 |
When Big Tech began turning to the bond market to finance its great AI buildout, my first thought was relatively simple.
Perhaps borrowing allowed these companies to preserve cash, maintain strategic flexibility, and continue returning capital to shareholders rather than paying for every data center and GPU directly from their own balance sheets.
But as I have been watching the AI arms race intensify—with the largest technology companies spending ever larger sums to secure chips, data centers, electricity, land, and computing capacity—I began to think about the other side of that trade.
What happens if this competition eventually produces clear winners and losers?
The question is not whether AI succeeds. Its industrial potential can be big.
But an industry can succeed spectacularly while some of the companies financing that success earn disappointing returns.
If that happens, what becomes of the enormous amount of long-term capital now flowing into Big Tech bonds? Where does that money go when investors begin distinguishing between stronger and weaker AI credits? And what might that mean for U.S. Treasuries, which compete for part of the same long-duration institutional capital?
This research began with that question.
Part I — What We Know
The AI Arms Race Is Moving Onto the Balance Sheet
For several years, the largest technology companies could present the AI investment cycle primarily as a cash-flow story.
They generated enormous operating cash flows. They used part of that cash to build data centers and purchase chips. They still had room for dividends, share repurchases, acquisitions, and strategic investments.
That description is no longer sufficient.
The cash generation remains extraordinary, but the scale of investment is rising even faster at several companies. Debt, leases, customer prepayments, supplier commitments, equity financing, and guarantees are increasingly being added to the funding mix.
The companies are participating in the same AI race. They are not financing that race in the same way.
The periods and accounting definitions differ, so this table is not a direct ranking of financial strength or AI efficiency. Nor does it tell us which company will eventually win the AI race.
It establishes a narrower point.
The AI infrastructure cycle has grown large enough that even the world’s most cash-rich companies are broadening their sources of capital.
Microsoft remains the clearest example of a company able to fund massive investment while retaining substantial financial flexibility. Oracle sits at the other end of the current spectrum: capital spending exceeded operating cash flow, free cash flow turned deeply negative, and external financing became essential.
Alphabet, Amazon, and Meta lie somewhere between those two positions. Their cash generation remains powerful, but capital spending is absorbing a much larger share of it.
NVIDIA is different again. It is a major beneficiary of the buildout rather than a traditional hyperscale operator, yet its financial role now extends beyond selling GPUs. It is also an issuer, investor, cloud customer, strategic financier, and guarantor within the AI ecosystem.
Current balance sheets are a baseline.
They are not predictions of future winners and losers.
Investors Are Treating Big Tech Debt Almost Like a New Safe Asset
This description is intentionally provocative—and requires an immediate qualification. Of course, Big Tech corporate bonds are NOT equivalent to the U.S. Treasuries.
Treasuries carry the backing of the U.S. government, are issued in dollars, trade in the deepest sovereign-bond market in the world, serve as collateral throughout the financial system, and receive unique regulatory treatment.
A corporate bond still carries default risk, downgrade risk, liquidity risk, and the possibility that its price will fall relative to Treasuries.
The narrower point is that highly rated technology-company bonds have become major sources of high-grade, long-duration assets. They offer pensions, insurers, and investment funds additional yield while allowing those investors to extend duration—the sensitivity of a bond portfolio to long-term interest rates.
The scale is becoming material.
Meta’s April offering introduced $25 billion of new bonds, including $14 billion maturing in 2046 or later. Its 2066 bond was priced at a spread of 1.47 percentage points above the comparable Treasury benchmark.
Amazon’s July offering was also $25 billion. Its 2066 bond was issued at a yield of approximately 6.30%, or 1.25 percentage points above its Treasury benchmark.
Alphabet followed with another $25 billion in August, again extending maturities to 2066. Oracle had already issued $25 billion of long-dated debt in February.
For perspective, the U.S. Treasury’s August quarterly refunding plan included a $16 billion 20-year bond auction and a $25 billion 30-year bond auction. https://home.treasury.gov/news/press-releases/sb0590
That does not mean Meta issued “as much debt as the Treasury.” The Treasury sells securities continuously across many maturities and finances needs on an entirely different scale.
It means a single Big Tech transaction can now introduce a meaningful amount of high-grade duration alongside individual Treasury long-end auctions.
These securities therefore compete for part of the same institutional capital pool.
That does not prove that Big Tech bonds are taking demand away from Treasuries. A pension or insurer can own both. New corporate issuance can also attract capital from cash, shorter-duration bonds, foreign securities, or other corporate issuers.
