AI debt: the battle for credit spills beyond Silicon Valley

Amazon, Alphabet and their peers are raising vast sums in bond markets. So is the US Treasury. That overlap deserves scrutiny, but it does not by itself explain rising yields. The clearest evidence so far concerns the price investors demand for taking technology companies’ risks. A broader increase in borrowing costs caused by AI remains a hypothesis to test.

On 15 September, the benchmark US ten-year Treasury yield touched 5.041% intraday, its highest since July 2007, according to Reuters. Investors were also revising monetary-policy expectations against a backdrop of oil-market tensions. Into this unsettled market comes a new question: is the financing required for artificial-intelligence infrastructure starting to make borrowing more expensive for governments, other businesses and, ultimately, households? [1]

It is a reasonable question. Answering it requires separating the common interest-rate component from the compensation demanded for lending to a particular borrower. Amazon’s bonds can become more expensive to issue without Washington paying more. Conversely, rising Treasury yields can increase Amazon’s financing costs even if investors have not changed their assessment of its creditworthiness.

The question is where higher borrowing costs are concentrated and how far they might spread.

Big cloud enters the bond market at scale

Hyperscalers are companies operating the computing infrastructure and cloud services that can support enormous workloads. The European study discussed here uses the term for Alphabet, Amazon, Meta, Microsoft and Oracle. Their bonds give investors contractual claims against the companies issuing them. [6][17]

Amazon’s debt note for the quarter ended 31 March 2026 provides a concrete example: $37 billion of dollar bonds and €14.5 billion of euro bonds issued in March alone. Keeping the original currencies avoids an unstated exchange-rate assumption. The stated use of proceeds is general corporate purposes, rather than a legally ring-fenced AI budget. These are issuance amounts, before subtracting repayments of other debt. [2]

Borrowing does not necessarily mean a company has run out of cash. Alphabet raised $31.1 billion in net proceeds from senior unsecured notes in the first quarter of 2026, while reporting $10.1 billion of positive free cash flow after capital expenditure, a non-GAAP measure. Net issuance proceeds are not net borrowing after debt repayments. That quarter also included substantial acquisitions. This historical snapshot is not a statement about September liquidity. It shows that debt issuance accompanies a wider set of capital-allocation decisions. [3]

Oracle’s latest results illustrate a different funding mix. For June–August 2026, its first fiscal quarter of 2027, it reported $23.1 billion in operating cash flow, $28.5 billion in capital expenditure and negative free cash flow of $5.4 billion, a non-GAAP measure. Operating inflows included $11.4 billion of customer prepayments with a significant financing component. Oracle also sold $20 billion of common stock before commissions. [4]

These distinctions matter. Every dollar of announced investment cannot be translated into a dollar of bonds waiting to be sold. Operations may provide some funding, customers may pay in advance and shareholders may contribute new capital. A bond issue may also replace debt falling due. New net financing and gross securities issuance are different quantities.

The Treasury is still looking for buyers

Public borrowing needs are substantial as well. In its 3 August estimate, the US Treasury projected $739 billion of privately-held net marketable borrowing for July–September 2026, followed by $628 billion for October–December. The projections depend partly on assumed quarter-end cash balances. They measure neither the budget deficit nor the gross volume of all Treasury auctions. [5]

Dividing Amazon’s March issuance by Washington’s third-quarter net borrowing would produce a misleading “market share”. The dates and definitions do not match. Setting these numbers alongside one another identifies borrowers that need willing holders of their securities; it does not justify a dramatic ratio.

Nor does the total size of a bond market settle the issue. Trillions of dollars of securities may already sit in portfolios whose owners do not intend to buy anything else today. The price of the next issue depends on the return that persuades an additional buyer to increase exposure : or to sell another asset to make room.

There is more than one price inside a borrowing rate

A useful approximation for a dollar corporate bond is a reference yield at a comparable maturity plus a premium specific to the security. That premium, the spread, compensates investors for risks including default and difficulty selling the bond. Contractual features such as early-redemption options also affect a meaningful comparison. [13][17]

Consider a purely illustrative example: a 5% benchmark yield plus a one-percentage-point spread produces a borrowing yield of roughly 6%. If investors demand a 1.5-point spread, the company’s cost rises to 6.5% with no change in the benchmark. Alternatively, the benchmark could rise to 5.5% while the spread remains unchanged. The result is again 6.5%. The final bill is identical; the mechanism is not.

