What the AI build-out is really priced on — the arithmetic behind the boom, in five charts you can push on.
Revised 26 Jul 2026: watch-list baselines updated with street consensus and company guides ahead of the 28–30 Jul earnings wave. Previously revised 23 Jul after Alphabet raised its capex guidance to $195–205B — the first live test of this framework. Launch assumptions stay visible as each earnings print revises them.
Five technology companies are spending over a trillion dollars a year building AI infrastructure — more than their earnings and free cash flow, with the gap covered by debt. The Bank for International Settlements, in its 2026 flagship report, compares the moment to canal mania, railway mania and the dot-com boom: genuine breakthroughs that attracted more capital than eventual returns could justify.
This page walks through one piece of arithmetic that rarely appears on a single screen: how much revenue the build-out requires, where that revenue could come from, and what each answer implies for wages, jobs and growth. Every figure is an order-of-magnitude model, not a forecast — and every assumption is on a slider, so you can disagree with it directly.
Fast-depreciating hardware turns a pile of capex into a permanent annual bill.
AI chips are not railway track: they earn for a few years, not a century. Spread the capex over the hardware's life, add operating costs (above all electricity), and add the return investors require for the bet to count as vindicated rather than merely survived. The sum is the revenue the sector must generate every year.
Even the friendliest corner of these sliders leaves a floor near a trillion dollars a year. The capex slider is live for a reason: Alphabet's 23 July guidance raise ($195–205B for 2026, with 2027 "significantly" higher) is already pushing the sector total toward the top of its range.
There are only two pools of money that revenue can come from — and they differ by an order of magnitude.
If AI is a productivity tool, its revenue comes from the pocket all software comes from: corporate IT budgets. If AI is a substitute for workers, its revenue comes from the wage bill. The blocks below are drawn to true scale.
The dashed outline is the revenue floor. In "tool" mode it is larger than the entire software pool: AI would need to swallow every software product on earth just to break even.
That is the inference in one picture. When required revenue exceeds what the tools pool can conceivably supply, the residual must be drawn from the wage pool. The scale of the bet is itself the substitution thesis: what is being priced is not a better spreadsheet but a payroll line.
Revenue drawn from the wage pool converts directly into displaced wages.
An AI vendor cannot charge a customer the full wage of the worker it replaces — the customer keeps most of the saving, or there would be no reason to buy. The vendor's slice is the capture rate. Divide required revenue by the capture rate and you get the wages that must be displaced to fund it.
"Jobs equivalent" is a scale check, not a layoff forecast: the same wage loss can arrive as slower hiring, downgraded roles or thinner raises. The benchmark is the displacement usually attributed to Chinese import competition in US manufacturing, 1999–2011 — whose political aftershocks are still being felt.
The BIS lays out two failure modes — and its own model shows they are connected.
The Bank for International Settlements' 2026 Annual Economic Report treats the boom's two endpoints with unusual directness for a central-bank publication.
Returns fail to cover the revenue floor. Overvalued assets reprice; the circular financing between chipmakers, model labs and their investors unwinds; the capex that has been carrying GDP growth is cut — which is itself the recession mechanism. The report's historical parallels: canals, railways, dot-com.
The revenue is extracted from the wage pool at scale. In the report's modelled scenario, labour's share of income slides from roughly 60% today toward 20% over four decades — a transfer the economist Adam Tooze glosses, after Keynes, as a "euthanasia of the working class." Unlike past machines, the report notes, AI competes with human cognition itself.
The twist buried in the report's own model: displaced workers are also lost consumers. In the BIS's "demand bottleneck" scenario, substitution shrinks the wage base that buys the output, firms find further automation unprofitable, and growth sinks below trend — labour badly hit and the investment disappointed, at the same time. Success by substitution can destroy the demand that was supposed to validate it. The two forks are one road.
Stylised paths after the BIS's Box C scenarios. The amber path tracks the red one for as long as substitution pays — then the shrinking wage base stops buying the output, further automation turns unprofitable, and growth sags below the very trend the investment was meant to beat. The end state combines both failure modes: labour displaced and returns disappointed.
Whether workers win or lose compresses into about 1.5 percentage points of GDP growth.
Suppose the bet pays off in full: capital income roughly doubles over fifteen years, from $24T to about $48T a year across the advanced economies. Capital's claim is fixed in dollars; labour receives whatever output is left. That residual structure makes the growth rate decisive — drag the dial.
One boundary condition makes the dial honest: it is geared to capital earning its floor return. In the BIS's no-substitution scenarios — business as usual, or a bounded productivity boost — labour remains an essential input and its share never moves, at any growth rate. The dial below only starts turning once the bet pays off and capital's enlarged claim is fixed; from that point, growth decides how much of the loss workers feel.
Move the dial and watch which lines respond. Capital's trajectory never changes — its claim is fixed by the bet. Everything else is residual.