What it does show is that the long-duration market is gaining a new and rapidly expanding source of supply.
Until now, investors have largely been willing to absorb it.
The harder question is what happens if they stop treating all major AI issuers as variations of the same high-quality story.
Everything up to this point describes what we can observe today.
What follows is a thought experiment.
Part II — What If?
AI Is a Competition, Not a Collective Victory
AI can succeed spectacularly without producing equally attractive returns for every company financing the buildout.
That is not a forecast that any current issuer will fail. It is simply the recognition that multiple companies are spending hundreds of billions of dollars to build overlapping infrastructure, train competing models, and win many of the same enterprise and consumer customers.
They cannot all be guaranteed identical pricing power, utilization rates, margins, or returns on invested capital.
The late-1990s internet and telecommunications boom offers one restrained historical comparison.
The internet was real. Fiber networks were necessary. Digital traffic eventually became foundational to the economy.
But not every company financing the infrastructure boom earned an attractive return. Some networks were overbuilt, pricing collapsed, weaker borrowers lost access to capital, and a number of telecom issuers eventually failed.
Today’s hyperscalers generally have far stronger cash flows, market positions, and balance sheets than many telecom challengers of that period.
The lesson is not that AI is “dot com bubble 2.0.”
It is that being right about a technology is not the same as being right about every company financing it.
What Happens When the First Laggards Appear?
In credit markets, a “loser” does not need to default.
It can simply be a company whose AI investment generates lower returns than bondholders expected.
Imagine that an issuer continues spending heavily, but its AI revenue grows more slowly than anticipated. Data-center utilization disappoints. Pricing becomes more competitive. Free cash flow stays weak for longer. Management then faces a choice between reducing investment, issuing more debt, cutting shareholder returns, or accepting pressure on its credit rating.
The transmission could look like this:
Lower-than-expected AI returns lead to weaker free cash flow. Weaker free cash flow produces greater financing discipline. Investors then demand a larger credit spread. Rating pressure may follow. Distress or default would appear only at the extreme end of that process.
Bond investors do not need to wait for bankruptcy to lose money.
A 20-, 30-, or 40-year corporate bond can fall substantially in price if investors decide that the company deserves a larger spread over Treasuries. The issuer may continue paying every coupon on time, but the market value of the existing bond can still decline.
This is why credit analysis is crucial even for companies with famous brands and investment-grade ratings.
S&P Global Ratings says it is monitoring AI monetization, demand durability, potential overcapacity, leverage, free cash flow, leases, guarantees, and other debt-like commitments as infrastructure investment rises. https://press.spglobal.com/2026-08-27-AI-Infrastructure-Investment-To-Exceed-1-3-Trillion-By-2027,-S-P-Global-Ratings-Says
That is evidence for the mechanism, not evidence that a credit crisis is imminent.
The market does not need to conclude that AI has failed.
It needs only to conclude that the returns will not be distributed evenly.
Where Will the Bond Money Go?
Suppose investors eventually become less comfortable holding the debt of one AI issuer.
Where will the capital move?
There is no reason to assume that it all goes into Treasuries.
Some investors may move into the bonds of companies viewed as stronger AI competitors. Others may prefer unrelated investment-grade sectors. Some may shorten duration, hold more cash, or move into government and agency securities.
The internet revolution era credit repricing points to this broader response.
As telecom and speculative-grade credit deteriorated in 2000–2002, spreads widened, downgrades increased, and weaker borrowers found it more difficult to issue long-term debt. At the same time, the market remained more receptive to stronger investment-grade issuers.
Investors moved up the credit spectrum.
That movement did not have one destination, and the period cannot be treated as a clean experiment. The U.S. economy was also experiencing a recession, Federal Reserve rate cuts, the September 11 attacks, and the failures of Enron and WorldCom.
Historical data therefore cannot tell us that a specific dollar left a telecom bond and entered a Treasury bond.
They establish something more modest.
When confidence in a heavily financed technology cycle weakens, capital becomes more discriminating. Stronger corporate bonds, government securities, and cash-like assets can all benefit while weaker credits face higher funding costs.
If a similar distinction eventually develops inside AI, Treasuries could be one beneficiary.
They would not necessarily be the only one—or even the largest one.
The Bigger Question: Is AI Demand as Independent as It Looks?
Before drawing a conclusion, there is one more feature of today’s AI boom worth examining.
Unlike the previous scenario, this one is already visible.
Some of the industry’s largest suppliers, customers, investors, and financiers are financially connected to one another. Capital can move into an AI company, help that company purchase computing capacity, and then return to the infrastructure provider as revenue.
