Sullivan Brothers · Working thesis · AI infrastructure economics · July 2026

The queue, not the river.

Roughly seven trillion dollars is being committed to AI infrastructure on the belief that today's revenue growth is a river — a demand stream that compounds forever. This paper argues much of it is a queue: a large but finite backlog of automatable work, draining once. The technology can succeed completely and the capital can still fail to pay. Historically, that is the normal outcome — and with a third of the S&P 500 now riding on it, the outcome is everyone's problem.

This is an illustrative working model, not investment advice or a prediction. Nothing here accounts for any individual's circumstances. Consult a licensed financial advisor before acting on anything you read below.

01The arithmetic nobody disputes

Start with the numbers both sides accept. The four largest cloud companies — Microsoft, Amazon, Google, and Meta, the firms Wall Street calls "hyperscalers" because they build computing at industrial scale — plan roughly $725 billion of capital spending in 2026 alone. Capital spending, or capex, is money for physical things: buildings, chips, power lines. That figure is up 77% from a record 2025, Wall Street's baseline runs to $7.6 trillion cumulative through 2031, and an increasing share of it is borrowed rather than paid from profits. Against all that, total identifiable revenue from actual AI customers runs at roughly $120–150 billion a year.

$725B
Big-four capex planned for 2026
$7.6T
Baseline cumulative spend, 2026–2031
~$150B
Annual AI end-customer revenue, mid-2026
~10x
Revenue growth needed in four years, while costs also grow

Even that $150 billion is flattered. Some of it is the same dollar counted twice — one vendor books its cloud partner's billings as its own revenue — and some of it isn't outside money at all. A striking share of the industry's reported growth comes from the players funding each other:

The circular money machine The chipmaker Nvidia — sells the shovels The cloud landlords Microsoft · Amazon · Google Meta · Oracle The AI labs OpenAI · Anthropic · xAI sell the intelligence buy chips ($100B+/yr) invests in its own customers commit $100B+ for compute also invest billions in the labs (equity, cloud credits)
Solid blue arrows are money booked as revenue. Dashed purple arrows are investment flowing back the other way. When the chipmaker invests in the labs that buy the clouds' capacity, and the clouds invest in the same labs that commit spending back to them, a meaningful slice of everyone's "growth" is the group's own capital circulating — not new money from outside customers. Every dollar of genuinely outside revenue supports several dollars of reported revenue along this loop. This pattern has a history: vendors financing their own customers was a hallmark of the telecom bubble in 1999, for the simple reason that real outside demand doesn't need to be subsidized into existence.

Now the cost side. Hardware doesn't stay bought — it becomes an annual expense through depreciation, which is just accounting's way of spreading a purchase over the years it's useful, like recognizing a fifth of a delivery truck's cost each year for five years. Run the AI hardware already installed through generous assumptions and the annual bill looks like this:

Cumulative AI capex, 2024–2026~$1.3T
Annual depreciation @ 5-yr life~$260B
Power, cooling, operations (+50%)~$130B
Annual carrying cost of what's already built~$390B
Revenue required at 60% margins, break-even only~$650B/yr
In plain English: the chips and buildings already purchased need to bring in roughly $650 billion a year just to cover their own costs — before earning a dime of profit, and before the even larger 2026–27 spending wave lands. Actual revenue is about a quarter of that. Earning a normal return on the full $7.6 trillion program requires something like $1.5–2 trillion of annual AI revenue by 2030 — an industry roughly the size of all business software on Earth, built from scratch in four years, selling a product whose price falls about 90% every year.

The bulls do not dispute this math. They dispute what it means. Their answer deserves to be stated at full strength before it is challenged.

The index exposure

This is not a niche trade. The AI-linked cohort — the cloud builders above, plus Nvidia — now accounts for roughly a third of the S&P 500, the heaviest concentration in the index's history. That makes the return on the spending described here the dominant variable in the broad market itself, and it means the ordinary index portfolio carries this bet by default rather than by decision. The concentration also works mechanically in both directions: passive inflows lifted the same handful of names on the way up, and redemptions would press on the same handful on the way down. Whatever one concludes about the buildout, there is no neutral position available to a holder of the index.

