Hot Tracking 021｜Banks Dump AI Data Center Debt They Just Underwrote: The Bubble Hypothesis Is Being Tested in Real-Time

First, fix the timeline. This time, it’s also two lines. The underlying event was a months-long process: according to Bloomberg, a banking consortium including JPMorgan and MUFG spent over six months trying to distribute $38 billion in data center construction debt; according to FT, the four major banks have been exploring for months how to transfer AI-related debt risks off their books. The second line is the dissemination node: on October 14, an analysis article summarizing these reports topped Hacker News’s hot list, with over 340 upvotes and more than 230 comments by October 15 morning. One statement was firmly established: this isn’t a “today’s default”; the banks’ sale of debt isn’t a breach of contract; it’s a months-long stress test on the AI infrastructure financing chain. The article’s own characterization is also restrained: the AI boom is built on the assumption that credit can remain open under reasonable conditions, and six months of this tug-of-war is the clearest evidence of this assumption being tested in real time, not a conclusion of collapse. All following facts are marked with sources of translation; opinions and characterizations are attributed individually.

What Happened

Core Fact 1: Banks are selling off their own insured debts. According to FT, JPMorgan, Morgan Stanley, SMBC, and MUFG are exploring “significant risk transfer”: the loans remain on bank books nominally, but the most risky layer of risk is transferred to external investors. The motivation, according to FT, is that these banks are approaching the upper limit of their internal holdings in a single AI-related asset. In plain terms: there’s too much money on this bank’s balance sheet tied to the same sector.

Core Fact 2: One debt has been stuck for six months. According to Bloomberg, a banking consortium including JPMorgan and MUFG took on $38 billion in construction debt, corresponding to data center projects in Texas and Wisconsin, leased to Oracle. This debt has been distributed for over six months without being cleared, and some banks are now willing to sell it at a discount to non-bank buyers, just to get rid of it. Just six months ago, according to the source text, the same banks were competing to insure data center loans from Oracle and Meta.

Core Fact 3: Credit pricing is getting expensive. According to Bloomberg’s market data, Oracle’s five-year CDS (default insurance cost) rose to about 151 basis points at the beginning of October, the highest since 2009; S&P downgraded Oracle’s rating to BBB- as early as July, up a notch from junk levels. On a macro level, according to Bloomberg, the yield on 10-year U.S. bonds reached 5.304% this month, the highest since 2002. Previously, $39 billion in government bond auctions were closed at 5.30%, the highest auction yield since 2000. The Fed’s September meeting minutes cited AI-related corporate borrowing as one of the drivers; Capital Economics argued that bigger factors are inflation and government deficits. No matter who’s right, for those currently seeking to finance data centers, the effect is the same: the risk-free rate, the anchor of all pricing, has just become much more expensive.

The components of these numbers are the words of industry insiders themselves. According to FT, Matthew Moniot, co-head of credit risk sharing at Man Group, said that the scale of these transactions “exceeds anything we have considered before.” The person who said this works in pricing such risks.

Banks Underwriting While Selling at a Discount: Where is the Structural Conflict

When you put these three things together, the conflict becomes clear. Six months ago, banks underwrote these transactions with real money, voting that “AI infrastructure financing is a good business.” Now, they explore risk transfer and sell at a discount, showing with their actions that “I don’t want to take on this exposure anymore.” The analysis in the source text is straightforward (this is an article analysis, not factual assertion): Banks won’t sell debts they are confident about.

What deserves more attention is the structural layer. According to the analysis, Oracle, Meta, xAI, and CoreWeave have spread their huge AI infrastructure expenditures to special purpose vehicles (SPVs), totaling approximately $120 billion. The intention was to keep the debt outside their balance sheets and better manage the risk. What happened? The risk didn’t disappear; it flowed back to the banks that originally financed the SPVs, which then transferred the risk to non-bank investors. After this series of structural maneuvers, the risk only changed names and holders. This section’s judgment is from an analytical perspective: The premise for SPVs to disperse risk is that the buyer understands what they are taking on. When professionals who price these risks say that the scale exceeds the model, this premise is already questionable.

