$38 billion AI loan required more than 20 banks and months to sell
When JPMorgan and MUFG arranged the largest debt package yet for AI data centers, the financing had to be spread across more than two dozen banks and investors — and months later, some lenders were still holding less than $1 billion unsold. The slow syndication, arriving just as CoreWeave's bonds fell to record lows, offers a concrete picture of how thin the financier roles behind the AI boom really are.
When JPMorgan Chase and Japan's Mitsubishi UFJ Financial Group (MUFG) put together roughly $38 billion in construction financing for two data center campuses tied to Oracle's expansion, the loan agreement itself was not the problem. The problem was selling the debt on. According to Gadget Review, citing Bloomberg, it took several months to spread the package into the market — and well into 2026, some lenders were still sitting with less than $1 billion that had not been placed.
The same calendar stretch included one of the most turbulent weeks for AI-related financial assets in 2026. CoreWeave's high-yield bonds fell to record-low levels, the company's shares dropped by double digits, the cost of credit protection against its debt set a record, and Nvidia-backed Australian Firmus scrapped an IPO that would have been the country's largest in decades.
The two events are not documented as causally linked — but together they show something more important than either alone: the credit market's appetite for AI infrastructure, even with top-tier banks as arrangers and first-priority security, has a practical limit.
The numbers behind the largest package
According to Bloomberg's reporting as relayed by Gadget Review, the package consisted of two senior secured credit facilities tied to Vantage Data Centers:
- Roughly $23.25 billion for a facility in Texas
- Roughly $14.75 billion in Wisconsin
- Total: around $38 billion
Both campuses are tied to Oracle's expansion plans — the company that has become one of the most important capacity suppliers for the major AI labs. Vantage Data Centers, which is building and operating the facilities, is the borrower; the Oracle connection is what makes the projects financially attractive in the first place.
What makes the story striking is not the sum itself, but what happened afterward. A syndication is the banks' exit from their balance sheets: the arrangers lend out or guarantee the financing first, then sell stakes to other banks, institutional investors and funds. The faster and cheaper this goes, the more conventional the risk. This package went slower and wider than large syndications normally do:
- The debt had to be spread across more than two dozen banks and investors, according to Bloomberg via Gadget Review.
- Pricing landed at around 2.5 percentage points over the relevant U.S. benchmark rate — a substantial risk premium for senior secured loans in a market where comparable infrastructure debt is often priced far tighter.
- Less than $1 billion remained unplaced well into 2026.
The exact dates when the syndication began and closed are not given in the sources — only that it took months and that parts of the package were still hanging on far into 2026. It is also worth noting that all specific figures here come via Gadget Review's summary of Bloomberg's reporting, not from primary documents.
Why breadth and price tell a story
Two details in the syndication carry the most informational weight.
The first is the number of participants. When a bank has to involve more than two dozen lenders in a single transaction, it means no single investor — or small group — was willing or able to take on enough exposure. That is not necessarily a sign of panic; $38 billion is an enormous sum for a market where data center debt has traditionally been a niche asset. But it means each new mega-package entails an ever harder placement job. The banks must find ever more buyers, and each new tier of investors usually demands better terms.
The second is the price. Two and a half percentage points over the benchmark for senior secured loans is not disastrous pricing — but it is far from the tight spreads that characterized infrastructure debt when capital was flooding in. Lenders are demanding to be paid for something, and what they are demanding payment for is most likely the combination of size, complexity and uncertainty around the revenue base.
The neocloud model: borrowing against future rental income
To understand why AI infrastructure debt is so sensitive, one has to understand the financing model behind the so-called "neocloud" companies — the data center firms that rent computing power to AI labs. As The Telegraph explains via Yahoo Finance, these companies have financed their growth by borrowing against promises from AI labs to rent future computing power from planned networks of data centers.
The mechanism works like this: a company like CoreWeave signs a deal with, say, an AI lab under which the lab will rent GPU capacity over several years. This revenue stream is used as security, or the basis for loans, that finance the construction of the data center itself. The debt is thus serviced by customers who have not yet paid — and in many cases, according to the Telegraph's description, by agreements that are not binding.
