An evening digest of the world press carried a story that at first glance looked like routine fare for the finance section: yields on ten-year US Treasury bonds hit 4.8%, thirty-year yields reached their highest since 2007, the national debt topped $40 trillion, and Treasury Secretary Bessent clumsily attempted to "calm" the market. The market churns out such headlines in batches, and usually they die in the feed within a day.
But the recap contained one phrase impossible to ignore: into the same pool of savings from which governments finance deficits has entered the private debt of AI hyperscalers — $132 billion in issuance from five tech corporations in seven months this year versus an average of $35 billion per year in 2020–2024. Demand for money is now shaped by both the Treasury and data centers. If this is true, then the "risk-free rate" — the foundational constant on which all asset valuations on the planet rest, from mortgages to pensions — has ceased being a constant and turned into a variable with a new term. I wanted to verify who wrote that term into the equation, how it works, and what happens if you subtract it. Spoiler: it's already happening — people just aren't talking about it in words yet, only in numbers on trading terminals.
I'll start with scale, because the scale itself upends the familiar picture. From 2020 to 2024, five hyperscalers — Alphabet, Amazon, Meta, Microsoft, and Oracle — issued an average of $35 billion in debt per year, and it was more of a hygiene procedure: refinance the old stuff, interest rates forgave everything, and capital expenditures were covered by operating cash flow.
In 2025 the figure jumped to $108 billion, and in the first eight months of 2026 hyperscalers threw roughly $194 billion onto the market — up 79% year-over-year, and the year isn't even finished (Reuters data via LSEG). Goldman Sachs expects approximately $250 billion in issuance by year-end and $400 billion in 2027. Vanguard estimates that total AI-ecosystem debt — including chipmakers, data center developers, and energy companies — could reach between $300 and $570 billion in 2026.
For comparison: the entire US high-yield bond market is approximately $1.4 trillion, and it took three decades to accumulate. Last year's new issuer is growing its borrowing faster than old markets gained weight over eras.
Next — the most interesting part: where the money goes and why it changes the very mechanics of the market. The US Treasury had a decades-perfected deal with global savings that economists call "exorbitant privilege": the whole world bought US government bonds not because their yields were exceptional, but because they were risk-free by definition — reserve assets, banking norms, the pricing base for everything. For this the US paid less than any other borrower of such scale would pay; the difference was a tacit subsidy to American power, paid by the rest of the world.
On the long end of the curve, this subsidy always depended on a narrow circle of "natural" buyers of long-dated paper: pension funds, insurers, central banks. And then in 2026 two things hit simultaneously. First: foreign official buyers — central banks and finance ministries — started reducing the share of Treasuries in reserves. Second: on that same long end appeared a competitor offering huge volumes of paper, paying slightly more, and formally no worse by rating — first-class investment grade AAA/AA.
Foreign private investors over the past 12 months put $390 billion net into American corporate bonds versus $329 billion into government bonds — a reversal of a multi-year proportion, calculated by Yardeni Research (Axios data). The capital didn't disappear — it changed addresses.
Yardeni formulated this into a phrase I'm ready to hang on the wall: the textbook used to say "if government borrows too much, it crowds out the private sector, and rates for business rise." In 2026 it's the opposite — hyperscalers are doing the "crowding," but their rates barely rise because buyers believe the AI story and follow them yield-agnostically, in Bessent's own phrasing: corporations "aren't particularly interested in how much they pay."
The result is that the market clears not through more expensive corporate debt but through rising Treasury yields — "Capital flowing into corporate bonds is capital not flowing into Treasuries, and Treasury yields have had to rise to clear the market." I've encountered this effect under the names "reverse crowding out" and "second issuance effect": the same dollar of savings that used to automatically flow into "risk-free" paper now chooses between it and paper backed by the promise of exponential growth in compute.
Thirty-year yields hitting 5.337% in mid-August — the highest since 2007 — this isn't "the market fears inflation." This is the price at which the Treasury wins the auction against data centers for pension fund money.
And here I caught myself looking at July 2007 in a funhouse mirror. Back then there was also excessive debt issued under an "unkillable" story (housing), also long-dated paper, also institutional buyers. The thesis "AI data centers are CDO 2.0" circulates online, and it's simultaneously sharp and imprecise.
Sharp when it comes to shadow: a significant portion of data center financing doesn't hit public markets or balance sheets at all — J.P. Morgan Asset Management counted approximately $1.4 trillion in already disclosed data center lease obligations (Hyperion), of which roughly $1.1 trillion are currently off-balance-sheet, plus another approximately $1.5 trillion in irrevocable procurement contracts for chips, infrastructure, and energy. Individual projects are financed through special-purpose entities: the giant Hyperion data center project raised $27 billion in debt through a "maze of special-purpose vehicles," CoreWeave raises loans for individual centers with non-recourse terms where if the customer refuses to pay, investors distribute losses by seniority and the parent company's assets remain untouchable. This is the anatomy of 2008 down to the last comma: each tranche has its own priority, and all depend on whether the end tenant pays.
