Why AI Debt Is Showing Up in Mortgage Rates

From the start of January to the middle of August, American companies sold a record $1.68 trillion of bonds, roughly 27 percent more than over the same stretch a year earlier. A large and fast-growing slice of that was raised to build artificial intelligence infrastructure: data centres, chips, power connections and the land they sit on. Bank of America puts AI-related issuance at about $220 billion so far this year, more than double last year's total. Goldman Sachs, using a wider definition that captures chipmakers, utilities and data-centre developers, puts the figure closer to $489 billion.

At the same time, the 30-year US Treasury yield climbed above 5.31 percent, its highest level since June 2007. A recent 30-year Treasury auction cleared at roughly 5.22 percent, the government's steepest borrowing cost at that maturity since 2001. The 10-year yield, which sets the tone for most household lending in America, has been trading near 4.70 to 4.75 percent.

The 30-year fixed mortgage averaged 6.65 percent in Freddie Mac's survey on 20 August, slightly higher than a year ago despite three Federal Reserve rate cuts in late 2025. Wall Street is now arguing, in public and with some heat, about how much of that is the AI buildout's fault.

In this article we explore what "duration supply" means, why long-dated data-centre financing competes with the US government for the same narrow pool of buyers, what the term premium is and why it is rising, the serious case that AI debt is only a secondary cause, and how all of it lands on a mortgage quote.

What Duration Supply Actually Means

A bond is a loan with a fixed schedule. The borrower pays interest for a set number of years and repays the principal at the end. The longer that schedule runs, the more the bond's price moves when market interest rates change. A 30-year bond loses far more value when yields rise one percentage point than a two-year bond does, because the investor is locked into the old rate for three decades rather than two years.

That sensitivity is called duration. It is measured in years, and it is best thought of as a unit of interest-rate risk rather than a unit of time.

When market participants talk about "duration supply", they mean the total quantity of interest-rate risk that the bond market is being asked to swallow in a given period. It is not the dollar amount of bonds issued. It is the dollar amount weighted by how long-dated those bonds are. Ten billion dollars of three-year notes barely registers. Ten billion of 30-year bonds is a meaningful ask.

The AI issuance wave is heavily skewed to the long end. Much of it carries maturities of 20 or 30 years, and one recent deal included a rare "century" bond maturing 100 years from issuance. So the duration supply being created is much larger than the headline dollar figures suggest.

The Buyers of Long-Dated Bonds Are a Narrow Club

There is no shortage of investors willing to buy short-dated paper. Money market funds, banks and corporate treasurers all need it. The market for 30-year bonds is different, and much smaller.

The natural owners of very long-dated debt are institutions with very long-dated obligations:

• Pension funds, which owe payments to retirees decades into the future and want assets whose cash flows arrive at roughly the same time
• Life insurers, for the same structural reason
• Foreign central banks and sovereign wealth funds, which hold Treasuries as reserves
• A handful of long-duration bond funds and liability-driven investment strategies

The practice of matching asset maturities to liability maturities is called liability matching, and it is why this buyer base is relatively price-insensitive but also relatively fixed in size. It does not double because supply doubles.

This is the mechanical heart of the crowding-out argument. A pension fund that needs 30-year assets to cover 30-year promises does not care enormously whether it buys them from the US Treasury or from a AA-rated technology company. If the technology company is offering more yield for what the fund judges to be tolerable extra risk, the Treasury has to sweeten its own terms to compete. Adding to that, foreign holdings of Treasuries fell in June, with the UK, China and Japan all trimming.

The Term Premium, Explained

A long-term interest rate can be split into two conceptual pieces. The first is the market's best guess at the average short-term policy rate over the life of the bond. The second is the extra compensation investors demand for locking money up rather than rolling over short-dated paper repeatedly. That second piece is the term premium.

It cannot be observed directly. It is inferred from statistical models, the best known being the New York Fed's ACM model and the Federal Reserve Board's Kim-Wright model. The two do not always agree on the exact number, which is why serious analysts treat term premium estimates as a range rather than a reading.

