Banks helped create the AI debt boom. By May, some were already asking other investors to take the riskiest pieces.
A syndicate including JPMorgan and MUFG had spent more than six months distributing $38 billion in construction loans for data centers in Texas and Wisconsin leased to Oracle. Some lenders tried to sell pieces at a discount. JPMorgan, Morgan Stanley and SMBC were also exploring private loan sales and significant risk transfers as their exposure pressed against internal limits. The Economic Times’ account of the Financial Times reporting quotes Man Group credit specialist Matthew Moniot describing the deal sizes as “out of scale to anything we’ve thought about, ever.” Startup Fortune brought the financing story back into view this week.
The banks’ behavior is the warning. They helped finance the buildout and now want less of its risk sitting on their own books. The AI boom is being sold as a self funding technological revolution while its financing risk is sliced up and pushed outward.
Oracle shows where cash stops being enough
Oracle has real demand. Its fiscal 2026 filing reported $638 billion in remaining performance obligations, up from $138 billion a year earlier. That number is not cash in the bank. Oracle expects to recognize only about 12 percent of it as revenue in the following twelve months, with most of the rest arriving over several years.
The same filing shows the financing gap. Oracle generated $32.0 billion in operating cash during fiscal 2026 and spent $55.7 billion on capital expenditures, leaving free cash flow negative by $23.7 billion. It issued $42.7 billion in senior notes and raised another $5 billion through mandatory convertible preferred stock. These are Oracle’s figures.
Oracle is one case in a much larger borrowing wave. Amazon, Alphabet, Meta, Microsoft and Oracle issued $121 billion in United States corporate bonds in 2025, compared with a $28 billion annual average from 2020 through 2024. Amazon, Alphabet, Meta and Microsoft later raised their combined 2026 capital spending plans to as much as $725 billion.
Epoch AI compared SEC filings from those four companies and Oracle. Its model found aggregate operating cash flow growing about 23 percent a year while cash capital spending grew about 70 percent. The lines cross around the third quarter of 2026, although changing the starting window moves the crossover between the second and fourth quarters. The model is a simple extrapolation with a wide margin. Even that qualified result shows why the industry has become so dependent on debt markets staying open.
Risk moves outward
A significant risk transfer lets a bank keep loans while transferring a defined layer of credit losses to investors. The bank can then receive regulatory capital relief. A Bank for International Settlements review says pension funds, sovereign wealth funds and endowments are the main indirect investors through specialized credit funds and diversified asset managers. The review also says public disclosures in the United States are too thin to estimate each bank’s use of the structures.
Risk transfer is literal. The risk leaves one balance sheet and appears on another. If dozens of banks are near their own concentration limits, distributing the exposure may protect each bank while making the system’s total exposure harder to see.
Meta’s El Paso project shows how quickly this debt travels. Meta and BlackRock created a venture in which BlackRock managed funds own 80 percent and Meta owns 20 percent. Meta’s own announcement says the campus will cost about $14 billion, with $12.5 billion financed through debt. By July, SEC portfolio filings showed Sopaipilla Investor bonds inside the Western Asset Income Fund and the Voya Global Bond Fund.
Within days of issuance, exposure to one Meta data center had moved from a project venture into registered bond funds. Someone can finance an AI infrastructure bet by buying a general income or global bond fund without choosing Meta’s data center economics directly.
The debt has real support
There are serious facts behind the industry’s confidence. Oracle reported $67.4 billion in fiscal 2026 revenue and $17.0 billion in net income. It also says $75 billion of hardware tied to large AI contracts was prepaid or supplied by customers, reducing the capital it must raise. Meta retained a 20 percent stake in the El Paso venture and provided a residual value guarantee with an initial threshold of about $13 billion. Meta still has billions at risk in the campus it will occupy.
Investors also bought the bonds. Sopaipilla Investor placed $12.547 billion of notes, and they traded above par in the early secondary market. Demand was real.
The comparison with 2008 has strict limits. Today’s assets and borrowers are different, and adjustable rate home loans are absent. The structure is the part worth remembering. Federal Reserve research found that risks thought to be separated from banks through off balance sheet vehicles returned to their sponsors under stress. A new legal wrapper can change who reports a risk. It cannot prevent many holders from rushing to sell at once.
Make the risk visible before it breaks
Regulators should map who holds the losses before any downturn forces the answer into public view. Banks using risk transfers for AI loans should disclose the referenced assets, retained exposure, counterparties and conditions that return risk to the originator. Funds holding project bonds should identify the tenant, lease assumptions, guarantees and concentration behind the security instead of burying it under an infrastructure label. Supervisors should stress the banks and the nonbank buyers together.
The industry’s current arrangement is convenient for everyone at the top. Oracle and Meta get capacity. Banks keep arranging loans without keeping every dollar of exposure. Asset managers get new yield to sell. The public gets a story about unstoppable progress with the financing details removed.
Banks helped create an AI debt market too large for their own concentration limits, and now they are looking for more places to put it. “Diversification” changes whose statement carries the risk. The risk itself remains.