What the arithmetic says about whether the AI buildout can ever pay for itself
There is a number at the centre of the artificial-intelligence boom that almost no one says out loud, because once you write it down the whole thing starts to look strange. So let us write it down.
The spending
In 2026, the four largest American technology companies — Microsoft, Alphabet (Google), Amazon and Meta — are guiding investors toward roughly $725 billion of combined capital expenditure in a single year, the overwhelming majority of it artificial-intelligence infrastructure: graphics chips, custom silicon, data-center buildings and the electricity to run them. That is up about 77 percent from roughly $410 billion the year before.
Looking further out, Goldman Sachs now expects those same four companies to spend a combined $5.3 trillion between 2025 and 2030.
To put $725 billion in one year into perspective: it is larger than the annual economic output of most countries on Earth.
The earning
Microsoft, the most transparent of the giants, recently disclosed that its entire AI business had reached an annual revenue run rate of $37 billion, growing a remarkable 123 percent year-on-year. OpenAI, the highest-profile pure-play, is running at roughly $25 billion. Adding contributions from Google, Amazon, Meta, Anthropic and others, the total identifiable revenue the world currently pays for AI itself is plausibly somewhere around $150 billion a year. (This last figure is an estimate; the companies do not all break out “AI revenue” cleanly, which is itself part of the problem.)
Spent in 2026: ~$725 billion. Earned from AI per year: ~$150 billion.
The industry is spending, in a single year, nearly five times what the entire world pays it for AI annually.
The asset that dies
A factory lasts forty years. A railway lasts a century. Those are investments in the ordinary sense: you pay once, the asset works for decades, and after it has paid for itself, everything it earns afterwards is profit. There is a harvest.
AI hardware is not like that.
Operator economics and secondary-market data paint a clear picture of rapid value decay. Analyses of top-of-the-line Nvidia H100 accelerators show strong returns in early years that turn sharply negative as newer generations arrive. On the resale market, an H100 typically trades for less than half the price of a new one by year three.
(This gap between accounting depreciation of five to six years and shorter real economic utility is not trivial: investor Michael Burry has argued that stretched schedules let the industry understate costs and overstate profits by roughly $176 billion between 2026 and 2028.)
The math, made simple
Here is the arithmetic the spending implies. None of it is exotic — it is the kind of sum you could do on the back of a napkin.
Cloud computing is a high-margin business, but not infinitely so. After you pay for electricity, cooling, networking, staff and the cost of the money itself, a strong operator keeps perhaps 30 cents of net profit on every dollar of revenue. That is a generous assumption — these are among the most profitable companies in history.
At a 30 percent net margin, recovering $1 of investment requires about $3.30 of revenue.
So the $725 billion spent in 2026 alone would need roughly $2.4 trillion of cumulative AI revenue to pay itself back. At today’s run rate of around $150 billion a year, that is about sixteen years of every dollar the world currently pays for AI — to recover a single year of building.
And here is where the word absurdity earns its place.
The treadmill
If the hardware were a factory, you would simply wait sixteen years and collect. But the hardware dies in three.
That means the $725 billion is not a one-time payment you recover at leisure. To stay competitive, you must buy it all again roughly every three years — and at the moment the bill is not merely repeating, it is growing.
Take the compute already deployed — well over $1.5 trillion built up across 2023–2026 — and assume a three-year life. Simply standing still, replacing what wears out, costs on the order of $500 billion every single year, forever, before a cent of growth or return.
At a 30 percent margin, covering that replacement bill alone would require roughly $1.6 trillion of AI revenue per year — more than ten times what the industry earns from AI today — just to keep the lights on and break even on the treadmill.
This is the heart of it. The investment never converts into an owned, paid-off asset. It behaves less like a building and more like a salary — a permanent, rising cost of merely continuing to exist in the business.
So is it madness?
Not necessarily — and it is worth being honest about why, because the bull case is real.
The correct comparison is not a factory or a railway. It is a semiconductor fab. Chipmakers have always lived on exactly this treadmill: each generation costs more than the last and is obsolete in a few years. It is survivable — TSMC runs that treadmill and is enormously profitable — but it is a filter. It rewards only three things: vast scale, relentlessly compounding demand, and genuine pricing power. It bankrupts everyone else.
