Watch this variable: AI's $725 billion bill against its revenue

This section of The Variables exists for numbers whose next few readings will settle large arguments, and no number currently qualifies like this pair. For 2026, the four hyperscalers, Amazon, Google, Meta, and Microsoft, have guided to a combined roughly $725 billion of capital spending, overwhelmingly on AI data centres: up about 77% from around $410 billion in 2025, which was itself a record that embarrassed every forecast made the year before¹. Amazon alone plans about $200 billion; Google $175-185 billion; Meta $115-135 billion; Microsoft $110-120 billion¹.

Against that bill stands the revenue of the thing being built for. OpenAI, the sector's largest pure vendor, crossed $20 billion in annualized revenue during 2025, booking about $13 billion for the year, and passed $40 billion annualized by mid-2026, among the fastest revenue scalings in business history, while losing more in 2025 than it earned². Add every rival lab and every cloud's AI line and the industry's direct AI revenue remains an order of magnitude below a single year's infrastructure bill. That gap is either the greatest anticipatory investment since the railways, or the largest misallocation since the fibre glut. The next eight quarters of this ratio will start to say which.

One caution before the scoreboard: the two numbers are not a like-for-like pair. Hyperscaler capex builds infrastructure that serves ordinary cloud, storage, and networking as well as AI, and OpenAI's revenue is one company's line, not the industry's. The ratio is a motivating signal, not a valuation, and infrastructure investment exceeding current revenue is what anticipatory building always looks like. The questions that actually settle the bet are narrower: whether utilisation of the new capacity is rising, whether inference revenue grows into it, whether margins on AI workloads improve, and whether the hyperscalers' own disclosures begin separating AI returns from the cloud business that hosts them. This page tracks the headline ratio because it is measurable quarterly, and reads it through those four questions.

The bull case, stated fairly

The spenders are not speculating with other people's money; they are the most profitable companies on earth, funding construction largely from operating cash flow that their existing businesses throw off regardless. Their argument runs: inference demand is doubling on curves visible in their own order books, the constraint on AI revenue is capacity, not appetite, and the fastest way to lose the next platform is to be the cloud that quoted eighteen-month waiting lists. On this reading the revenue gap is a timing artefact, the same one the railways, electrification, and the internet each displayed, infrastructure by nature preceding the traffic that justifies it. They also note, correctly, that a data centre is not a sunk cost in the fibre-glut sense: compute is fungible across workloads, and the buildings, power contracts, and cooling outlive any one chip generation.

The bear case is equally concrete. Depreciation is the bull case's shadow: GPUs age like fish, not wine, and $725 billion of kit depreciating over five years must be earned back at margins that inference's price collapse keeps compressing. Financing has begun migrating from cash flow to debt and off-balance-sheet vehicles, the classic late-cycle signature. And the demand curves the spenders cite are partly circular: labs buy compute with money invested by the companies selling the compute, a loop that flatters everyone's growth until the day it does not. Both cases are serious. That is what makes the ratio a variable rather than a talking point.

Scale demands its comparisons, because $725 billion has stopped meaning anything unaided. It approaches the world's annual military spending growth, exceeds the global fertilizer, shipping-fuel, and battery markets combined, and represents, in a single year's guidance from four firms, more than the CHIPS Act, the EU's chips act, and every national AI strategy on earth stacked together. Construction of this magnitude is already macroeconomics: analysts attribute measurable fractions of US GDP growth to data-centre building, and the towns hosting it experience the boom the way oil towns once did, land, wages, substations, and school budgets moving together. Whatever the ratio's resolution, the spending is already a real economy, with real payrolls, that a correction would really end.

The concentration reading matters as much as the sum. Four boards, advised by overlapping bankers, are making correlated bets with a combined sum that private markets could not assemble and most governments would not dare; the history of infrastructure manias is substantially a history of exactly this correlation, every railway baron reading the same traffic projections. The mitigating difference is that these four fund from cash flows rather than leverage and can each stop unilaterally without a creditor's permission. The aggravating similarity is that none can stop first without conceding the platform, which is how rational actors ride a correlated bet past its fundamentals. Both dynamics are in the price of every reading this page will publish.

