Big Tech Will Spend About $700 Billion on AI Infrastructure in 2026. The Revenue Is Not Close.
Current guidance across the largest cloud and AI infrastructure providers puts combined 2026 capital spending somewhere around six hundred and sixty to seven hundred billion dollars, roughly double last year. Google in the region of two hundred billion, Amazon similar, Microsoft not far behind, Meta well over a hundred, Oracle around fifty. Tracked AI-attributable spending has gone from about thirty billion in 2020 to several hundred billion now.
Revenue at the pure-play model developers is growing fast and remains a fraction of the infrastructure being deployed on their behalf. This is the gap everyone points at, and the way people point at it is usually wrong.
The common framing treats it as a single-year profit and loss problem, as though the spending has to be justified by this year’s revenue. It does not. Capital expenditure depreciates across five to six years on the accounting, and the useful life of a data center shell is measured in decades even if the silicon inside it turns over faster. The right comparison is cumulative spend against the revenue expected across the asset life, and on that basis the gap is a bet rather than a discrepancy.
What makes it a serious bet rather than an obviously bad one is that the buyers are funding it largely from operating cash flow at businesses that print money doing something else entirely. Search advertising, retail, enterprise software. If AI revenue disappoints, these companies write down assets and continue. That is a very different risk profile from the telecom buildout people keep comparing it to, where the spending was done by companies whose only business was the thing being built.
The genuine constraints are physical and they are already binding. Power is the main one. Grid interconnection gaps run to gigawatts at individual providers, and a substantial share of planned 2026 projects face delay or cancellation on that basis alone. Supply chains for transformers, turbines, and high-bandwidth memory are tight. Announced capacity and energized capacity are diverging.
The financial spillover is now macro-relevant. Debt issuance connected to the buildout competes for the same capital as government borrowing, which is one of the reasons long-dated yields hit multi-decade highs this month. An investor with no view on artificial intelligence at all is nonetheless paying for the buildout through their mortgage rate.