The extraordinary investment required to build the infrastructure behind artificial intelligence is increasingly becoming a credit story. Data centres, power generation, specialised chips, cloud infrastructure and the networks connecting them require capital on a scale that can no longer be comfortably absorbed by the cash flows and balance sheets of the technology companies driving the boom. Banks, private credit managers, bond investors, infrastructure funds and special-purpose financing vehicles are consequently becoming more important sources of capital. That expansion does not mean an AI credit crisis is inevitable, nor does it make every loan connected to AI inherently risky. It does, however, create a more complicated concentration problem than conventional lending metrics may reveal. A lender financing a data centre, another backing a power project and a third providing capital against GPU infrastructure may appear to hold exposures to different borrowers and even different industries. Economically, however, all three investments could depend on the same underlying assumption: that a relatively small group of technology companies will continue spending enormous amounts on AI and that demand for computing capacity will grow sufficiently to justify what is being built. As AI financing spreads across credit markets, understanding that shared dependency may become as important as analysing the individual borrower.
AI infrastructure is moving beyond Big Tech’s balance sheets
The first phase of the generative AI investment cycle was supported largely by companies with exceptionally strong balance sheets. Amazon, Alphabet, Meta and Microsoft could deploy tens of billions of dollars into chips and data centres from their own resources, making the AI infrastructure boom appear principally to be a question of corporate capital expenditure.
The scale is changing that equation. J.P. Morgan estimates hyperscaler capital expenditure will reach $697 billion in 2026 alone, while research from Columbia Business School describes a financing ecosystem that increasingly includes third-party developers, real estate investment trusts, private credit and structured finance. The physical requirements are equally broad: AI infrastructure extends beyond the data centre building itself to power generation and transmission, cooling systems, networking equipment and the specialised processors providing the underlying compute.
As those capital requirements expand, external financing becomes increasingly important. Research published by Carlyle argues that the AI compute buildout could require roughly $1 trillion from private credit, an amount equivalent to more than half of the industry’s current assets under management. The structures being developed to provide that capital range from construction and power financing to GPU-backed asset financing, off-balance-sheet special-purpose vehicles and senior secured lending to emerging cloud providers.
This is an important transition. The financial risk attached to AI infrastructure is no longer contained principally within the companies making the investment. It is gradually being distributed through the wider credit system.
The concentration may be hidden by the borrower
Traditional credit concentration is relatively straightforward to conceptualise. A bank measures how much it has lent to a particular company, sector, geography or type of borrower and establishes limits designed to prevent a single exposure from becoming disproportionately important.
AI infrastructure complicates that approach because apparently different credits can share the same economic dependency.
A data-centre developer may derive much of its future revenue from a hyperscaler tenant. A power project may have been developed specifically to serve rapidly expanding data-centre demand. A financing vehicle might be secured against GPUs whose economic value depends on demand for AI compute and the pace at which newer processors make existing hardware less competitive. A specialised cloud provider may rely heavily on a small number of AI customers. Meanwhile, private equity and infrastructure investors may own businesses whose valuations depend on precisely the same investment cycle.
None of these exposures necessarily appears identical in a portfolio.
Yet they may all deteriorate for the same reason.
This is why the conventional question, “How much AI lending do we have?”, may eventually prove insufficient. Financial institutions may need to ask a more difficult question: how much of our portfolio ultimately depends on continued AI infrastructure spending?
The distinction matters because concentration is not simply about lending repeatedly to the same borrower. It can also arise when different borrowers depend on the same economic engine.
The SaaS lending cycle offers an uncomfortable lesson
Carlyle’s warning is particularly interesting because it draws on private credit’s experience financing software companies during the software-as-a-service boom.
Software appeared unusually attractive to lenders. Recurring subscriptions created predictable revenues, margins could be high and individual companies seemed to serve sufficiently different markets that their risks appeared relatively uncorrelated. Private equity sponsors also saw substantial upside, creating a steady pipeline of acquisitions requiring debt financing.