What matters is the direction. After roughly five years hovering near or below zero during the era of quantitative easing, the term premium has moved decisively back into positive territory. Investors are once again being paid, visibly, to accept duration risk.

This is the channel through which supply becomes a yield story. More long-dated paper competing for the same limited pool of long-horizon buyers raises the price of duration. That shows up as a wider term premium, and a wider term premium raises long yields even when expectations for Fed policy are unchanged.

The Case That AI Is Driving Yields

Bank of America's economists have written that AI borrowing is potentially crowding out long-end Treasury demand and has played a major role in the rise of yields. They estimate that the surge in corporate debt sales, together with heavier issuance of mortgage-backed securities, pushed 10-year rates up by roughly 0.3 percentage points this year.

Barclays takes a similar view of the mechanism, framing the rise in rates as less about inflation and more about the budget deficit, AI-related issuance competing with Treasuries, and a higher term premium. Anshul Pradhan, the bank's head of US rates research, noted that three separate data releases in August argued for lower yields and long-end yields rose anyway.

The scale is the reason this is being taken seriously. Vanguard's numbers show the five hyperscalers issued roughly $35 billion a year on average between 2020 and 2024, then $93 billion in 2025, and around $132 billion so far this year including a single multi-tranche offering of about $53 billion. Estimates for total AI-related issuance across the wider ecosystem in 2026 run from roughly $300 billion to $570 billion.

And it is not a one-year event. Hyperscaler capital expenditure is projected to approach $800 billion this year and to exceed $1 trillion annually from 2027 through 2030. JPMorgan strategists put total AI infrastructure spending at $5.5 trillion through 2030. Morgan Stanley estimates $3.2 trillion by 2028, of which roughly $1.75 trillion would need to come from credit markets.

The Case That AI Is a Secondary Cause

The sceptical argument is strong, and it starts with proportion. The high technology sector accounts for only about 12.8 percent of US bond issuance in 2026. Financial companies remain by far the largest borrowers at roughly 45.2 percent. AI is the fastest-growing source of supply, not the largest.

The mechanism itself is contested. Corporate and government bonds are not perfect substitutes, the buyer pool is not genuinely fixed, and the market has absorbed record supply this year at credit spreads that many managers describe as attractive entry points rather than signs of distress. Treating some of the highest-rated borrowers in the investment-grade index as a source of systemic stress arguably confuses a supply story with a credit story.

Meanwhile the fiscal picture speaks for itself. The July federal deficit came in at $432.3 billion, the largest monthly shortfall since March 2021, taking the year-to-date total to nearly $1.8 trillion. Interest on a national debt approaching $40 trillion has cost roughly $1.2 trillion this year. A recent $42 billion 10-year note auction cleared at 4.68 percent, the highest in 19 years, and BMO has noted that five of the previous seven 20-year auctions tailed, meaning they priced at a higher yield than the market indicated beforehand.

There is also a credibility question. Reuters columnist Jamie McGeever has argued that the more important trigger for higher borrowing costs in recent months has been the new Fed chair, Kevin Warsh, and specifically remarks that appeared to downplay inflation risks at a time when inflation has run above the 2 percent target for five years. Add an inflationary oil shock from the conflict with Iran, rising Japanese yields dragging global rates along, and subdued foreign official demand, and there are several forces each capable on its own of moving the long end.

How This Reaches a Mortgage Quote

American mortgage rates are not set by the Fed. They track the 10-year Treasury yield plus a spread, because a 30-year mortgage is typically repaid or refinanced long before maturity, making the 10-year the closest comparison. Over the long run that spread has averaged somewhere in the region of 1.7 to 1.8 percentage points.

The spread exists because a mortgage is riskier for an investor than a Treasury in a specific way. Borrowers can refinance when rates fall and will not when rates rise, which means the lender gets the money back at the worst possible moment. That is prepayment risk, and investors charge for it.

