The Hook. In one of the evening digests, a phrase from Capital Economics caught my eye—one an engineer accustomed to reading financial summaries through an infrastructure lens simply can't ignore: "Hyperscaler borrowing for AI infrastructure is competing for the same buyers as sovereigns." One sentence—but hidden within it is possibly the most important structural shift in global debt markets since China began building up its Treasury holdings in the 2000s. If this is true—and it is true, as confirmed nearly simultaneously by Dallas Fed, BIS, Moody's, S&P, and PIMCO—then we're witnessing a story no one is telling as a unified narrative: for the first time in thirty years, the largest technology corporations have shifted from "earn and spend" to "borrow and build," and now they're competing with their own government for global capital.
On the surface, no one sees this, because everyone's used to "Big Tech = fortress balance sheets." Indeed: Microsoft, Alphabet, Amazon, and Meta were still financing their capex almost entirely from operating cashflow in 2024, and the debt share of financing was around 9%. By mid-2026, that figure jumped to 32%—in other words, over eighteen months the debt component grew 3.5x. In absolute terms: five hyperscalers issued $121 billion in corporate bonds in 2025—more than they issued cumulatively over the previous five years. And in the first half of 2026—already $159 billion, meaning the 2025 record was beaten by mid-year.
But that's just the tip. The real story is in off-balance-sheet "shadow debt": hyperscalers are shifting risk into joint venture structures with private credit, and BIS directly calls this "shadow borrowing"—obligations economically equivalent to debt but living outside GAAP balance sheets. The numbers Moody's analysts pulled out in July 2026 look insane: $1.2 trillion in total lease obligations across six hyperscalers, of which $820 billion are leases on data centers that haven't even started construction yet. That is, these are contracts to rent capacity that physically doesn't exist.
To understand why this matters for the debt market, we first need to see the scale of AI capex in perspective. According to Morgan Stanley (June 2026), the combined capex of the four largest hyperscalers—Alphabet, Amazon, Microsoft, and Meta—will be around $700 billion in 2026, nearly double their combined 2025 spending. Moody's in July 2026 gave a broader estimate: $785 billion across six hyperscalers (including Oracle and Nvidia) in 2026 and around $1 trillion in 2027. FactSet adds another angle: accounting for finance leases and customer prepayments, calendar-2026 capex approaches $800 billion, and by FY28 will exceed $900 billion.
To put this figure in understandable terms: $700 billion is more than Switzerland's entire annual GDP ($884 billion in 2024) or the Netherlands ($1.1 trillion). That is, five American corporations are spending on GPU farms, fiber optics, and power substations an amount equivalent to the GDP of a small developed European country. Every year. And this amount is growing 70–80% annually, which according to FactSet is "the largest annual increase of the cycle."
The most important thing in this picture isn't the absolute numbers, but the structural shift in funding sources. Here's how Moody's describes it in research from July 23, 2026: "Previously, these companies relied on asset-light structures centered around software, intellectual property, and scalable cloud services that required modest capital investment. The transition from asset-light to asset-heavy models requires unprecedented volumes of investment and capital raising." This formulation is key to everything that follows: for thirty years Silicon Valley created the world's most valuable companies using a model where software requires almost no capital to replicate. Generative AI broke that model—it requires physical hardware: warehouses packed with expensive and power-hungry servers and chips.
The specific numbers Moody's pulled out in July 2026 look like this:
Looking at historical bond issuance perspective (per Bank of America, via Axis Intelligence):
| Year | Combined Bond Issuance, 5 Hyperscalers |
|---|---|
| 2020–2024 (annual avg) | $28 billion |
| 2025 | $121 billion |
| 2026 (1H) | $159 billion |
This is a four-fold increase in average annual issuance volume over eighteen months, and 2026, according to Morgan Stanley's forecast, will end at $570 billion in global AI-related debt issuance (four times 2025). Meanwhile, debt's share of capex financing (Axis Data Center Debt Intensity Ratio) grew from 2% in 2022 to 17% in 2025—a genuine structural shift from self-financing to debt financing of infrastructure.
And now we come to the most interesting part. On August 18, 2026, the yield on 30-year U.S. Treasuries exceeded 5.31% for the first time since June 2007 (CNBC, 18.08.2026). Over two months, since late June, the long end of the curve added more than 40 basis points. CNBC directly names three causes: (1) worsening fiscal deficit (in July the budget gap hit $432.3 billion—the highest since March 2021), (2) inflation stuck above the Fed's 2% target, (3) "rash of corporate debt issuance competing with Treasuries for investors' favor"—corporate debt issues competing with Treasuries for investor preference.
That third reason is the subtlest. Before 2025, corporate debt was niche relative to Treasuries: pension funds, insurance companies, foreign central banks could hold enormous Treasury positions, and absorption capacity was so large that any corporate issuance was a drop in the ocean. Now, when five hyperscalers issued $159 billion in just the first half of 2026, and this continues to accelerate, the drop has become a substantial share of growth in global long-duration debt.
