Who is actually paying for the AI data center boom

The bill is now bigger than the cash flow

For most of the past decade, the largest US technology companies paid for their growth out of pocket. That is changing. Microsoft, Amazon, Alphabet, and Meta have guided to roughly $720 billion to $745 billion in combined capital spending for 2026, according to analyst tallies of their earnings guidance. That is up from about $410 billion in 2025 and roughly $226 billion in 2024. In two years the number has more than tripled, and almost all of the increase is tied to artificial intelligence: graphics processors, custom chips, and the data centers and power to run them.

The scale is easier to grasp company by company. Amazon's capital spending reached about $131.8 billion in 2025, up from $83 billion a year earlier, and CEO Andy Jassy told investors the company plans about $200 billion in 2026, "predominantly in AWS, because we have very high demand." Microsoft said it would spend around $80 billion building AI-capable data centers in its fiscal 2025 and has signaled continued growth. Alphabet raised its 2026 guidance toward $195 billion to $205 billion during the year. Meta guided to $115 billion to $135 billion for 2026 and later lifted the top of that range.

Where the money is going

The spending is concentrated in a narrow set of inputs. Alphabet told investors that in one 2026 quarter roughly 60 percent of its technical infrastructure spend went to servers and the other 40 percent to data centers and networking gear. Amazon says the majority of its outlay is for AWS, and most of that is AI. The companies describe the demand as real and immediate. Jassy's line, "we are monetizing capacity as fast as we can install it," has become the standard defense of the numbers.

Demand does appear to be growing. Google Cloud's revenue rose 82 percent year over year in one 2026 quarter, and the company reported a cloud backlog above $500 billion. The open question is not whether customers are buying, but whether the revenue will grow fast enough, and at a high enough margin, to cover what the infrastructure costs to build and, later, to replace as the chips age.

The new habit: borrowing for it

The clearest sign that the math has changed is where the cash is coming from. FactSet reported in 2026 that free cash flow is expected to fall close to zero or turn negative in fiscal 2026 for every one of the big spenders except Alphabet and Microsoft. To fill the gap, the companies have turned to the bond market at a pace they never needed before.

In 2025 the five largest hyperscalers, Amazon, Alphabet, Meta, Microsoft, and Oracle, issued about $121 billion in US corporate bonds, against an average of roughly $28 billion a year between 2020 and 2024. Meta's $30 billion offering in October 2025 was the largest corporate bond sale in about two years, and Alphabet followed with $25 billion in November. By FactSet's measure, debt rose from about 9 percent of these companies' capital spending in fiscal 2024 to about 32 percent in the year through mid-2026. Goldman Sachs has projected that Big Tech will fund more than a third of its AI investment with debt in 2027.

The off-balance-sheet turn

Debt sold in a company's own name at least shows up on its balance sheet. The more striking development is the money that does not. In October 2025 Meta arranged about $30 billion for its Hyperion data center campus in Louisiana through a special purpose vehicle it owns only about 20 percent of, with Blue Owl Capital holding the rest. The vehicle raised roughly $27 billion in senior secured notes due 2049, owns the facility, and leases it back to Meta. The structure lets Meta control the site while keeping most of the borrowing off its own books.

It is not cost-free. Press reports and Meta's own filings describe a residual value guarantee of up to about $28 billion, a promise to compensate investors if the data center is worth less than a set threshold, disclosed in the footnotes rather than recorded as a liability. Lawyers who follow the sector have flagged these arrangements as a growing source of hidden risk, and analysts estimate off-balance-sheet lease commitments across the five companies now run into the hundreds of billions of dollars.

The question the banks keep asking

With the spending now funded partly by borrowing, the payback question has sharpened. Goldman Sachs estimated in September 2026 that the hyperscalers need to generate roughly $300 billion in annual AI revenue to break even on what it projects as about $800 billion of AI infrastructure spending in 2026, counting Oracle alongside the big four. By Goldman's reading, AI cloud revenue was running only about $70 billion above its pre-AI trend, leaving a gap of roughly $230 billion a year still to be closed. The bank's strategists called the investment cycle large relative to the economy, comparing it to the railroad buildout of the late 1800s.

Consultants put the longer-run bar higher. Bain & Company estimated in September 2026 that the industry will need about $6 trillion in annual revenue by 2031 to justify the data centers its own demand forecasts imply. Existing AI services, Bain said, would supply only part of that, leaving roughly $4 trillion in new revenue that does not yet exist, from products still being built. None of this is proof the bet will fail. It is a measure of how much has to go right.

What to watch

The useful signals from here are concrete. Watch whether free cash flow keeps shrinking or recovers, whether cloud revenue growth holds the pace the backlogs imply, and whether more capacity moves into off-balance-sheet vehicles, which make the true leverage harder to see. The companies insist demand justifies the spend, and their revenue is in fact growing. What has not yet been demonstrated is the return. For now, a rising share of the AI buildout is being paid for with borrowed money against revenue that is still, in the banks' own arithmetic, mostly a forecast.