That does not make the demand artificial.
It does make the system more intertwined than headline revenue growth alone might suggest.
Microsoft and OpenAI provide the most familiar example. Under their April 2026 agreement, Microsoft remains a major OpenAI shareholder and its primary cloud partner, while OpenAI has contracted to purchase an incremental $250 billion of Azure services.
The relationship is no longer as exclusive as it once was: OpenAI can serve products across other cloud platforms. But Microsoft remains both an investor in the customer and the principal infrastructure provider receiving much of its computing expenditure.
Amazon and Anthropic show a similar interaction in a different form.
Amazon’s June 2026 SEC filing discloses two $5 billion investments in Anthropic and a financing facility whose availability is linked partly to Amazon delivering AWS computing capacity. After the second investment, the facility’s remaining maximum availability was reduced to $15 billion.
Anthropic, meanwhile, uses AWS infrastructure and Amazon’s AI chips.
Amazon capital and financing support Anthropic. Anthropic purchases Amazon computing capacity. The arrangement can strengthen both parties, but the investment relationship and the demand relationship cannot be viewed as entirely separate.
Oracle represents another structure.
In its fiscal 2026 results, Oracle disclosed that customers had provided approximately $75 billion through prepayments or customer-supplied GPUs and related hardware. https://www.oracle.com/news/announcement/q4fy26-earnings-release-2026-06-10/
Here, future demand helps finance the infrastructure required to satisfy that demand.
This can reduce Oracle’s upfront funding burden and demonstrate strong customer commitment. It also means that customer financing, infrastructure construction, and future cloud revenue are part of the same commercial arrangement.
NVIDIA’s role is broader.
Its July 2026 quarterly filing disclosed $36 billion of AI cloud-service commitments and maximum gross guarantees of approximately $3.5 billion associated with land, power, and facilities for selected AI cloud providers.
NVIDIA also disclosed a separate contingent guarantee, capped at $105 billion, connected to a 4.25-gigawatt OpenAI and SB Energy data-center project.
That $105 billion ceiling is not current funded debt, a probable loss estimate, or an amount NVIDIA has already paid. It is a maximum contingent exposure whose realization depends on the agreement’s conditions and future project performance.
The significance lies in the structure.
NVIDIA can supply GPUs, invest in AI companies, purchase cloud capacity, support infrastructure financing, and provide guarantees connected to projects that deploy NVIDIA technology.
Supplier, financier, investor, and customer relationships are beginning to overlap.
None of these arrangements proves “fake demand” or “circular revenue.” The computing capacity is real, the facilities must be built, and customers are entering legally binding commercial agreements.
But interconnected financing can amplify a cycle in both directions.
During expansion, investment supports infrastructure, infrastructure supports demand, and demand supports further investment.
If returns diverge later, investors will need to determine how much exposure sits inside ordinary debt, leases, purchase commitments, equity investments, guarantees, and customer-financing arrangements.
Headline capital spending will not tell the whole story.
The Verdict: AI Success Will Not Be Shared Equally
The evidence does not support an AI-crash prediction.
It does not support a Big Tech default call or a straightforward bullish case for Treasuries.
It supports a more useful conclusion.
The bond market is financing an AI race whose eventual returns are unlikely to be distributed evenly.
Today, the largest participants remain extraordinarily profitable. Their debt has generally been issued with strong investment-grade ratings. Demand for their bonds demonstrates that investors still regard the group as capable of funding a long-duration infrastructure cycle.
But the financing structures are already diverging.
Some companies can cover investment comfortably from operating cash flow. Others are absorbing most of their cash generation or spending beyond it. Some rely more heavily on debt. Others use customer prepayments, leases, supplier commitments, equity investments, or guarantees.
At the same time, the AI ecosystem is becoming financially interconnected. Companies can be suppliers, customers, investors, lenders, and guarantors to one another.
That interconnectedness does not invalidate the demand.
It increases the importance of understanding where the ultimate economic return is being earned—and who bears the risk if it disappoints.
If AI returns begin to diverge, the first signal may not be a default.
It may be a credit spread.
Investors may begin demanding more yield from issuers whose capital spending appears less productive, while continuing to fund stronger competitors. Capital may rotate into better corporate credits, Treasuries, or shorter-duration assets rather than leaving the AI credit market in one clean movement.
The point is not to predict which company wins or where every dollar eventually goes.
It is to recognize the distinction that the current investment boom can easily obscure.
AI may succeed spectacularly.
That does not mean every company financing the boom—or every investor funding it—will share equally in that success.