Imprecise — when it comes to the quality of the top layer: hyperscalers themselves are — for now — exceptional credits with margins and interest coverage any industrial borrower would envy, as Vanguard emphasizes. The difference between "CDO" and simply "new first-class borrower" comes down to one term: sustainability of tenant cash flows. As long as OpenAI and its clones pay for compute with real money, the structure stands. The question the market has now begun viewing with suspicion: what happens to this pyramid when AI monetization diverges from its ambitions. Oracle CDS spikes and "fatigue" in new issues (78 of 91 hyperscaler bonds in 2026 traded below placement price) — the market has already started asking this question in whispers.
Official Washington's reaction deserves a separate paragraph, because this is a rare case where macroeconomics turns into a political thriller in real time. Bessent — a former macro trader, Druckenmiller's student, someone who understands how the long end works — made intervention his rule: first the failed yen intervention (the market perceived it as capitulation), then a buyback program for long bonds doubled from $2 to $4 billion per operation, then, according to CNBC reports, an idea was being developed to use the Treasury General Account — literally the federal government's checking account with nearly $1 trillion on hand — to finance further buybacks.
Armstrong and Martin from FT Unhedged put it precisely: this isn't quantitative easing, where the central bank creates money; this is the government buying its own debt with tax dollars, and the signal to the market from this isn't "everything's under control," but "Washington is nervous."
The climax — a quote that will enter textbooks: asked whether he had directed Bessent's interventions, Trump answered: "We have many types of intervention — that's one. The ultimate intervention is our military. And if we have to use that, we will." FT made this the headline of their podcast "Bombing the bond market," and although no one seriously discusses the military option, the very fact that the US president says on camera that the army is a tool for working with the bond market — that's a diagnosis. The tools on the Treasury Secretary's desk have run out, and everyone sees it.
Also telling is that Fed Chair Kevin Warsh — a proponent of expensive money being a feature, not a bug: the Fed's silence about the "malfunctioning" Treasury market is the loudest thing it could say. And Druckenmiller publicly dissected his student's policy in a thread that Axios describes as criticism of interference in the "healthy functioning of the market." Teacher against student against the backdrop of $40 trillion — a Hollywood screenwriter wouldn't dare write this.
Now about gold, because without it the dedollarization picture looks like a horror story, but with it — like precise diagnostics. Summer headlines reported that gold for the first time surpassed Treasuries as the world's largest reserve asset, and the internet predictably sang "the end of the dollar." I dug into a fresh Fed researchers' note (September 2026), and it broke down this narrative into molecules: yes, global gold reserves are worth about $5.1 trillion and formally exceeded foreign official holdings in Treasuries, but this is mostly revaluation — the metal's price soared on private demand (ETF inflows since late 2024), not on central bank buying; central bank purchases since 2022 explain only a portion.
Plus a statistical bug: the US holds 22% of the world's gold and can't hold its own Treasuries, which skews the comparison. By June 2026, foreign official holdings in Treasuries again surpassed gold excluding the American bullion.
But the subtle point the note indirectly confirms: EM central banks did buy gold at elevated rates starting in 2022 — and this happens precisely from the moment of freezing Russian reserves, meaning the realization that a "risk-free" reserve asset can be a weapon. An expensive dollar, gold as insurance against weaponization, AI bonds as an alternative to Treasuries — three separate symptoms of one disease: the world has begun diversifying from the single point of failure.
I don't believe in imminent collapse, and honestly, all sources — even the most nervous ones — don't believe in it either: hyperscalers are still paying, Treasuries are still trading, "we're not hitting the panic button" sounds even from those who've seen five crises.
But what happened in summer 2026 is more important than any specific crash: the risk-free rate stopped being an engineering constant and became a subject of bargaining. For half a century the entire financial architecture was built on the tacit assumption that the world pays extra for US "risk-freeness" — and now the US must honestly pay the market for it, in competition with data centers, gold, and fear. I think a future historian of finance will record this summer as the moment when AI finally did what neither Facebook nor Google could: it changed the price of money for the state itself. Top-tier irony: the same industry promising "productivity that will pay off the national debt" has itself become the second-most voracious consumer of global savings.
If the promise is fulfilled — the story will end happily with high rates as a bonus. If not — we'll learn what a financial crisis sounds like when the battle for capital isn't between countries but between the state and a machine that wants to grow. The stakes, as usual, will be set not by Washington and not by Silicon Valley, but by that same narrow circle of buyers of long money. And for the first time in fifty years they're asking not "how much does the Treasury pay," but "what else is on the table."