The Compute-Capex Mirage: $725 Billion Goes Into the Ground. What Comes Back Out?

Start with the number, because the number is not in dispute. Amazon, Google, Microsoft, and Meta have guided to roughly $725 billion of combined capital expenditure for 2026 — up seventy-seven percent from last year's $410 billion. Around three-quarters of it is AI infrastructure: GPUs, custom silicon, and the buildings and power to run them. Analysts already sketch a trillion-dollar year for 2027. These are filed, audited figures and public guidance. Nothing about the spending is a mirage.

The question is what the spending is called — and what it earns.

On every earnings call, the outflow arrives wrapped in the same language: strategic pivot, capacity expansion, AI transformation. Accounting treats it accordingly. Capital expenditure is not an expense that hits this quarter's operating margin; it is an investment that depreciates over years. So a company can pour tens of billions into hardware, tell the market it is buying the future, and keep its current profitability looking largely intact — while the actual question, is the AI business earning anything, gets deferred to the depreciation schedule. That is not fraud; it is how the rules work. But it means the income statement is structurally flattering to exactly this kind of spending, and everyone in the boardroom knows it. Whether the useful life assigned to a GPU matches the pace at which it becomes obsolete is one of the sharpest open questions on this desk — stretch the schedule and today's earnings look better while tomorrow absorbs the cost. Our Depreciation signal tracks precisely that gap.

Now put the desk's own meters beside the guidance. Our divergence read — market narrative against filings ground truth — sits at +4.06 standard deviations, the widest in the series. Our recycling ratio finds that for every dollar of outside equity actually reaching the model layer, roughly 15.5 dollars of compute has been committed — meaning the demand at the center of the boom is substantially funded by the vendors and partners inside it. And the payoff board shows adoption that is real but partial: most industries can show demonstrable use, far fewer can show returns that resemble the invoice. The cost side of this cycle is concrete, contracted, and enormous. The return side is a promise with a fiscal year attached that keeps moving.

Did AI do this, or did we?

The machines did not demand $725 billion. A specific competitive logic did: no executive is willing to be the one who underbuilt if the boom is real, and the market — so far — punishes hesitation more than it punishes spending. When the same market also lets that spending sit on the balance sheet as an asset rather than a cost, the rational move for every individual player is to build. The mirage is not any single company's capex; each firm's numbers are real. The mirage is the aggregate assumption that all of this capacity can be justified by revenue that does not yet exist — that everyone's spreadsheet can be right at once.

What we are not claiming: that the spending is fake, that a crash is scheduled, or that the infrastructure will never earn. The dot-com fiber glut eventually carried the internet — after the companies that laid it were gone. The honest statement is narrower and harder: the gap between committed capital and realized AI revenue is the widest measured number in this cycle, it is widening, and the accounting treatment ensures it can widen for a long time before any income statement forces the question.

So we hold the question open, with the meter running. When realized revenue starts closing the gap, our numbers will show it, and we will say so. Until then, $725 billion is going into the ground this year — and "the silence where the return should answer" is still the loudest sound in the filings.