That makes the debt highly sensitive to a single assumption: that demand for AI computing power stays high enough, long enough, for rental income to pay off the loan. If demand falls — or merely belief in it — both the value of the underlying contracts and the market terms for all similar debt fall with it. There is no physical collateral value in the classic sense; the "security" is largely other companies' intentions.
The Vantage/Oracle package differs somewhat from this model — it is construction financing for data center building, and the Oracle connection gives a more established counterparty picture than a freshly minted neocloud contract. But the principle is the same: the values rest on future rental income from a concentrated group of AI players, and the market has begun pricing that in.
The week it all happened: bond rout, share falls and a pulled listing
The syndication was reported in mid-October 2026 — precisely when AI debt came under visible market stress.
CoreWeave's bonds. According to the Telegraph, the price of some of CoreWeave's high-yield bonds fell from 85 cents on the Friday to as low as 77 cents on the Wednesday — a record low. Yields on some bonds climbed as high as 17 percent. CoreWeave is valued at around $42 billion on the stock market and carries $35 billion in debt; on Tuesday, the cost of credit default swaps (CDS) — insurance against default — hit a record high against that debt.
The equity market. On Wednesday, CoreWeave shares fell roughly 15 percent, as part of a two-day AI-driven decline driven by concerns about debt-financed expansion and rising borrowing costs, according to Yahoo Finance. The following Thursday, the market recovered much of the loss as U.S. Treasury yields fell and chip stocks trimmed their losses.
Management's response. CoreWeave's chief executive Mike Intrator told employees on Wednesday that the market was behaving "irrationally," Bloomberg reported, as relayed by the Telegraph via MSN.
The Firmus listing. Nvidia-backed Firmus, which had been aiming for an IPO at a valuation of up to A$44 billion and a raise of up to A$7 billion to build AI data centers in Australia and the Pacific — the largest listing in Australia in decades — withdrew, citing market conditions, according to the Telegraph.
The sequence of events points in the same direction: capital markets have begun demanding significantly better terms to finance AI infrastructure — whether as debt (CoreWeave), equity (Firmus) or bank syndication (the Vantage/Oracle package).
What the autumn sell-off actually proves — and what it does not
Here it is important to be precise about what the sources actually document.
There is no documented causal link between the slow Vantage/Oracle syndication and the CoreWeave rout. They became known in the same period, and both point to the same underlying friction in AI financing — but there is no source-based connection between them, and none of the sources claims one.
Indeed, there are several competing explanations for the sell-off itself, and they are not settled. Rising U.S. Treasury yields and general AI nervousness are one candidate — and the yield decline on Thursday, which coincided with markets recovering, points in that direction. The scrapped Firmus listing is another: the Telegraph reported that the collapse unnerved AI investors, but it is unclear to what extent it actually drove CoreWeave's bond fall, or whether it was mere calendar coincidence. Analysts disagreed on the weighting, and the sources provide no basis for saying one explanation wins.
It is also worth noting that the sell-off was partly reversed by Thursday. This was not a collapse, but a sharp repricing under significant uncertainty.
What to watch
Two things deserve attention going forward.
The first is the pace of future syndications. If the next big data center package sells faster and tighter than the Vantage/Oracle deal, that indicates the market has absorbed this size. If the next packages also take months, require broader participation and higher spreads, that is a clear indication that credit appetite is a hard constraint on the pace of build-out — whatever the AI labs promise.
The second is repricing of existing Oracle-linked credit facilities and other large data center loans. If secondary debt positions in that sector come under further price pressure, it will show up in banks' willingness to take on new, large packages.
It is also worth noting: none of the sources suggests the Vantage/Oracle financing failed. It was placed — but slowly, widely and with a significant risk premium. That is the only claim that can actually be documented: the market can finance AI infrastructure at record scale, but it demands ever more payment, ever more hands and ever more time. For an industry whose build-out plans assume billion-dollar sums around the world every quarter, that friction is not a footnote. It is part of the equation that determines how fast the boom can actually be built.