The government-bond yield itself can also be decomposed. The approach presented by the New York Fed separates expected future short-term rates from a term premium: compensation for the uncertainty involved in lending for longer rather than repeatedly reinvesting at short maturities. This premium is estimated using a model. It is not a directly observable item on a market screen. [8]

AI-related borrowing could put pressure on that premium if it increased the amount of interest-rate risk investors had to absorb. But AI could also change expectations for growth, inflation or monetary policy. The claim that “AI is pushing up rates” therefore covers several potential mechanisms, with very different economic implications.

What investors need from a bond

A bond manager working with a fixed budget must make choices. An attractive new technology bond might lead to smaller purchases of industrial debt, or a reduction in government-bond holdings. Competition would then operate through prices: other issuers might have to offer better terms. That is a plausible chain of events, not an automatic one.

Treasuries and corporate bonds do not always perform the same job. Under US liquidity rules for the banks subject to them, Treasury securities qualify as Level 1 liquid assets; qualifying corporate debt falls within Level 2B. Switching between them is therefore not neutral for an institution building its regulatory liquidity buffer. [10]

Maturity also matters. An insurer expecting to pay benefits far into the future may want long-dated assets. A fund required to hold short-term government securities faces a different choice. Even bonds with similar final maturities can have different interest-rate sensitivity. That sensitivity, known as duration, depends partly on the timing of payments. Adding up issuance without examining the risk being absorbed gives an incomplete picture of competition. [18]

Research by Greenwood, Hanson and Stein, published in the Journal of Finance in 2010, offers a useful counterpoint. Historically, companies sometimes issued more long-term debt when governments favoured shorter borrowing. The markets can therefore complement one another too. That older result says nothing about the size of an AI effect in 2026; it explains why maturity composition matters alongside volume. [12]

Money also does not disappear inside a data centre. Borrowers pay suppliers, employees or other creditors, and those funds continue circulating. Across the economy, savings flows and balance sheets change; bank lending can itself create deposits, as the Bank of England explains. None of this removes capital constraints or lenders’ caution. The scarce resource may be the willingness and capacity to bear risk for twenty years, rather than an immutable quantity of cash. [11]

European demand holds up through the first wave

The ECB blog study published on 31 August, using observations through 20 August, shows rising credit premiums for Amazon, Alphabet and Microsoft in its chart, alongside resilient demand for euro-area issuers. Those premiums are measured against interest-rate swaps, contracts that exchange interest payment streams, rather than against US Treasuries. Some reportedly adjusted their issuance calendars. The authors discuss possible later spillovers rather than describing them as already established. [6]

That distinction changes the policy interpretation. A higher premium charged to a cloud giant may indicate that investors are differentiating risks more carefully. It does not establish that a European manufacturer has lost market access. Even a change in issuance timing may be a successful adaptation rather than evidence of exclusion.

Speaking in Vienna on 14 September, Christine Lagarde nevertheless emphasised the possibility of international transmission. She noted that market participants attributed part of the increase in US long-term real yields to AI-related borrowing. That warning deserves attention, but it is not a causal ECB estimate expressed in basis points. [7]

Both messages can be true. One study describes observed adjustments in a segment of Europe’s bond market. A policy speech warns about a larger, sustained financing wave. Moving from one argument to the other requires specifying the horizon, rather than dropping the qualification.

Reading the rise in Treasury yields

A simple check is to compare several maturities over the same dates. In the H.15 release published on 15 September, the available 8–14 September window shows the two-year nominal yield rising from 4.39% to 4.65%, against 4.80% to 4.97% for the ten-year. The thirty-year moved from 5.25% to 5.34%. Those are increases of 26, 17 and 9 basis points, respectively. One basis point is one-hundredth of a percentage point. [9]

Over this short window, the largest increases were not concentrated at the longest maturities. The curve calls for a broader reading: policy expectations, inflation and risk premiums can move together. This window cannot isolate or rule out a contribution from AI-related borrowing.

These are interpolated constant-maturity yields from the Treasury curve. They should not be confused with the intraday peak on the benchmark security quoted at the beginning. The 4.97% observation is for 14 September; the 5.041% intraday observation was reported on 15 September. [9][1]

Identifying a distinct technology-issuance effect would require examining price changes around unexpected announcements, controlling for macroeconomic news and accounting for hedging transactions. Observing that yields rise while companies borrow cannot distinguish cause, consequence and coincident timing.

Euro bonds open another funding route

A reverse Yankee is a bond issued by a US company in a foreign currency, including the euro. It may reach a different investor base or match the company’s euro revenues. A lower coupon in euros than in dollars does not, by itself, establish a funding saving: converting and hedging future payments against exchange-rate movements has a cost. The ECB’s June 2025 analysis explains that comparison. [13]

For a lender with euro liabilities, a euro-denominated bond removes the currency mismatch on the security’s contractual payments. It does not remove the borrower’s business risk. Problems in American cloud computing can therefore reach a European portfolio without first passing through a decline in the dollar.