It’s important to note immediately: Selling at a discount does not equal defaulting; risk transfer is a normal tool for banks, and selling debts at a discount does not mean panic. The conflict lies not in the act of “selling,” but in the time gap of “hurriedly selling just after underwriting,” and the duration of unsoldness: six months.

Official Position Does Not Match (Verbal and Behavioral Audit)

The audit in this article is not about a company’s official statement, but rather the discrepancy between the words and actions of two groups.

The first group is the banks. Their public narrative remains that “AI infrastructure financing is progressing normally,” with underwriting, lending, and support all in place. In reality, they have transferred the most dangerous layer of risk off their books. Neither side lies; underwriting is a fact, and transferring is also a fact. But when put together, it becomes the standard form of “more words than books.” The motivation reported by FT (approaching internal holding limits) indicates that this is not a single bank being bearish on AI, but rather the entire banking community’s appetite for a single sector, having reached its own limit.

Group two, Oracle. On the expansion narrative: It is reported that Oracle aims to raise more than $40 billion in debt and equity this fiscal year, continuing its support for facilities in Texas, Wisconsin, and New Mexico after issuing $18 billion in bonds in September last year. On the credit signal side: CDS has risen to 151 basis points, its highest since 2009 (according to market data from Bloomberg), with the rating labeled junk grade (S&P, July). Another point is that Oracle has sued the Wisconsin Public Service Commission because a rule may force it to pay over $7 billion in collateral for the Port Washington data center campus, and this conflict has contributed to S&P’s downgrade. While Oracle is expanding its AI business by issuing bonds, it is also seeing insurance costs reach a 15-year high and fighting regulatory lawsuits. This gap is the most sharp part of this article. The market doesn’t listen to stories; it only prices risks, and the current pricing indicates that borrowers’ stories are increasingly not trusted by lenders.

What are the comments saying?

More than two hundred comments on HN can be roughly divided into several factions. Below is a summary, not representing the stance of this article.

The 2008 analogy faction has the most influence. Several commenters directly compared the mechanism to the 2008 financial crisis, with user ceejayoz responding to the statement “This is just normal banking business” by saying: Indeed, sometimes they create global financial crises in this process. User UncleOxidant recalled his feelings in the summer of 2007: The complex operations related to subprime loans already made him feel that something was going wrong, so he shorted a little, making some profit and losing some, but later lost interest. It must be noted: The 2008 analogy is the critical perspective in the comment section, which this article adopts as one of its perspectives, without making a conclusion.

The cooling faction also has many members. User Aurornis believes this is a classic rate fallacy: Selling bonds is originally a banking business. Once sold, new loans are issued, and as long as there’s profit, buyers will be found. Everyone is just led astray by the assumption that “everything related to data centers is bad news.” User lewdwig goes deeper: Any bank underwriting $40 billion in loans would conduct syndicate distribution. Not doing so would be reckless; what’s truly abnormal isn’t selling, but the distribution that usually takes weeks to complete and this time took six months. User rsamtravis bluntly says this sounds like selling anxiety. Banks don’t make money by collecting debt; they make money by issuing loans that can be sold. The reminder from this faction is valuable: Normal mechanisms can coexist with abnormal signals. What’s abnormal isn’t “selling,” but “being unable to sell.”

The price-following faction provides a standard for judgment. User dragontamer says the key is to see at what price the bank sells: In 2007, debt prices collapsed and destroyed the economy. Now bond prices are indeed declining, but they’re not at a panic level yet. User nick486 adds that risk pricing is subjective. If the seller discounts sufficiently, both buyers and sellers can feel like they’ve made a profit.