Dallas Fed in February 2026 made a formal estimate: $300 billion in AI-related investment-grade bonds in 2026 = 360 billion in 10-year equivalents by duration supply, which represents about 1/8 of total duration supply from U.S. Treasury. This, note, is not all corporate debt—this is only the AI-related portion. If you add oil & gas, banks, telecom, healthcare, and all other corporate markets, U.S. corporate debt is already comparable to Treasuries in contribution to duration supply, which has never happened in history.
And this creates what Dallas Fed directly calls a "crowding-out effect"—the displacement effect. When AI buildout pulls global institutional buyers (pension funds, insurers) into hyperscaler bonds rated A/AA, those same buyers have less capital for Treasuries. To attract them back, the U.S. Treasury must raise yields. This is the exact mechanism through which AI capex pressures the long end of the curve.
In March 2026, the Bank for International Settlements (BIS) published a formal analysis in its Quarterly Review that directly introduces the term "shadow borrowing"—obligations economically equivalent to debt but mostly living outside corporate balance sheets.
The mechanism is simple and elegant: a hyperscaler (say, Microsoft) doesn't build a data center itself. Instead, it creates a joint venture with a specialized private credit fund. The JV raises debt through private placement, the hyperscaler contributes a minority equity stake and assumes a long-term lease obligation or take-or-pay contract. Debt is serviced from lease payments, held by private credit funds and institutional investors, sometimes receives investment-grade rating thanks to hyperscaler guarantees. Economically, this replaces upfront capex with multi-year operating expense—almost like operating leases replace real estate purchases.
But here's what's important: from a risk perspective, this is debt. If the project doesn't hit targets, if AI demand reverses, if OpenAI goes bankrupt—the JV debt holders will take losses. And JV debt holders aren't pension funds directly, but private credit funds, which in turn may be subject to procyclical outflows. BIS warns directly: "These arrangements strengthen links between hyperscalers and non-bank investors such as private credit vehicles and insurers. Banks support the vehicles with funding lines, potentially creating new shock transmission channels—e.g., via refinancing pressures at the vehicle level, procyclical shifts in private credit appetite or the activation of guarantees."
In other words, we're creating a financial system where AI buildout risk is smeared across non-bank channels that in a crisis can behave as unstably as in 2008. BIS Bulletin 120 (January 2026) shows that outstanding private credit to AI companies grew from nearly zero to over $200 billion, and in 2025 alone private credit funds issued $40 billion in new loans to the AI sector (for comparison: $3 billion in 2010).
And here's where it gets slippery. Moody's in July 2026 noted "structural circularity within the AI boom": some of the multi-billion-dollar hyperscaler backlogs stem from strategic deals with pre-IPO AI labs—OpenAI and Anthropic. The scheme looks like this:
S&P in its Oracle press release on July 9, 2026 stated directly: "OpenAI represents approximately half of the $638 billion RPO." And added: "If OpenAI were unable to pay Oracle, we believe Oracle could be left with massive data center lease contracts it cannot break or would be forced to sublease on less favorable terms."
On August 17, 2026, Reuters reported that Nvidia will provide up to $105 billion in guarantees for financing OpenAI's Ohio data center. This is the largest corporate guarantee in tech industry history. Essentially, Nvidia, whose GPUs sit inside every one of these data centers, is guaranteeing that OpenAI will be able to pay for them. This creates a feedback loop where Nvidia's stock value (on which the entire AI narrative rests) implicitly backs debt obligations through which that value is being created. This is circular financing in pure form—those very "circular deals" that Bloomberg made into a separate graphics page in January 2026.
CNBC on July 24, 2026 drew the line: "Moody's also pointed to structural circularity within the AI boom." In other words, rating agencies are now openly calling this circular, without hiding behind euphemisms.
The final layer of the story is the return of bond vigilantes. Ed Yardeni coined this term in the early 1980s to describe bondholders who "strike" against poor fiscal policy by selling Treasuries and pushing yields up. In 1993, when Bill Clinton was just starting his first administration, 30-year Treasuries traded above 7.8%, and analysts talked about how "the bond market beat Clinton" (to which Democratic strategist James Carville famously responded: "I used to think if there was reincarnation, I wanted to come back as the president or the pope. But now I want to be the bond market. You can intimidate everybody").
Now, in August 2026, 30-year Treasuries broke 5.31%—the highest since June 2007. And again there's talk of bond vigilantes. But the situation fundamentally differs from 1993. Back then, the bond market was reacting to one source of pressure—the government's fiscal deficit. Now it's reacting to three independent sources simultaneously:
And here's the key point almost no one says out loud. In May 2026, Fortune interviewed Guy LeBas, chief fixed-income strategist at Janney Montgomery Scott, and he said something many market pros repeat but rarely write in headlines: "Bond vigilantes do not exist. The bond market has become too large and too dominated by non-discretionary buyers like pension funds for a handful of participants to form a message. The pure explanation is mostly automatic: momentum funds that buy when prices rise and sell when they fall."
