Banks that financed the AI data center boom now want out of that debt

The lenders who built the AI data center boom's loan book are trying to sell pieces of it before anyone else notices the price is moving.

JPMorgan Chase and Mitsubishi UFJ Financial Group have spent more than six months trying to offload a record $38 billion construction loan tied to Oracle-linked data centers in Texas and Wisconsin. According to reporting on the syndication, some lenders are now willing to sell slices of that debt to non-bank buyers at a discount, just to get it off their books. More than two dozen banks and investors already joined the deal. There's still less than $1 billion left to place, but the fact that anyone needs a discount to move the last scraps tells you where sentiment has drifted.

The Texas and Wisconsin projects were developed by Vantage Data Centers for Oracle, to power OpenAI's compute needs. They were structured as two senior secured credit facilities: $23.25 billion for Texas, $14.75 billion for Wisconsin, priced about 2.5 percentage points over the benchmark rate with four-year maturities. It was supposed to be the clean version of AI infrastructure lending: investment-grade sponsor, hyperscale tenant, collateral you could point to. Banks are still having trouble selling it.

While that $38 billion syndication grinds along, CoreWeave has become the market's preferred stress gauge for the whole AI lending complex. The company priced bonds due 2032 at 9.625% back in June. By mid-October those same bonds were trading at a record 14.5% yield, according to reporting cited by Octus. That's a jump of roughly five percentage points in four months, with no change to the underlying bond - just a change in how much the market trusts getting paid back. CoreWeave stock fell about 13% in a single session around that move.

Credit default swaps on CoreWeave tell a similar story with more drama in the middle. Spreads spiked to 8.81 percentage points in December 2025, the market pricing in real default risk, before easing back toward roughly 4.5 points more recently. That's still a distressed-adjacent level for a company HSBC has flagged for high funding needs and rising debt costs. CoreWeave's total debt reached $35 billion as of June 30, up from $21 billion at year-end 2025. Interest expense alone hit $640 million last quarter, north of a quarter of revenue, while the company stays cash-flow negative even as its top line grows.

Here's the thing about a lender trying to sell debt at a discount: it's not an act of panic, it's a confession. You don't cut the price on something you still believe in.

Capital is already voting with its feet. Investors are rotating toward data center debt backed by investment-grade hyperscalers like Microsoft and Google, and away from weaker credits further down the chain, including the kind of debt-heavy, single-tenant-dependent structures CoreWeave and some of the newer AI-infra borrowers represent. That's the quiet part of this story: the banks that underwrote the boom aren't betting the whole thing collapses. They're just no longer willing to hold the riskier tranches themselves.

The IMF put a number on what's at stake. Its October 13 Global Financial Stability Report warned that weaker or less durable AI investment and profitability could trigger a correction in AI-linked equities and credit. Its stress tests found that institutions holding 22% of global banking assets would fall into a "weak" capital category if an AI disappointment hit alongside elevated rates and an energy shock. The Fund also flagged that hyperscalers have funneled roughly $300 billion in equity and $650 billion in debt into weaker AI companies, with partners committing $1.4 trillion in future revenue. That's exactly the kind of interlocking exposure that turns one bad quarter into a system-wide problem.

None of this means the lending stops. Blue Owl's Stack Infrastructure was still out seeking a $5.9 billion syndicated loan for a Melbourne project as recently as July, and a Morgan Stanley-led consortium prepared a $15 billion bond sale in August to shed pre-built AI construction risk of its own. The money keeps moving. It's just moving faster toward the exits than it used to, and the banks doing the underwriting are the ones who'd know first.

Also read: ASML says chipmakers are already booking tools for 2029 and 2030 • South Korea's Kospi sinks 7.57% as chip fears compound a brutal week • TSMC posts record profit as AI chip demand defies market debt fears

This article is posted in Financial Markets News, check it out for more related stories.