The phrase "the bill is coming due" entered the AI conversation in June 2026, when the largest U.S. technology stocks shed about $2.7 trillion in combined market value over the month. That figure is a loss of stock-market value, not a spending total — but it landed because of the spending total behind it.

The scale of the build-out

The numbers are large by any measure. Combined capital expenditure — capex, the money spent on long-lived physical assets such as buildings, servers and chips, as opposed to day-to-day operating costs — at Alphabet, Amazon, Meta, Microsoft and Oracle rose from $162.3 billion in 2022 to $448.3 billion in 2025, according to data compiled by Visual Capitalist. For 2026, the four largest "hyperscalers" — the cloud giants that build and rent out massive data-center capacity — are guiding toward roughly $725 billion combined, up about 77% from 2025, Tom's Hardware reported.

Cumulatively, the bill runs far higher. Goldman Sachs in June raised its forecast for combined Meta, Microsoft, Amazon and Alphabet capex to about $5.3 trillion from fiscal 2025 through 2030, and analyst forecasts for total industry AI capex now top $1 trillion in 2027, per CNBC.

How it's being paid for

For years this build-out was funded from operating cash flow. That is changing. The Bank for International Settlements has flagged a shift toward debt and off-balance-sheet financing. Tech firms have moved more than $120 billion of AI data-center debt off their balance sheets using special purpose vehicles, or SPVs — separate legal entities that borrow against a specific project. Meta's roughly $30 billion Louisiana campus, for example, was financed via a joint venture with Blue Owl Capital that issued about $27 billion in debt through an SPV; Meta leases the finished site back, converting capex into an operating cost. The mechanics echo our earlier coverage of the AI-linked private credit and debt structures funding capital-hungry projects like SpaceX's.

The free-cash-flow squeeze

The pressure shows up in free cash flow — the operating cash left after capex. Analysts project hyperscaler free cash flow falling sharply as the build-out gets more expensive and component shortages spanning memory, chips, networking gear and power persist. Higher spending also feeds depreciation, the accounting practice of spreading an asset's cost across its useful life. Several firms have stretched server depreciation schedules from three or four years toward six, collectively trimming reported depreciation by an estimated $18 billion a year. Critics, including investor Michael Burry, argue that with Nvidia refreshing chip designs every 18 to 24 months, six-year schedules may flatter earnings.

When does it pay off?

That is the unresolved question, and views diverge. Goldman Sachs's Jim Covello has argued the economics look more questionable than two years ago, with returns still thin. A widely cited MIT report found about 95% of enterprise generative-AI pilots showed no measurable profit impact, though that finding has been contested. Bulls counter that cloud-revenue growth at Amazon and Microsoft is reaccelerating and that the build-out is still early.

As the original analysis put it, the biggest AI names "are no longer trading solely on the promise of future revenue. They are trading on the cost of getting there." Which scenario prevails — durable returns or a costly overshoot — is not yet settled. This is analysis, not investment advice.