On June 30, 2026, Tobias Adrian, director of the International Monetary Fund's Monetary and Capital Markets Department, told Bloomberg that AI leverage is more worrying than AI valuations. The previous day, the Bank for International Settlements had released its Annual Economic Report naming the AI capex boom and its increasingly fragile financing structures as one of four interlocking pressure points requiring immediate policy attention. Two institutions that rarely agree on anything published the same conclusion in the same week: the risk in the AI trade is not that stocks are overvalued. It is that the entire buildout is being financed with borrowed money, and the debt will outlast the assets by decades.
The numbers are difficult to overstate. Amazon, Alphabet, Meta, Microsoft, and Oracle issued $159 billion in corporate bonds in the first five months of 2026, 47 percent more than the same period last year. From 2020 through 2024, these five companies averaged roughly $28 billion per year in bond issuance, or about $140 billion total over five years. They exceeded that in five months. The companies that defined an era by hoarding cash (Apple alone once held $261 billion, Microsoft kept $130 billion on hand) are now borrowing at rates that would make a utility company blush.
Amazon set the pace in March with a $54 billion global bond sale, split between approximately $37 billion in U.S. dollar tranches and a debut euro-denominated offering of roughly $17 billion. Investor demand for the U.S. portion alone reached $126 billion, more than three times the amount available. The proceeds are earmarked for what Amazon plans to be $200 billion in capital expenditure this year, nearly all of it directed at AI data centers, custom chips, and networking infrastructure.
Nvidia followed on June 15 with $25 billion in bonds, its first debt sale since 2021, back when it generated $27 billion in annual revenue. Revenue has since grown to $216 billion. The offering was seven tranches deep, with the longest maturing in 2056. Thirty years from now. Investor demand reached $85 billion. At the long end, the spread tightened from 90 basis points over Treasuries to 65. Thirty years. The company is financing the AI buildout with debt that will mature when today's GPUs have been scrap metal for a quarter century.
Oracle came in February with $25 billion of its own. But Oracle's credit rating is Baa2 from Moody's, BBB from S&P, just two notches above junk. In November 2025, Barclays downgraded Oracle's debt to underweight and warned the company could fall to BBB-minus, the lowest investment-grade rung before the fall. Oracle is financing about 50 percent of its $50 billion 2026 capex plan through debt issuance. Meanwhile, Microsoft carries a AAA rating, Alphabet AA-plus, and Amazon and Meta both carry AA-minus. The same trade is running through five companies at five very different credit qualities, and the bond market is treating them as one bet.
Combined hyperscaler capital expenditure is projected above $600 billion for 2026: Amazon at $200 billion, Alphabet at $180 billion, Meta at $125 to $145 billion, Microsoft at approximately $120 billion, and Oracle at $50 billion. These figures exceed what the entire U.S. airline industry spent on aircraft in the prior decade. Capital intensity now runs between 45 and 57 percent of revenue, levels that historically belonged to oil majors and telecom operators, not software companies. At several of these firms, capital expenditure now exceeds internal cash generation, which is precisely why they went to the bond market.
The core problem is not the borrowing. It is the mismatch between how long the debt lasts and how long the assets last. GPUs, the central asset class in every AI data center, have an actual competitive life of one to two years at the frontier, according to Princeton's Center for Information Technology Policy. Companies depreciate them over five to six years. GPUs lose approximately 64 percent of their rental value within eighteen months. Two-thirds of Microsoft's capex in its most recent quarter went to what it classifies as short-lived assets: GPUs and CPUs that depreciate over three to five years.
When Nvidia issued a 30-year bond, it was selling investors a claim on cash flows stretching to 2056. The GPUs that bond finances will be replaced by 2029 at the latest. The data centers housing those GPUs may last 15 to 20 years, but the computing equipment inside them turns over three to four times within that envelope. The bond market is pricing AI infrastructure as if it were a bridge or a pipeline, assets that produce steady returns for decades. In reality, the assets behave more like commercial aircraft: high upfront cost, aggressive depreciation, and constant pressure to upgrade to the next generation.
Then there is the debt you cannot see. The BIS Quarterly Review published in March 2026 warned that hyperscalers are using off-balance-sheet financing structures to fund AI infrastructure without reporting it as debt. Tens of billions have flowed through these channels. The largest disclosed deal is Meta's $27 billion private credit joint venture with Blue Owl Capital for the Hyperion data center campus in Louisiana. Oracle has its own multibillion-dollar project-finance arrangements for AI facilities in Texas and Wisconsin. These structures amount to what the BIS calls shadow borrowing: obligations that are economically identical to debt but live outside corporate balance sheets.
The BIS warned that guarantees embedded in these structures could activate unexpectedly. If private credit flows shift in a procyclical manner, with lenders pulling back precisely when conditions deteriorate, the off-balance-sheet obligations snap onto the balance sheet at the worst possible moment. This is not a theoretical concern. It is the mechanism that turned the 2008 housing crisis from a real estate correction into a financial crisis: structured vehicles that were nominally separate from bank balance sheets came home when the market turned.
CNBC reported in February 2026 that Big Tech's AI bond issuance has shattered what it called an unspoken contract with investors. For two decades, the investment case for owning these stocks rested partly on their fortress balance sheets. Apple, Microsoft, and Alphabet generated so much cash that they returned hundreds of billions to shareholders through buybacks while still accumulating reserves. Investors accepted lower dividend yields because these companies did not need external capital. That contract is over. The five largest technology companies are now among the most active issuers in the investment-grade bond market, and they are borrowing not because they want to optimize their capital structure but because they have to. The AI buildout costs more than even their cash flows can cover.
The closest historical analogy is not the dot-com bubble. It is the railroad boom of the 1870s, when the most advanced technology companies of their era built transformative infrastructure financed primarily through bonds, not equity. Railroads consumed 60 percent of all securities listed on the New York Stock Exchange. When revenues failed to cover debt service, the crash came not through falling stock prices but through bond defaults that cascaded into bank failures. Jay Cooke and Company, the most prestigious bank on Wall Street, failed in September 1873 because it could not sell Northern Pacific Railway bonds. The stock market closed for ten days. The depression lasted six years.
The AI buildout may prove more profitable than the railroad buildout. It may not. But the financing structure is rhyming. The debt is long. The assets are short. The cash flows are projected. The lenders are enthusiastic. And the two institutions whose job is to worry about financial stability, the IMF and the BIS, published their warnings the same week, the way doctors nod at each other across a patient's chart.
According to J.P. Morgan, sixty percent of the AI data center capacity planned for completion by 2027 has not yet broken ground. The borrowing to finance it has already happened. The duration mismatch is not a risk that might emerge. It is a risk that has already been written into the bond market, at thirty-year maturities, at spreads so tight they imply the market believes these companies will generate sufficient cash to service this debt through the year 2056. The AI boom's real risk was never the stock price. It was always the duration.