The AI boom is is becoming a major source of fixed-income supply — and the bond market may not be fully prepared for what comes next.
For most of the past decade, the AI boom was financed primarily with equity and operating cash flow.
Microsoft, Amazon, Alphabet, Meta, Oracle and Nvidia are increasingly turning to debt markets to finance the enormous capital spending required to build AI infrastructure. At the same time, data-center developers are raising capital through project finance, private credit, ABS and other structured vehicles.
Hyperscaler debt issuance, which averaged roughly $35 billion a year between 2020 and 2024, jumped to around $93 billion in 2025. By mid-2026, the five hyperscalers had already issued roughly $132 billion. Broader estimates for AI-related debt issuance in 2026 range from roughly $300 billion to more than $500 billion, depending on what is included.
J.P. Morgan estimates that hyperscaler issuance alone could reach around $279 billion this year. The Dallas Fed estimates that this could translate into as much as $360 billion of new 10-year-equivalent duration supply — roughly one-eighth of the duration supply generated by U.S. Treasury issuance.
That is a remarkable change in a very short period of time. And I think the important question is no longer whether these companies can borrow. It is whether the bond market can absorb everything that needs to be financed.
Hyperscalers can issue investment-grade bonds. Data-center developers can use project finance. Private credit funds can finance construction. ABS markets can finance contracted infrastructure. Different structures. Different risks. But ultimately, much of it competes for the same pool of institutional capital. That is where this becomes a fixed-income story.
The crowding-out problem
Think about a pension fund deciding where to put its next $500 million.
It can buy Treasuries.
It can buy a Microsoft bond.
It can buy an investment-grade utility.
It can buy a data-center financing.
Or it can allocate to private credit.
The underlying capital is not infinite. And the supply is arriving from several directions at once. First, there is the competition between sovereign and corporate duration. A Treasury does not have to offer a credit spread. A Microsoft or Alphabet bond does.
But if the supply of highly rated technology debt becomes large enough, investors may demand more compensation to absorb it. That can affect the relative pricing of other investment-grade issuers, and potentially Treasuries themselves.
We are already seeing signs of this. U.S. technology companies have become increasingly active in the euro bond market, where they now account for close to 10% of gross new corporate issuance outside the financial sector. The ECB has warned that the growing presence of U.S. technology issuers could crowd out other borrowers and put upward pressure on financing costs.
Second, there is competition within investment-grade credit. Microsoft, Amazon, Alphabet and Meta can absorb enormous amounts of investor demand because they are among the highest-quality corporate borrowers in the world.
But the investor’s allocation is finite. If the hyperscalers issue another $20 billion, somebody else may have to pay a little more to get their deal done. That is how crowding-out works in practice. It doesn’t require anyone to be rejected. It simply means the price of capital moves. Third, there is the competition further down the credit spectrum.
Data centers are increasingly being financed through project finance, private credit and structured products. These transactions can offer higher yields than traditional investment-grade bonds, but they also carry risks that a Microsoft bond simply does not.
It’s not just more debt. It’s different debt.
A 10-year Microsoft bond is primarily a bet on Microsoft’s ability and willingness to service its debt. A data-center project-finance loan is something else entirely. It is a bet on construction being completed on time. It is a bet on power being delivered. It is a bet on the tenant remaining solvent and continuing to pay. It is a bet on utilization. It is a bet on the residual value of a highly specialized asset. And it is a bet that the assumptions made when the project was financed will still look reasonable several years later.
Construction risk.
Completion risk.
Cost-overrun risk.
Tenant concentration.
Power availability.
Technology obsolescence.
These risks do not disappear simply because the ultimate customer happens to be a highly rated technology company. This is where I think investors need to be careful. There is a tendency to treat anything connected to Microsoft, Amazon, Google or Nvidia as if it carries the credit characteristics of the hyperscaler itself. It doesn’t.
A Microsoft bond is not the same thing as a loan secured by a data center whose economics depend on a single tenant, a particular power connection and a construction schedule. The credit risk sits in a different place. And I suspect some of that distinction is currently being priced too generously.