So the chain runs like this. Heavy long-dated issuance, from both the Treasury and the AI complex, raises the price of duration. That widens the term premium. The term premium lifts the 10-year yield. The mortgage rate is the 10-year plus a spread. The Fed can hold its policy rate at 3.50 to 3.75 percent, as it has all year, and a homebuyer will still be quoted 6.65 percent.

The same logic runs through auto loans, credit card APRs, commercial property refinancing and the discount rate applied to long-duration equities.

What the Credit Market Thinks of the Borrowers

On conventional measures the hyperscalers are excellent credits. Morgan Stanley data for the first quarter of 2026 shows total leverage of about 1.3 times, net leverage of 0.5 times, cash covering 128 percent of debt and a median rating of AA minus. The wider non-financial investment-grade universe carries total leverage nearer 2.4 times at a BBB rating.

But the market is charging more anyway, and the tells are in the plumbing rather than the ratings:

• The median new-issue concession, the extra yield a borrower offers to get a deal done relative to its existing bonds, rose to around 12 basis points in 2026 from about 2.25 basis points in 2025. A basis point is one hundredth of a percentage point.
• A Reuters analysis of LSEG data found 78 of 91 comparable hyperscaler bonds issued in 2026 were trading at higher yields in late July than at issuance, meaning early buyers were sitting on losses.
• Credit default swap spreads across the group have widened materially even as the broader investment-grade market has barely moved. A credit default swap is insurance against a borrower defaulting, and its price is a clean read on perceived credit risk.
• Oracle has become the market's designated worry. S&P cut its long-term rating a notch to BBB minus, the lowest investment-grade rung, citing the financial burden of expanded AI investment.

There is also the question of what is not on the balance sheet. A meaningful share of AI capacity is being financed through long-term leases and purchase commitments that do not appear as debt in the conventional sense. Rating agencies have so far assessed hyperscalers on reported balance sheets. If that treatment changes, spreads would likely widen. The Bank for International Settlements has warned that if the AI boom fades, bond markets would be particularly exposed.

Key Takeaways for Investors

• Duration supply, not dollar volume, is the variable that matters. Long-dated issuance consumes a scarce resource: the willingness of a narrow set of institutions to bear decades of interest-rate risk.
• The crowding-out effect is real but modest. Bank of America's estimate of roughly 0.3 percentage points on the 10-year from corporate and mortgage-bond supply is meaningful without being the main story.
• Deficits, inflation above target for five years, an oil shock and doubts about Fed credibility remain the primary drivers of the long end. AI issuance is an accelerant, not the fire.
• Credit quality and supply pressure are different things. Hyperscaler balance sheets are strong on reported metrics, yet wider new-issue concessions and secondary underperformance show the market is charging more for the sheer volume.
• Oracle is the exception rather than the rule, and the concentration of concern in one name is itself informative.
• The transmission to household borrowing costs is mechanical and does not require the Fed to do anything. Policy rates and mortgage rates have visibly decoupled this year.

Conclusion

The most useful thing to take from this episode is not a view on whether AI debt is or is not pushing up yields. It is the recognition that the bond market has stopped behaving like a discounting machine for Fed policy and started behaving like an ordinary market clearing an unusually large amount of a specific good: long-dated interest-rate risk.

For roughly fifteen years, central bank asset purchases removed duration from the market and the term premium sat near zero or below. That era ended, and the AI buildout arrived at precisely the moment the government's own borrowing needs were expanding. Two enormous long-term borrowers are now bidding for the same finite pool of patient capital, and the price of that capital is being set in public, at auction, several times a month.

What makes this structural rather than cyclical is that neither borrower is going away. Federal deficits are unprecedented outside a crisis, and hyperscaler capital expenditure is forecast to exceed a trillion dollars a year through the end of the decade. The cost of long money is likely to stay higher than the last two decades conditioned anyone to expect, and that repricing will keep showing up in places that appear to have nothing to do with data centres.