The problem was that lenders were not merely financing this market. Their willingness to lend helped shape it.
When credit became readily available and underwriting criteria widened, private equity sponsors could support higher acquisition valuations and pursue transactions that might not have worked under more restrictive financing conditions. Carlyle argues that underwriting guidelines therefore influence which deals sponsors pursue rather than simply responding passively to whatever transactions arrive.
Generative AI subsequently introduced another weakness into the software credit thesis: technological disruption could affect large numbers of supposedly independent software businesses at the same time. Risks that looked idiosyncratic turned out to contain a common factor.
That lesson is particularly relevant to AI infrastructure because the potential financing requirement is considerably larger.
Credit can do more than fund an investment cycle. Abundant credit can accelerate it.
If lenders become increasingly comfortable financing data centres, GPU clusters and associated infrastructure, developers gain access to more capital. Greater financing capacity enables more projects to proceed, supports asset valuations and attracts further equity investment. Those valuations can, in turn, support additional borrowing.
The result can become a reinforcing cycle in which optimism creates financing capacity and financing capacity helps turn optimism into physical investment.
AI risk is spreading through several layers of finance
The structure of the AI buildout makes the ultimate location of risk increasingly difficult to identify. Financing can sit on corporate balance sheets, inside joint ventures, at individual data-centre projects or within special-purpose vehicles. Debt can subsequently migrate from banks into bond markets or private portfolios.
That movement is already visible. Banks led by Morgan Stanley have been preparing to refinance $15 billion of debt associated with a Google-backed data-centre project in Texas through the bond market, allowing the lenders to reduce their balance-sheet exposure and create capacity for further AI infrastructure financing.
Off-balance-sheet structures add another layer. Reuters reported in September that Morgan Stanley estimates roughly $3 trillion of financing and leasing arrangements connected to Nvidia, Broadcom and major hyperscalers are being kept outside conventional corporate balance sheets. Meta’s Hyperion data-centre development in Louisiana illustrates how these structures can work: financing for the project includes bonds issued through a separate vehicle rather than simply appearing as conventional Meta corporate debt.
Risk has not disappeared in these structures. It has changed location.
That distinction is critical for banks and regulators. Transferring a loan into the bond market, placing assets inside a financing vehicle or bringing private capital into a project can reduce direct exposure for an individual institution. Across the financial system, however, the underlying economic exposure remains.
Even diversification within AI may not be diversification
One of the more difficult questions for credit committees will therefore be deciding what actually constitutes diversification.
A portfolio containing loans to data centres in several states, multiple cloud companies, different power projects and several GPU financing vehicles may appear diversified when assessed through conventional sector or borrower classifications.
But suppose those businesses ultimately depend on demand from the same four or five hyperscalers.
Or suppose their economics assume that AI model development will continue requiring exponentially greater amounts of compute.
Or that data-centre utilisation remains high enough to support lease payments and debt service.
Or that today’s expensive GPUs retain sufficient economic value before the next generation of hardware arrives.
In each case, apparently distinct credits contain a common underlying factor.
There are already signs that credit investors are thinking this way. Reuters reported that some institutional investors are approaching single-name exposure limits after aggregating bonds issued by technology companies with related data-centre financing structures. Investors have also demanded wider spreads from AI-related issuers even where the underlying companies retain strong balance sheets, partly because markets are being asked to absorb such large volumes of debt.
The implication is important: AI credit concentration may ultimately need to be measured by dependency rather than borrower.
The collateral creates another challenge
Credit risk becomes even more complicated when financing moves from buildings and contracted cash flows towards the computing hardware itself.
GPU-backed lending is attractive for obvious reasons. Advanced processors are extraordinarily valuable assets and demand for computing capacity remains intense. Turning those assets into financeable collateral could unlock another large source of capital for AI infrastructure.
But GPUs are not conventional long-lived infrastructure.