I’ve seen this movie before
I have seen capital rush into sectors where the underlying technology was real, the demand was real and the projects were real, but where the returns ultimately failed to justify the amount of capital deployed.
Telecom around 1999–2001 is the obvious example. The internet was real. The bandwidth demand was real. The technology worked. The problem was that too much infrastructure was financed at the same time. Shale provides another example. The technology worked. Production increased dramatically.
But for periods of time, the economics were poor because capital flooded into the sector and returns failed to compensate investors for the risks they were taking. AI infrastructure could be different. But the financing dynamic looks familiar.
My base case
I see three possible outcomes. My base case, with roughly a 50% probability, is orderly absorption. The market ultimately absorbs the additional supply, but not without demanding a higher price for doing so. Yields remain structurally higher, investment-grade spreads drift wider and new-issue concessions become more common as issuers compete for investor attention. The result would not be a disorderly market or a credit event; it would simply be a gradual increase in the cost of capital across the market.
I assign roughly 35% to selective indigestion. In this scenario, investors become increasingly discriminating about the structures financing the data-center buildout. The best hyperscaler bonds continue to attract strong demand, while weaker projects are forced to offer wider spreads, stronger covenants or greater equity cushions. Project finance and data-center ABS could therefore experience significant spread widening even while the investment-grade bonds issued directly by the hyperscalers remain relatively well bid.
My final scenario, with roughly a 15% probability, is broad crowding-out. AI-related borrowing becomes large enough that incremental investor demand simply cannot keep pace with the supply coming to market. Treasury auctions begin to show signs of weaker demand, investment-grade and high-yield spreads widen materially, and lower-rated borrowers face significantly higher refinancing costs. At the same time, weaker data-center projects may find that they can no longer raise debt on the terms assumed in their original financing models. Importantly, none of these scenarios requires an AI crash. The adjustment can happen simply because the supply of credit grows faster than the pool of capital available to absorb it.
What I am watching
I am watching the Treasury market first, particularly auction bid-to-cover ratios, dealer takedown and the behaviour of long-dated yields and the term premium. I am then looking at the corporate market: hyperscaler issuance calendars, the concessions required to clear new deals and the 10-year spreads of Microsoft, Alphabet and Amazon relative to Treasuries. Increasingly, however, I am also focused on the financing structures sitting behind the data-center buildout, including project-finance spreads, data-center ABS, private-credit underwriting, construction financing, covenant packages, tenant concentration and completion schedules. This is where I expect the more interesting credit risks to emerge.
There are already signs that investors are becoming more demanding. Reuters reported in August that AI-related debt issuance had reached around $220 billion in 2026, while technology bond spreads had widened relative to the broader investment-grade market. Amazon’s $25 billion bond deal, for example, priced at materially wider spreads than comparable issuance a year earlier. That is not a crisis. It is something more interesting: evidence that the market is beginning to ask for greater compensation for the amount of capital it is being asked to provide.
The downside is not a crash. It is a grind.
My downside case is not 2008. It is a slow and persistent increase in the cost of capital, which may ultimately be more important for markets precisely because it does not require a dramatic event. If hundreds of billions of dollars of additional long-duration debt continue to come to market, investors will eventually demand a higher return for absorbing that supply. The adjustment could take the form of wider credit spreads, higher Treasury yields, larger new-issue concessions, more selective underwriting, higher equity requirements or simply fewer projects getting financed.
For credit investors, that creates two very different sets of opportunities. At the top of the market, hyperscaler bonds may remain among the highest-quality corporate credits available. Further down the capital structure, however, the opportunity may lie in being properly compensated for risks that are currently being treated as secondary: construction risk, completion risk, refinancing risk and tenant concentration. Ultimately, investors also need to consider whether the AI infrastructure being built today will generate the returns assumed in the financing models.
The AI buildout is real. The demand is real, and the technology is real. None of that guarantees, however, that every dollar of debt financing will generate an adequate return. That is the part of the story I think the equity market is still largely ignoring. The AI boom is increasingly becoming a fixed-income story, and the most important consequence may not be a wave of defaults, but a prolonged repricing of the cost of capital.
Alessandro Rovelli
Rovelli Research
Independent credit investment research