There is equity exposure as well. In its 17 August study, the ECB put euro-area households’ exposure to US technology equities at around €440 billion; the holdings data in the corresponding chart refer to the third quarter of 2025. This is a stock measured at market value, not the amount recently lent to hyperscalers. [14]

A fund’s domicile, a bond’s currency and the nationality of the ultimate investor are three different pieces of information. A euro bond issue does not establish that all subscribers are Europeans, or that the proceeds will be spent in the United States. Our investigation into European funds with American portfolios explored this distinction.

Europe’s question is consequently more specific: what financial exposure should it accept, what infrastructure should it build, and what economic benefits should it obtain? Financing an American company does not necessarily destroy European wealth. Lenders may earn returns and European customers may benefit from its services. Lagarde herself acknowledges the benefits of adopting imported technology. Strategic dependence and investment profitability are nevertheless separate questions. [7]

Look past the tenant’s name to the actual borrower

Not all AI-related debt is issued by the large parent companies. Researchers at the Bank for International Settlements (BIS) also describe vehicles that finance data centres and service their debt using leases or commitments to purchase capacity. Guarantees and risk allocation depend on the contract. A hyperscaler’s involvement does not automatically turn project debt into an obligation of that parent company. [15]

Imagine a building occupied by a large cloud provider. The building’s lender must assess the lease term, termination conditions, future investment needs and available legal recourse. A direct bond issued by the parent company would create a different claim. Two investments marketed as “AI exposure” may therefore provide very different protections.

These arrangements explain why competition extends beyond publicly traded bonds: the same insurer may choose between government debt, corporate bonds and private infrastructure finance. But adding the project loan, every promised lease payment and the tenant’s debt would risk double counting. The journey from a GPU to a financial claim was central to our investigation into the debt behind AI.

How the financing wave could develop

Orderly absorption remains possible. New investors arrive, issuance is spread out, companies adjust maturities and accept higher financing costs. The additional burden remains concentrated on particular issuers. Supporting evidence would be sustained demand for other borrowers even during large technology deals. This scenario does not require credit spreads to remain unchanged.

Portfolio-driven spillovers would emerge if buyers reached their exposure limits and reduced other positions to absorb technology debt. Pressure could spread to comparable companies and then to other market segments. Warning signs would include larger pricing concessions, delayed deals and persistent relative increases in borrowing costs for companies with no direct AI connection. The additional step from corporate credit to sovereign yields would still need separate evidence.

An investment-cycle reversal could produce a less intuitive result. Disappointing revenue might push up technology credit premiums while prompting a flight towards safety and weaker growth expectations. Treasury yields could then fall while AI companies paid more to borrow. This outcome would depend partly on inflation and continued confidence in government debt. The BIS highlights the possibility of a boom financed against ambitious future earnings turning into a retrenchment. [16]

Conversely, genuinely productive AI could support investment and increase the expected return on capital, contributing to higher real interest rates. That would be a different story from a financing crisis. The BIS scenarios underline that the long-run outcome depends on productivity gains and the demand needed to make them profitable. [16][19]

The bill would arrive through new borrowing and refinancing

For a borrower outside technology, the practical exposure lies in new financing, floating-rate liabilities and upcoming refinancing dates. An existing fixed-rate contract does not increase its coupon simply because investors now demand a higher market yield. Its resale value, however, may fall immediately. [17]

The sensitivity is straightforward. On €1 billion borrowed for one year, an additional 0.50 percentage point means €5 million in extra interest, using simple interest, before fees and tax, with no principal amortisation. This illustrates the effect of a rate change. It does not estimate a current AI-related surcharge.

For a new French mortgage, transmission would be more indirect still: the bank’s own funding cost, euro reference rates, the customer’s credit risk, competition and the commercial margin all matter. Attaching an “American data centre” label to a higher monthly payment would require tracing those intermediate links.

The decisive indicator will therefore not be the next record bond issue in isolation. It will be who is paying more, against which benchmark, over what maturity and following what new information. More demanding risk pricing must also be distinguished from genuinely impaired access to credit.

Large technology companies have added substantial funding demand to a market already serving governments. Their borrowing capacity may help deepen that market; their growing risks may also make its terms less accommodating. So far, the European pressures examined here are concentrated in technology debt. Whether they spread will depend on investor demand, further issuance and the revenue these projects actually generate.

Sources and methodology

Limitations

Information cut-off: 16 September 2026, before that day’s FOMC decision. Scenarios are conditional and carry no assigned probabilities. The rate comparison uses a consistent H.15 window; the corporate examples are not a sector-wide total. No number of basis points is causally attributed to AI.