How to read the next data points

Watch the ratio's two ends separately, because they answer different questions. The revenue end tests adoption: annualized run-rates at the labs, AI line items in cloud earnings, and, more telling than either, the boring enterprise metrics, seat renewals, usage per customer, the share of pilots surviving to production. Doubling runrates justify patience; decelerating ones reprice everything downstream, from Nvidia's order book to the gas turbines Georgia is permitting. The capex end tests conviction: guidance revisions are the spenders' own confession, and the first hyperscaler to cut is worth a hundred analyst notes. So far every revision has been upward¹.

Watch the seams between them too. The electricity piece tracks whether the grid can even deliver the guided build-out, with a fifth of projects at risk of connection delays; the chip pieces track the sold-out 2-nanometre capacity that converts capex guidance into physical wafers; the inference-cost piece tracks the deflation that simultaneously grows usage and shrinks margin per use. The capex-revenue ratio is not one number but the summary of that whole system, which is exactly why it earns the watchlist's top slot.

The revenue side needs one more decomposition, because 'AI revenue' hides three different qualities of dollar. The cleanest is end-customer subscription and usage revenue, the $20-to-$40-billion run-rates at the labs and the copilot lines in enterprise software: real demand, priced monthly, cancellable, and therefore informative². The middle tier is cloud AI revenue, some of which is labs renting compute funded by their own investors, informative only net of that loop. The murkiest is committed-contract announcements, the hundred-billion-dollar compute deals whose press releases outrun any delivery schedule; they measure conviction, not consumption. The watch discipline is to weight the first tier, discount the second, and read the third as sentiment, which most headline coverage inverts.

History offers the base rates. The railway manias built too much, ruined their financiers, and left the track that industrialised two continents; the electricity build-out overshot regionally and still electrified everything; telecoms fibre bankrupted its layers and bequeathed the internet its backbone at ten cents on the dollar. Anticipatory infrastructure booms tend to be right about the future and wrong about the timing, punishing the builders and rewarding the users. The question this variable will answer is not whether AI matters, the usage curves have settled that, but who pays for the decade between the concrete and the cash flows: shareholders, creditors, or, through the productivity the kit enables, nobody at all. That last outcome has never happened before, which is the precise content of the phrase 'this time is different', and why it earns a watchlist rather than a verdict.

What would settle it

Three resolutions are possible, and each has a signature. Convergence: AI revenue compounds into the hundreds of billions while capex growth moderates, the ratio closing from both ends, the railway outcome. Correction: revenue decelerates first, guidance cuts follow within two quarters, and the adjustment cascades through chips, power, and construction, the fibre outcome, painful and, like fibre, probably leaving infrastructure the next decade gratefully inherits. Or the muddle: revenue grows fast but not fast enough, returns on the marginal data centre sag quietly, and the spending normalises into a permanently larger, lower-return cost of doing cloud business, the outcome nobody narrates because it has no crash date. Assign your own weights. This page will publish the readings.

One asymmetry keeps the correction scenario from being a rerun of 2000, and one keeps it from being dismissed. Unlike the dot-com names, the four spenders earn hundreds of billions annually from businesses that exist regardless, so a bust would wound valuations, suppliers, and towns without touching solvency at the centre; the system's leverage sits in the second ring, the neoclouds, the GPU-backed loans, the power developers building against letters of intent. But unlike 2000's investors, today's have that history in living memory and keep spending anyway, which means either the evidence of demand is genuinely better this time, or the fear of missing the platform has simply outgrown the fear of writing it off. Both propositions are testable against the next four quarters, which is the point of the page.

A closing note on method, since this piece inaugurates the section's format. Variables to Watch entries commit to numbers in advance, the readings that would change our mind, so that hindsight cannot quietly rewrite what we expected. For this one: sustained doubling of lab run-rates through 2027 and unrevised capex guidance reads as convergence; a first guidance cut plus decelerating run-rates reads as correction; anything else is the muddle. The numbers arrive quarterly, four earnings calls at a time, and the largest capital allocation in corporate history will be legible in them to anyone who keeps the two columns side by side. We will be here, keeping them.