02The bull case, stated fairly

The bull case rests on three claims, each with real evidence behind it.

The market is labor, not software. AI is priced against the work it replaces, and the global wage bill for knowledge work runs in the tens of trillions. Capturing even a low single-digit share dwarfs the capex. The demand signal is visible today: contracted future cloud spending ("backlog") in the hundreds of billions, enterprise accounts spending over $1 million a year doubling in months, the fastest revenue ramps in the history of software.

Cheaper computing means more computing, not less. Every time the price of a unit of intelligence falls, consumption has risen faster — the same dynamic that made a century of ever-cheaper electricity a century of growing electricity revenue. Casual chatbot users with capacity to spare are not the market; autonomous "agents" that work for hours on a task are, and an agent consumes thousands of times what a chat question does.

The frontier keeps receding. Yesterday's capability becomes nearly free, but nobody's demand sits still at yesterday's capability. Harder reasoning, longer tasks, video, and autonomy keep absorbing the efficiency gains, so the price of the best available work holds even as old work deflates to zero.

This is a coherent position. The thesis below does not defeat it by denying demand. It defeats it — if it does — by examining what the demand is made of.

03Three mechanisms working against the capital

Mechanism one

The deflation engine eats the asset base

Computing efficiency improves roughly tenfold per year. The bulls read this as a demand subsidy. But it cuts twice: the same curve that makes AI cheap makes last year's hardware uncompetitive. The cloud companies book their chips as five- or six-year assets; if the economic reality is closer to two or three — because a 2028 chip does the work of five 2026 chips — then true annual costs are 50–80% higher than the reported numbers, and today's committed orders are purchases of obsolescence. Falling compute prices don't destroy AI demand. They destroy the balance sheets of whoever owns last year's compute.

Mechanism two

The capture problem: usage up, revenue down

Even where demand is genuinely unlimited, revenue requires someone to hold price — and look at who would have to do the holding. Microsoft, Amazon, Google, Meta, and Oracle are racing to build nearly identical warehouses of nearly identical chips. OpenAI, Anthropic, Google, Meta, and xAI sell increasingly interchangeable intelligence, metered by the token — the kilowatt-hour of AI. Their customers, meanwhile, are businesses whose own prices are deflating, because their competitors bought the same AI. Five sellers of a commodity, selling to buyers who are themselves losing pricing power, on a cost curve falling 90% a year: that structure passes the windfall through to end consumers. The technology wins; nobody keeps the money. This is not a hypothetical. It is the modal outcome for transformative infrastructure:

Fiber optic, 1998–2002Traffic ↑ relentlessly · Capital destroyed

The internet was real. Demand eventually exceeded even the bubble buildout — used by firms that bought the fiber out of bankruptcy for cents on the dollar.

Long-distance callingMinutes ↑ · Revenue → ~0

Usage exploded as prices fell. Industry revenue nearly vanished. Consumers captured everything.

Digital photographyPhotos ↑ ~1000x · Industry revenue collapsed

The activity grew a thousandfold while the revenue pool shrank. Kodak died inside photography's greatest boom.

Anyone who sold computers in the 1980s has watched this movie from the inside: dominance looks permanent right up until the product standardizes, and then the margins belong to no one and the benefits belong to the buyers. Tokens are standardizing faster than any product in the history of the industry. The infrastructure bet requires not just that AI transforms everything — it requires the transformers to keep the money, against a century of precedent that they usually don't.

