Who Is Actually Paying for the AI Boom?

Amazon's free cash flow over the last twelve months was negative $42.4 billion (see Sources, [6]). In the latest quarter, Alphabet and Meta went negative too (see Sources, [1], [3]). Some of the richest companies on Earth are now spending more on the AI buildout than their businesses bring in, so someone else's money has to fill the gap.

That money comes in four layers: bond investors, private lenders behind off-balance-sheet deals, shareholders and chip suppliers, and finally the customers who are meant to pay it all back. Here is who is in each layer, and how much of the risk has reached ordinary savers.

The bill

Amazon expects to spend $220 billion on capital projects in 2026, and to spend meaningfully more in 2027 (see Sources, [6]). Alphabet guides to $195 billion to $205 billion for 2026 and expects a significant increase in 2027 (see Sources, [2]). Meta guides to $135 billion to $145 billion (see Sources, [3]).

Free cash flow is what remains after a company pays for that spending, and the latest quarter shows how far the bill has run ahead of income:

Alphabet took in $42.29 billion from operations and spent $51.355 billion on property and equipment (see Sources, [1]). Meta took in $32.79 billion and spent $34.91 billion, including finance lease principal (see Sources, [3]). Oracle's capital spending in the quarter was $28.5 billion, up from $8.5 billion a year earlier (see Sources, [8]).

This is not everyone, and it is not every period. Microsoft stayed positive (see Sources, [7]). Over a full year, Alphabet is still positive at $19.747 billion, and it holds $265.098 billion in cash and marketable securities (see Sources, [1]). But the companies themselves say spending is heading up, not down.

Layer one: the bond market

The first source of outside money is lenders. By early October 2026, five companies (Amazon, Alphabet, Meta, Microsoft and Oracle) had issued $218.1 billion of investment-grade bonds, against $107.3 billion in all of 2025 (see Sources, [9]). Meta alone sold $25 billion of senior unsecured notes in May (see Sources, [4]).

These borrowers now make up 4.7% of the FTSE US BIG Corporate index by market value, and about $500 billion of their debt was outstanding at the end of August (see Sources, [9]). That matters because an index fund does not choose whether to lend. As a borrower's weight grows, the fund holds more of its bonds.

The European Central Bank says passive, benchmark-tracking investors may mechanically rebalance toward these companies, and it reports strong demand from pension funds and insurers for their euro-denominated bonds (see Sources, [11]). So anyone holding a broad bond fund may already be lending to the buildout.

Vanguard says the compensation for that concentration is thin, and that the channels could correlate if the buildout disappoints. It also says plainly that nothing here is a solvency warning (see Sources, [10]). These are strong borrowers. The concern is how much of one bet sits in a broad fund, not whether these companies can pay.

Layer two: off the balance sheet

The second layer is harder to see. Hyperion, Meta's data center campus in Louisiana, is owned 80% by funds managed by Blue Owl and 20% by Meta, with about $27 billion in development costs. Part of Blue Owl's capital comes from debt issued to PIMCO and other bond investors (see Sources, [5]).

Meta rents the campus on leases with a four-year initial term and options to extend (see Sources, [5]). Lenders accept that short term on a project of this size because Meta also provides a residual value guarantee: if it leaves and the conditions are met, it makes a capped cash payment to the venture, for the first 16 years of operations (see Sources, [5]).

Across its two unconsolidated data center ventures, Meta's maximum guarantee reached $41.493 billion at September 30, 2026, up from $28.091 billion at the end of 2025. Its total maximum exposure to loss is $68.404 billion. Meta says guarantee payments are not probable, so no liability has been recorded (see Sources, [4]).

Then there are leases that have been signed but have not started. Meta puts those at about $355.73 billion, commencing through 2036 with terms of up to 30 years (see Sources, [4]). By our arithmetic, that is more than twice the top of its 2026 capital spending guidance.

Layer three: shareholders and suppliers

The third layer is the owners. Meta bought back $3.33 billion of its stock in Q3 2025, and none in Q3 2026 (see Sources, [3]).

Then it went the other way. In October, Meta raised its equity plan to $97.5 billion, its largest ever: $25 billion of Class A stock, $12.5 billion of mandatory convertible preferred shares, and $60 billion planned through an at-the-market program in non-voting Class C shares (see Sources, [12]).

Nvidia said it would invest $3 billion in that offering and AMD $1 billion (see Sources, [13]). Both sell AI chips, so suppliers are now helping to fund a buyer of AI chips. Nvidia also put $30 billion into OpenAI in late February 2026 (see Sources, [14]). Whether that is smart strategy or circular financing, the suppliers are now part of the funding.

Layer four: the customer

All of this works only if customers pay, and right now the revenue is real. Google Cloud grew 103% to $30.7 billion in the quarter (see Sources, [1]). Azure grew 48% (see Sources, [7]). AWS grew 39% to $45.98 billion (see Sources, [6]).

Contracts are larger still. Google Cloud reports a backlog of $561 billion, with just over 55% expected to become revenue within 24 months (see Sources, [2]). Oracle reports $664 billion of remaining performance obligations (see Sources, [8]).

But a backlog is a promise, and one customer sits behind many of these promises. Sacra lists about $300 billion of OpenAI commitments to Oracle, $250 billion to Azure and $138 billion to AWS (see Sources, [15]).

Sacra estimates OpenAI's annualized revenue at $50 billion in September 2026, and it cites a Financial Times report of a projected $278 billion in cumulative negative free cash flow over 2026 to 2030 (see Sources, [15]). So part of the money meant to repay this buildout comes from a company that still has to raise money itself. These are third-party estimates, not figures from OpenAI's accounts.

So who is actually paying?

Today, it is lenders, shareholders and index funds. Tomorrow, it is whoever buys the AI services.

If demand arrives, shareholders win and lenders collect their fixed payments. If it falls short, the loss runs through the same chain, and part of that chain is ordinary savings. None of this says the boom will fail. It says the risk has moved out of a few company budgets and into the wider financial system.

The same financing pattern shows up one level down: The Billion-Dollar Business of Renting Nvidia GPUs looks at how GPU clouds fund themselves, and How Data Centers Actually Make Money covers the leases on the other side of these deals. For what all that money buys, see Where the $60 Billion Goes in One AI Data Center.