At trend growth (2%), capital's doubled claim can only be met by cutting wages in absolute terms. At 3.2%, wages exactly match their historical trend. Above 3.5% — a rate advanced economies last sustained in the postwar catch-up era — everyone gains, even as the share shifts. Each percentage point of growth is worth roughly 7–8 points of labour share.
The scenario that gives this page its title runs longer than the dial above: the BIS models labour's share sliding toward 20% over four decades, not fifteen years. Year by year the increment is small — roughly a point a year, never a cliff — which is what keeps the path politically quiet. Cumulatively it is nothing like the past: about ten times the drift advanced economies have absorbed since 1980.
And it rests on a stronger premise than the dial above. A one-time doubling of capital's claim needs only substitution — AI taking over a large share of existing tasks. A share that keeps falling for four decades needs the automation frontier itself to keep moving, which in the BIS's model happens only in its transformative branch: AI improving AI. The dial is a bet on substitution; this chart is a bet on recursion.
The historical line is a stylised advanced-economy average. The distance between the two endings — 56% versus 20% — is the size of the claim the scenario makes on the future.
But a share of income is also a share of leverage. Labour's bargaining power — the credibility of the strike, the cost of replacing a workforce, the political weight of the pay packet — scales with its claim on production. In a 20%-share economy that leverage largely dissolves: firms can produce without workers, so withdrawing labour loses its threat; household demand is sustained by transfers rather than wages; and workers bargain with the state for redistribution rather than with employers for pay. The wage bargain gives way to the transfer bargain — a different political economy, arrived at one imperceptible year at a time.
Set the dial to 3.5% and every indicator a worker actually experiences turns positive: wages grow 2.5% a year, above the ~2% historical trend — the best sustained pay run since the postwar decades. In lived terms, this is prosperity.
Over the same fifteen years, labour cedes eight points of income share (60% to 52%) while capital income doubles, compounding at nearly 5% a year. And capital's gains do not stay as income: they accumulate as assets, concentrating ownership even as pay packets swell. Absolute experience registers a win; relative position records a retreat. The tension stays invisible for exactly as long as growth holds — if it ever falters, the share losses are already banked, and the bargaining power that might have contested them has drained away in the good years.
The boom is unusually datable: its core asset expires on a schedule.
Three clocks are running. Markets can lose patience with the revenue gap at any time — a risk sharpened by the 2026 inflation shock and the rate rises now being priced (for the geopolitical backdrop, see Geo-V). The hardware bought in 2024–25 must be replaced around 2027–29, when a second capex wave has to be funded from revenue rather than faith. And the debt issued in 2025–26 rolls over around 2028–30. The clocks converge.
The near-term test is already underway. The numbers below, most within weeks of this page's publication, are the first readings on which fork the economy is on. The current baselines matter: growth rates at this scale decay every year, so the early prints must massively overshoot the ~40% multi-year average the revenue floor requires — even a spectacular-sounding 60–70% could mean the slope is breaking early.
| Date | The number | What it must show |
|---|---|---|
| 22 Jul — reported | Alphabet Q2 — Cloud +82% (expected +64%), backlog $514B; capex guide raised to $195–205B; shares −5% after hours. Full analysis → | Demand cleared emphatically; no capex trim — but the market punished the spend. The patience clock ticked. |
| 29 Jul 2026 | Microsoft Q2 — AI run-rate, last print $37B, +123% yoy; street at $87.6B revenue, EPS $4.24 | The cleanest single number in AI. Deceleration below ~80–90% starts breaking the required slope. |
| 29 Jul 2026 | Meta Q2 — capex guide $125–145B, lifted on component costs; own revenue guide $58–61B | AI monetised invisibly through ads; watch spending conviction, not an "AI" line. |
| Early Aug | Amazon Q2 — AWS growth, last print 28%, fastest in 15 quarters; guided sales $194–199B | Must hold or accelerate; slipping back to low-20s undercuts the demand story. |
| 7 Aug, monthly | US jobs report — white-collar and entry-level hiring; June: +57,000, unemployment 4.2% | The other ledger. Substitution revenue must eventually appear here — or it isn't substitution. |
| 26 Aug (est.) | Nvidia Q2 — guided $91B data-centre revenue | The capex-side thermometer: a soft guide means hyperscalers are easing off. |
| Autumn 2026 | Anthropic S-1 published pre-roadshow; late-2026 listing planned | First audited pure-play AI revenue trajectory. |
| 2026–27 | OpenAI listing — filed June 2026; delay to 2027 under discussion | Baseline: $20B annualised revenue at end-2025 against multi-billion quarterly burn. |
The signal is in the divergences, not the single prints. If the earnings columns keep beating while the payroll column stays flat, the "tool" world is quietly winning — and the arithmetic above says that world runs out of software budget around 2028. If both move together, the substitution rotation has begun. Either way, by roughly 2028 the depreciation wall forces the answer: the sector proves it can take wages, or it cannot.