Aircraft, property and industrial equipment have decades of historical information allowing lenders to estimate useful lives, residual values and recovery rates. AI processors exist within a technology cycle where a new generation of hardware can materially change the economics of the previous one.
That makes the value of the collateral partly dependent on technological progress.
The issue is already generating debate between technology companies and financiers. Reuters reported this week that lenders considering chip-backed financing are taking a more conservative view of GPU values than Nvidia, with some banks using depreciation assumptions of roughly three to four years and seeking stronger guarantees or revenue commitments from investment-grade customers.
This does not mean GPUs cannot become a credible financing asset class. It means lenders are being asked to price residual-value and technological-obsolescence risks for which there is comparatively little historical evidence.
A slowdown does not need to produce defaults to matter
It is tempting to frame concerns around AI financing as a binary question: either the AI boom continues or an investment bubble bursts.
Credit markets rarely require such dramatic outcomes for problems to emerge.
Suppose AI adoption continues growing, but more slowly than today’s infrastructure plans assume. Data-centre utilisation could take longer to reach projected levels. Hyperscalers could moderate future capital expenditure. Power projects designed around aggressive demand forecasts could face delays. GPU residual values could fall faster than lenders expected.
Borrowers might remain solvent while the economics of new financing deteriorate.
Lenders would respond by demanding higher spreads, lowering loan-to-value ratios, requiring stronger guarantees or simply declining projects that would previously have received funding. Refinancing could become more expensive. Private asset valuations could adjust more slowly than public credit, but they would eventually have to reflect the changed environment.
There are already signs that markets are becoming more discriminating. KKR recently warned that the AI investment cycle is creating unusually concentrated exposure in credit markets and estimated that around one-fifth of the investment-grade bond index could eventually carry AI-related risk. Meanwhile, corporate bond investors have begun demanding greater concessions from AI-related issuers as the volume of debt hitting the market expands.
The danger, therefore, does not require a collapse in artificial intelligence. A meaningful revision to expectations could be enough to expose where leverage and concentration have accumulated.
The real question is where the risk ultimately sits
None of this makes the AI infrastructure boom inherently unsustainable. Artificial intelligence could produce substantial productivity gains, demand for compute could continue expanding and many of the data centres and power assets being constructed today may generate reliable cash flows for years.
The financial question is different.
As an investment cycle grows large enough to require trillions of dollars of external capital, the way that capital is structured begins to matter almost as much as the technology being financed. Research from Columbia Business School argues that the shift towards outside financing has increased leverage and redistributed AI infrastructure risk across banks, private lenders, real estate investors and specialised financing structures. The concern is not that financial distress is inevitable, but that uncertain demand, technological change and leverage create meaningful downside if expectations are revised.
For banks and private credit managers, this makes AI a portfolio-management problem as much as an underwriting opportunity.
The challenge will be recognising when apparently different loans are ultimately expressions of the same bet.
What this means for the industry
- AI exposure needs a broader definition. Measuring direct loans to technology companies may substantially understate exposure when data centres, power projects, GPU financing and infrastructure vehicles ultimately depend on the same AI investment cycle.
- Concentration should increasingly be measured by economic dependency. Different borrowers and asset classes can still share the same underlying source of repayment risk.
- Credit availability can amplify the investment cycle. Lending standards do not merely respond to AI infrastructure demand. Greater financing capacity can enable more projects and support higher valuations.
- Banks can transfer exposure without eliminating system-wide risk. Moving AI infrastructure loans into bonds, private credit or special-purpose vehicles changes who holds the risk rather than making it disappear.
- GPU financing introduces an unfamiliar collateral problem. Lenders must assess technological obsolescence and residual values without the long historical datasets available for traditional infrastructure assets.
- The stress scenario does not require an AI collapse. Slower demand growth, lower utilisation or reduced hyperscaler spending could tighten financing conditions and expose concentration long before widespread defaults appear.
