Why the AI Credit Crunch Is Already Here

The AI Boom Hits a Wall It Didn’t See Coming

The narrative that drove tech stocks through most of 2025 was simple: AI requires enormous capital, so whoever can borrow the most wins. The assumption was that lenders would keep handing out money at reasonable rates because the demand side — hyperscalers, data center developers, neoclouds — was too eager to spend to say no.

That assumption is unraveling.

Ten-year U.S. Treasuries are trading above 5.3% and continuing to sell off. That may sound like abstract macroeconomics to most readers, but it matters directly for every AI data center project that needs to issue debt. When the risk-free benchmark rate climbs, every spread above it becomes more expensive. And spreads are widening at the same time. The double squeeze is hitting AI borrowers harder than anyone outside fixed-income circles seems to realize.

Oracle is the canary. Its $18 billion September bond sale cost roughly $7 billion more in interest than it would have at the time of issuance, purely because both Treasury yields and Oracle-specific spreads moved against it. Back then, spreads sat between 105 and 165 basis points. Today they range from 171 to 282 basis points. Oracle is investment-grade. It is still paying a penalty. CoreWeave, which doesn’t carry that same rating cushion, is pricing its debt at spreads between 672 and 882 basis points — effective yields of 11.53% on its shortest maturities and 13.49% on its longest. These are numbers that make project finance viability genuinely uncertain.

Banks Are Getting Picky. Again.

According to reporting from The Information, bond markets are hardening for lower-rated borrowers. CleanSpark, which is developing a data center for Meta, had to offer investors significant concessions earlier this month to close its financing. Major lenders — Société Générale, Sumitomo Mitsui Banking Corp., and Mitsubishi UFJ Financial Group — are becoming more selective on data center loans, people familiar with the deals said. These are the same institutions that lined up for seven and seventeen data center transactions respectively at the tail end of 2025, according to Ed Zitron’s end-of-year analysis.

The shift from eager participant to cautious gatekeeper happened fast. Blue Owl’s $10 billion commitment to AI data center projects alongside Primary Digital Infrastructure was reportedly approved in fifteen minutes during their first in-person meeting two years ago. That kind of speed signals enthusiasm, not due diligence. When the funding environment warms, deals move quickly. When it cools, the same deals look very different.

The Math No One Wants to Do Out Loud

Here is the uncomfortable arithmetic: Anthropic sits on $413 billion in non-cancellable compute contracts. OpenAI, when it eventually goes public, will need at least $50 billion in debt annually just to fund its growth trajectory. CoreWeave needs $102 billion through 2030. SB Energy alone has $174 billion in debt to raise for its OpenAI data center project — debt that carries NVIDIA backstop language but still faces the same market reality.

At 9% to 14% borrowing costs, the economics of new data center projects become questionable even before you account for power delays, local regulatory pushback, or the possibility that hyperscaler utilization doesn’t fill out fast enough to cover the interest payments. The $32 trillion data center spending projection for 2050, frequently cited as justification for today’s buildout pace, looks increasingly like fiction dressed in engineering terminology.

There is a structural problem beneath the cyclical one. U.S. Treasury issuance is on a trajectory that guarantees more competition for investor capital. The federal deficit sits around $2 trillion annually, and debt service payments now consume 14% of federal spending. Any scenario in which the government cuts Social Security, Medicaid, or military spending to reduce that deficit is politically implausible. That means Treasury supply keeps growing regardless of what the Federal Reserve does or what Treasury Secretary Janet Yellen decides about auction timing.

Who Wins. Who Loses.

The winners right now are investors with cash who can pick deals at elevated yields. The losers are the companies that priced their borrowing based on 2024 assumptions and are now facing 2026 reality. Project Jupiter, Oracle’s flagship data center initiative, has already required a force majeure notice tied to power constraints. If similar projects stall because debt refinancing costs make the math impossible, the ripple effects extend beyond individual balance sheets — they reach into the semiconductor supply chain, the construction labor market, and the grid capacity calculations that underpin the entire AI infrastructure thesis.

Hyperscalers like Google, Amazon, and Meta are expected to issue roughly $400 billion in bonds in 2027. Their investment-grade ratings give them pricing advantages that CoreWeave and similar borrowers simply do not have. The gap between what Google pays and what a virgin data center project pays is the gap between a project that gets built and one that doesn’t.

The next few months will clarify whether this is a temporary spike in rates or the beginning of a structural repricing of AI infrastructure risk. Ten-year Treasuries are reacting to inflation expectations driven by supply chain disruption from the wars in Ukraine and Iran, to fiscal uncertainty, and to the sheer volume of debt coming to market. None of those headwinds are disappearing. If anything, they are compounding.

The AI capital rally is not over. But the party is about to get more expensive for everyone except the people who already own the building.