Mechanism three

The queue: today's growth may be a backlog draining, not a baseline compounding

Most office work is computationally simple — structured information moving between systems, the spreadsheet and its software equivalents. Two properties follow. First, the computing cost of replicating it is small: automating the routine output of the entire American office plausibly costs on the order of $200 billion a year at current prices, before any deflation — less than the datacenters cost annually to depreciate. Second, and more important, automation compiles. An AI agent that figures out a monthly reconciliation doesn't re-think it every month afterward; it writes the program once, and from then on the task runs as ordinary software, outside the AI economy entirely — leaving only a residue of monitoring and exception handling. AI doesn't just do the work cheaper. It converts the work into a form that no longer needs AI.

Under this reading, the revenue ramps everyone extrapolates are not a demand curve. They are a migration flow: the one-time cost of moving a finite backlog of automatable work into automated form. Migration flows look exactly like exponential growth while the backlog drains. Then they cliff. The demand was never fake. It was a queue.

Every historical infrastructure bust featured real, measurable, exploding usage right through the collapse of the capital that financed it. The technology succeeding and the capital being destroyed are compatible outcomes — historically, the default pairing.

04Where the thesis is vulnerable — the honest ledger

A thesis that can't be wrong isn't a thesis. Four objections have real force, and the model below carries each as an explicit dial rather than pretending it away.

Demand outside the queue may be enormous. Bounded-task logic doesn't govern everything. Adversarial spending — cybersecurity, algorithmic markets, model-versus-model arms races — escalates with the opponent, not with a task list. Frontier research buys maximum intelligence, not minimum cost. Robotics and real-time systems, if they arrive, burn continuous computing that can't be scripted away, because the physical world keeps changing. The buildout's actual repayment depends on these — which is itself remarkable, since none of them is what "enterprise AI adoption" means on an earnings call.
Even discovery may not compile. The strongest bull refuge is that scientific research resists being turned into a script. The rebuttal, taken seriously here: screening-type discovery already has been — protein folding, materials search, drug-candidate filtering. Whether open-ended hypothesis generation follows is unresolved, so it's a toggle in the model, not an assumption.
The maintenance residue might be a river after all. Compiled workflows still need verification, exception handling, and audit — and in regulated industries, trust requirements may keep AI checking AI indefinitely. If that residue is large, the queue leaves behind permanent revenue. If it's small, it doesn't. This single parameter separates the scenarios more than any other.
Timing: organizational slack slows the drain. The capability to automate the backlog exists today — any solo operator now doing the work of a former team can verify it. But inside larger firms, the gains pool invisibly as slack: the hour's work stretches back into the week, because no employee volunteers for the layoff and no manager shrinks their own department. Companies restructure on recession timelines, not technology timelines. This delays the migration — extending the revenue flow and helping the spenders — but it also means the queue drains in a lurch when a downturn finally forces the question, rather than in a smooth curve.
What the thesis concedes, and what it keeps. Conceded: AI usage will very likely keep growing for years; adoption is real; the backlog is enormous; nothing here predicts a crash on a date. Kept: growth is not the question — capture is. The sharpest version of the thesis needs only this: revenue per unit of work falls faster than units grow, in a market where no one can hold price, on a spending schedule financed with debt that cannot wait. Supply is being committed years ahead against demand that is partly a one-time migration, partly slack-gated, and wholly deflating.

05The model

All of the above, as arithmetic. The backlog drains at a migration rate that ramps in over several years (organizational slack gates early adoption); vendors take a one-time cut while each workflow is scaffolded; compiled work leaves a maintenance residue; elastic demand grows outside the queue. Price deflation begins compounding in 2027 rather than immediately — realized prices hold while capacity is scarce, as they are now, and erode once competition matures. The installed hardware base sets the revenue bar (dashed line) that the stack must clear. Three presets, or set every dial yourself.