In other words, the "return of bond vigilantes" isn't an ideological investor rebellion against fiscal policy, but an automatic algorithmic reaction to the convergence of three independent structural pressures. This is simultaneously less dramatic and more dangerous: you can't negotiate with algorithms like Carville negotiated with the market in 1993 through deficit reduction.
I should note an important counterargument. PIMCO in June 2026 published analysis titled "AI Financing Needs Do Not Override Cyclical Drivers of Yield"—and their position radically differs from what Bloomberg and CNBC write. Their thesis: "Structural pressures from AI buildout are real, but they're growing slowly and aren't driving yields right now." PIMCO points out that 30-year yields are currently driven by cyclical factors: fiscal deficit, inflation, OPEC oil premium, Fed uncertainty.
And there's logic to this: $570 billion in global AI debt in 2026 is a lot, but it's still only ~3.5% of the global debt market ($150+ trillion). Dallas Fed states directly that AI bonds represent ~12.5% of duration supply from Treasuries. That's significant but not dominant. The real story is in the trajectory: if 2026 is the start of acceleration, then 2027–2028 with $900 billion capex and growing debt share could create exactly that critical mass that transforms "structural pressure" into "dominant factor."
Also, an important nuance emphasized by both BIS and Moody's: the sharpest part of the problem isn't investment-grade bonds from Microsoft/Amazon, but speculative debt from CoreWeave, Oracle, and pre-IPO labs. CoreWeave with a Ba3 rating finances GPU farms through complex private debt structures. Oracle after S&P's downgrade to BBB- (one notch above junk) is trying to raise $40 billion by FY27 (June 2026) through a mix of debt and equity. If one of these players runs into trouble—credit spreads will widen, mortgage rates will go up, and the real economy will feel pressure through housing costs, not through the stock market.
This story isn't about AI as such. It's about the financial architecture being created right now, in real time, before our eyes—and almost no one is discussing it as a unified narrative.
The main inversion I see: five years ago the narrative was "tech giants have fortress balance sheets, they finance growth from cashflow, they're not subject to debt cycles." In 2026 it turns out that the world's most valuable companies have become the world's largest borrowers, not because they need to plug holes (like banks in 2008), but because investment appetite exceeded the business's generating capacity by 4x. Microsoft, which recently took pride in $100+ billion operating cashflow, now has to issue bonds and raise equity (Alphabet raised $84.75 billion in one round in June 2026) to finance its own future.
The second inversion is structural. Treasuries were the "risk-free asset" by design: the U.S. government could borrow practically infinitely because it was the only borrower of such scale and creditworthiness. Now, when hyperscalers issue investment-grade bonds in comparable amounts with better real economics behind them, the "risk-free" U.S. Treasury's share in the global debt portfolio is starting to erode. And the more this erosion, the higher the premium Treasury must pay to retain those same institutional buyers. This is a slow, structural, but unstoppable process—and 30-year Treasuries above 5.3% in August 2026, in my view, are not the peak, but only the beginning.
The third inversion is the most alarming. Circular financing (Nvidia guarantees OpenAI debt, OpenAI pays Oracle, Oracle buys GPUs from Nvidia) creates a system where Nvidia's stock value implicitly supports debt obligations through which that value was created. In a normal financial system, debt and equity are different instruments with different risks and different holders. Here they're woven into a feedback loop, and the only way this loop can break without catastrophe is if AI actually pays back all the investments. If it doesn't—we'll see the first AI infrastructure debt crisis in history, in scale comparable to the dotcom crash of 2000–2002, only this time through the debt channel, not equity.
What personally hooked me most. When I started digging into this story, I expected to find another confirmation of the cyclical "AI bubble" narrative. I found a structural shift that will survive regardless of whether AI pays off. Even if tomorrow it turns out that AGI won't be created and GPT-6 is just a statistical trick, the physical infrastructure—data centers, fiber optics, power substations—will remain standing. And with it will remain $1.2 trillion in lease obligations, $460 billion in direct debt, $200 billion in private credit. This infrastructure will find other uses—cloud computing, SaaS, enterprise AI, crypto mining (irony, right?). But the debts will remain. And these debts will be serviced from future cashflows that may not exist. In this sense, AI capex 2024–2026 is a bet that can't be cancelled, even if the bet proves wrong.
If there's one chart worth remembering from this story, it's debt's share of capex financing at hyperscalers: 9% in FY24 → 32% LTM mid-2026 → forecast >40% by 2027. This is a time series that explains everything else—both bond vigilantes and rising 30-year Treasuries and Oracle's downgrade and why Alphabet raised $85 billion in equity instead of living on cashflow.
P.S. What I didn't have time to verify but is definitely worth further digging.