The perspective of the takeover party comes from user michaeljx, who said he saw an article in FT or Bloomberg a few days ago discussing that now is a good time to buy private equity company credits, implying that non-bank buyers are being encouraged to enter the market. User iririririr questions the quality of the debt itself, believing that Oracle’s large-scale financing shouldn’t be normalized. This is the personal opinion of the commenter, only recorded here.

Three things ordinary users can do

First, don’t read “bank sale of AI debt” as “AI bubble burst.” The article’s own characterization assumes it is being tested in real time, not a collapse judgment; sales, discounts, and risk transfers do not equal default. Look at price signals, not headline sentiment—headlines are common in such topics.

Second, watch a few thermometers more honest than stock prices. One is the subscription multiples of AI bond issuances, which, according to Bloomberg, have shrunk from about 5 times in February to less than 2 times in July. The second is Oracle’s CDS, whether 151 basis points will continue to rise. The third is how much the 38 billion yuan construction debt will ultimately be discounted and cleared. These indicators precede narrative shifts; news headlines are always lagging.

Third, if you hold AI-related stocks or funds, understand that the risk chain is longer than what you see on market software. Approximately 120 billion US dollars in AI infrastructure spending is spread across SPV structures. Your exposure may have been transferred to unknown non-bank buyers through multiple layers, or conversely, the risks transferred by others may be flowing towards your holdings. When you don’t understand the structure, treat your position as “knowing what you don’t know” rather than “having thoroughly studied it.”

Conclusion

Moderation statement: This article does not accuse any party. Banks using risk transfers are legal and conventional tools; sales do not equal a bearish judgment. Oracle’s bond issuance expansion and worsening credit signals can both be true; the downgrade is S&P’s judgment, not ours. The 2008 analogy is for the comment section; this article adopts it as one of its critical perspectives, clearly making no conclusion. The debate over attribution is presented side by side; this article does not judge whether the Federal Reserve memo or Capital Economics is right. Verified facts are: the 38 billion yuan construction debt has been distributed for over six months (according to Bloomberg), four major banks are exploring significant risk transfers (according to FT), Oracle CDS is about 151 basis points (according to Bloomberg’s market data), BBB- rating from S&P, 10-year US Treasury bond yield 5.304% (according to Bloomberg). AI credit is always open—this assumption is being tested in real time. This is the article’s characterization and its tone. Two key points to watch: one, whether Oracle’s more than 40 billion yuan debt and equity financing this fiscal year will be completed as planned, scaled back, or fail—these are harder signals than CDS. The second is whether the 38 billion yuan construction debt will ultimately be distributed and discounted; the clearing price of this debt will become a pricing reference for the entire AI debt market. If any of these occur, v2 will be released.

Reference Sources

• HN story 50092664: Algolia data, 10-14T18:32Z listed, 10-15 morning snapshot 348▲/231💬, text uses “over 340▲, more than 230 comments” format
• StartupFortune aggregate article (published on 10-14, title “Banks are now trying to offload the AI data center debt they just underwrote”): Only for aggregation and citation, all hard numbers in the text marked with FT/Bloomberg/S&P/Fed, CCIR/Capital Economics sources
• Financial Times: Four banks transfer significant risk, Moniot interview (text marked “according to FT reports”)
• Bloomberg: 38 billion construction debt distributed in six months, Oracle CDS 151bp, order book reduced by 5 to 2 times, 10-year US Treasury rate 5.304% and 39 billion bidding (text marked “according to Bloomberg reports”)
• S&P: Oracle rating downgraded to BBB- (July)
• Fed September meeting minutes (AI corporate lending attribution); Capital Economics (counter-attribution), both sides listed in text
• HN comment pool: 32 substantive comments, ceejayoz/UncleOxidant/Aurornis/lewdwig/rsamtravis/dragontamer/nick486/michaeljx/iririririr all self-descriptions by commenters, each attributed in text
• For the original article site recommendation, see the statement “Oracle CDS 203bp” separately, which conflicts with the text’s 151bp, likely an old draft not adopted