16.1 |
| Consumer Sentiment (YoY%) | -11.2 | -10.5 | -18.5 | -14.2 | -4.6 | -6.5 | -12.5 | -21.3 | -28.5 | -29.0 | -24.0 | -21.4 |
| Inflation & Money Supply | ||||||||||||
| CPI (YoY%) | 3.3 | 3.4 | 4.2 | 3.8 | 3.3 | 2.4 | 2.4 | 2.6 | 2.7 | 2.8 | 3.1 | 2.9 |
| Core PCE (YoY%) | 3.5 | 3.4 | 3.6 | 3.5 | 3.4 | 2.9 | 2.9 | 2.8 | 2.6 | 2.5 | 2.8 | 3.1 |
| M2 Money Supply (YoY%) | 5.0 | 5.2 | 5.1 | 4.3 | 4.2 | 4.6 | 4.1 | 4.1 | 4.1 | 4.6 | 4.7 | 4.8 |
On the levels framework, growth against zero and inflation against 2%, the economy sits where it has since January: Quadrant II, the inflationary boom, with the growth model at 0.34 and CPI at 3.6% on the three-month basis the map uses. The inflation axis ticked lower this month for the first time since February, which is the turn last month’s edition was waiting for. It arrived in the same week oil traded back above $90.
The disinflation in the exhibit is genuine and already dated. If the Hormuz premium persists, the inflation axis could turn back up within a couple of prints, and the question stops being whether the map reaches Goldilocks and starts being whether growth holds while prices climb again. Quadrant III as we define it requires growth below zero, which is a long way from 0.34, so stagflation is not the base case. Rising inflation against slowing growth would move the reading in that direction, and that combination is the one to watch.
When capital itself is repricing, what matters is balance-sheet strength, cash generation, and whether a business earns a return above its cost of financing. The hyperscalers are both large borrowers and large cash generators. Borrowing alone therefore tells us little; the question is what that investment earns. Energy is a major risk to further disinflation, valuation now turns on the long end, and the September meeting arrives with three Fed governors already on record wanting a hike.
Growth is broadening and the model is at its best level since late 2024, but two variables that could shape the next two quarters sit outside the data we just published: the price of oil and the price of long-duration money. Quadrant II, with the risks now running in both directions.
Disclosures. These materials have been prepared solely for informational and educational purposes and do not constitute investment advice or a recommendation to make or dispose of any investment or to engage in any particular investment strategy. Information and data shown were obtained from sources believed to be reliable, but accuracy is not guaranteed. All investments involve risk, including the potential loss of principal. Past performance is not indicative of future results. References to specific securities and issuers are for illustrative purposes only and are not intended as recommendations. Third-party research and market commentary, including material published by Citadel Securities, Bank of America, Jefferies, Goldman Sachs Research, ING Research, SemiAnalysis, Deutsche Bank Research, Bloomberg, and The Information, are cited for context and do not represent the views of RQA. Certain third-party figures are reproduced as summary statistics with attribution. Token throughput figures are directional company disclosures and are not audited. Richmond Quantitative Advisors, LLC is an SEC-registered investment adviser; registration does not imply a certain level of skill or training.