Two forces can save the AI buildout:
- Each new chip generation does far more useful work per dollar. If performance-per-dollar keeps improving faster than the treadmill spins, the same revenue covers an ever-larger pile of computing, and the companies grow into the spending rather than drowning under it. This is the entire optimistic case.
- Old chips do not go to zero — they cascade. Yesterday’s frontier training chip becomes today’s cheaper workhorse for everyday tasks, earning a declining tail rather than dropping straight to scrap.
If demand and efficiency both compound fast enough, for long enough, the silicon underneath becomes a consumable while the real, durable asset — the software franchise, the subscriptions, the user lock-in, the data — quietly becomes valuable. That is the bet. Stated plainly, it is this: that performance-per-dollar and demand will both keep compounding faster than the hardware decays.
Who pays if the bet is wrong
The danger is not evenly shared.
The cash-engine giants — Google with its advertising, Microsoft with Office and Azure, Amazon with retail and cloud — can fund the treadmill out of existing profits more or less indefinitely, treating AI as a defensive cost of staying relevant. For them the likely outcome is not collapse but permanently thinner margins: a golden age of profitability quietly traded away for the privilege of staying in the race.
The fragile links are the debt-funded and pure-play builders. Borrowed money does not care that your collateral is worthless in three years; it still demands repayment on schedule. The moment rapid obsolescence meets heavy leverage, ordinary investment risk turns into solvency risk. The cautionary precedent is the telecom and fiber-optic boom of the late 1990s: the fiber that was laid eventually got used and proved enormously valuable — but most of the investors who paid for it went bankrupt first. The assets outlived the capital.
The bottom line
Strip away the excitement and the math says something simple and uncomfortable:
The world’s most valuable companies are spending trillions on assets that expire in three years, to chase a revenue stream currently a fraction of the cost, on the faith that efficiency and demand will compound fast enough to outrun the decay — and they are smoothing the accounting in the meantime to make the present look calmer than it is.
It might work. For two or three of them, it probably will. But it is not an investment in the comfortable old sense of the word. It is a treadmill set to a steepening incline, and everyone has agreed to keep running, because the one thing more dangerous than the cost of running is being the first to step off.
A note on the numbers
This article mixes two kinds of figures, and it is worth keeping them apart.
Hard, sourced facts (capital-spending guidance, the $5.3 trillion forecast, Microsoft’s $37 billion run rate, the H100 resale and depreciation context, the Burry estimate) come from company disclosures and reporting listed below.
Illustrative arithmetic (the $3.30-of-revenue-per-dollar rule, the ~$2.4 trillion payback, the ~$500 billion-a-year treadmill, the ~$1.6 trillion break-even) are deliberately simple back-of-the-envelope models built on a stated 30 percent net-margin assumption and a three-year hardware life. Change those assumptions and the numbers move — but the shape of the problem does not. The “~$150 billion of total AI revenue” figure is an estimate, because the companies do not all report AI revenue separately.
Sources
- Combined 2026 capital-expenditure guidance (~$725 billion, up ~77% from ~$410 billion in 2025): Financial Times Q1-earnings compilation, reported by Tom’s Hardware and Yahoo Finance, April 2026.
- $5.3 trillion combined capex 2025–2030 (Microsoft, Alphabet, Amazon, Meta): Goldman Sachs, reported by Yahoo Finance, 2026.
- Microsoft AI revenue run rate of $37 billion, up 123% year-on-year: Microsoft FY26 Q3 earnings release and GeekWire, April 2026.
- OpenAI run rate of roughly $25 billion (mid-2026): FutureSearch / Sacra revenue analyses, 2026.
- Nvidia H100 resale value below half by year three and depreciation debate: secondary-market data and analyses (e.g., Silicon Data, Hasrate Index reports, 2025–2026); Michael Burry comments via CNBC, November 2025; Deep Quarry / Substack, December 2025.
- Hyperscaler cloud gross margins (~70%) used to support the ~30% net-margin assumption: Microsoft cloud results, 2025–2026.
Figures are accurate as of mid-2026 and reflect guidance and estimates that the companies themselves revise frequently.