Backlog remaining: —Compiled: —
The queueRoutine work not yet automated. Finite. Every task leaves once.
The token phaseAI agents scaffold the workflow. Today's revenue lives here.
CompiledOrdinary software plus oversight. Exits the AI economy, minus a residue.
Demand — the queue
$12T
60%
16%
6%
1.2%
Demand — outside the queue
$60B
40%
Pricing and capture
30%
Supply — the cost side
1.00x
4.0 yr
Peak AI revenue year
Revenue, 2030
Gap vs break-even, 2030
Queue 90% drained by
Migration flow (the drain) Maintenance residue Elastic demand Actual reported revenue, 2022–26 Required to break even (dashed)

Historical actuals are estimates assembled from company disclosures and reporting: ~$1B (2022, ChatGPT launches in November), ~$5B (2023), ~$15B (2024), ~$55B (2025), ~$140B annualized run rate (mid-2026) — mixing fiscal actuals, exit run rates, and gross-basis reporting, so treat as directional. The dashed requirement line is computed from reported capex throughout, so its history is firm even where revenue history is soft. Model components begin in 2025.

Queue depletion

The flat left side is a finding, not a gap in the data: through 2024, cumulative AI revenue was a rounding error against the labor pool. Whatever has happened to stock prices, the actual drain of the backlog has barely begun — which is precisely why the revenue ramp still has years of apparent strength ahead of it under either scenario.

Note what the "queue" preset shows: revenue keeps growing through roughly 2027 — matching the pace observable today — before it rolls over. Demand never disappears; the bust, where it occurs, comes from the cost side arriving faster than the capture. This is the uncomfortable implication: for the next several quarters, the queue and the river produce the same data. Strong growth now confirms neither side. Only the watch-list indicators below distinguish them before the divergence arrives.

06What would settle it — the watch list

Neither side gets to be unfalsifiable. These are the observable indicators that distinguish the river from the queue in real time — none of them is "the demand curve," because the demand curve looked spectacular in March 2000 too.

IndicatorRiver saysQueue says
Utilization of deployed capacityStays tight; electricity remains the binding constraintSlackens as compiled workflows exit; idle capacity appears
Realized price per token (blended)Holds up via premium frontier tiersRaces down; providers can't hold price on a commodity
How automated workflows runCompiled tasks keep re-invoking AI models (fat residue)Scripts run as ordinary software without models (thin residue)
Depreciation schedules5–6 year chip lives survive auditor scrutinyLives get shortened; a restatement marks the top
Datacenter debt marketsFinancings complete; borrowing costs stay tightSpreads widen, a deal fails — credit breaks before stocks do
Cloud-giant guidance languageCapex plans reaffirmed every quarter"Disciplined," "optimizing," "phasing" — one flinch gives all of them cover
Corporate disclosureAI spending grows as output growsThe first Fortune 500 CFO explains a 20% headcount cut with flat output — the queue drained in a lurch

07The thesis in one paragraph

AI does not need to fail for the buildout to fail. It needs only this: that a meaningful share of today's revenue is the one-time cost of migrating a finite backlog of routine work; that automated work compiles into ordinary software and exits the AI economy; that competition among interchangeable providers passes the deflating cost of intelligence through to end customers rather than to infrastructure owners; and that the debt-financed capacity was committed two years ahead of a demand schedule with all three of those properties. Under those conditions — each individually well-precedented — usage explodes, the technology remakes the economy, consumers capture the surplus, and the capital is destroyed. That is not a paradox. It is what transformative infrastructure has usually done to the people who financed it.

Calibration sources: hyperscaler 2026 capex guidance ~$700–775B (Q1 2026 earnings calls); Goldman Sachs cumulative baseline $7.6T, 2026–2031; hyperscaler debt issuance $108B (2025), $1.5T projected; AI end-revenue run rates ~$120–150B annualized mid-2026 (Anthropic ~$30B gross-basis Apr 2026; OpenAI ~$25B Feb 2026; Microsoft AI ~$37B), with gross/net double-counting caveat; index concentration per S&P Dow Jones data, 2026. This model is illustrative and directional, not investment advice — consult a licensed advisor before acting on any of it. Historical analogies (fiber 1998–2002, long-distance telephony, digital photography) describe capture dynamics, not predictions. Sullivan Brothers Inc. · Minneapolis · July 2026