BLOG
Shiwen Dialogue Episode 30 | Banks Start Offloading Their Underwritten AI Debt: The Financing Window for Startups is Narrowing
Introduction: Telling stories for three years, banks start calculating risks
Shiwen: These days, there has been a lot of discussion about AI finance, with articles on bonds, ratings, and interest rates coming one after another. What angle should we discuss today?
Yongliang: We’ll focus on one angle closest to practitioners: the financing environment. We won’t go into macro interest rates and credit ratings today; we’ll save that for another article. On October 14th, an article on “banks selling off AI data center debt they just underwrote” made it onto Hacker News’ top stories, receiving over 340 upvotes and more than 230 comments by the morning of October 15th.
Shiwen: Let’s clarify the facts: who is selling and what are they selling?
Yongliang: According to FT, four banks—JPMorgan, Morgan Stanley, SMBC, and MUFG—are exploring a transaction called “material risk transfer”: the loans remain on their books, but the most risky part is transferred to external investors. Why? Because they have reached the upper limit of their internal holdings in a single AI-related asset. Half a year ago, these banks were rushing to underwrite data center loans; now, the same banks are trying to get the debt out of their books. This shift is the most worth discussing today.
Shiwen: From telling stories to calculating risks.
Yongliang: Yes, this is my judgment, supported by facts later on. I’ve led a technical team for seventeen years and have seen credit conditions shift from loose to tight. This time, the shift happens within banks, which are closest to money. My personal experience and estimates will be explained in detail below.
Q1: $38 billion sold for six months, still unsold
Shiwen: Is there a specific example showing how stuck this situation is?
Yongliang: Yes. According to Bloomberg, a bank consortium including JPMorgan and MUFG has spent over six months trying to sell $38 billion in construction debt, related to data center projects in Texas and Wisconsin, which will be leased to Oracle after completion. Now, some lenders are willing to sell at a discount to non-bank buyers, only asking for delivery.
Shiwen: What does “six months” mean?
Yongliang: People in the comments who understand the situation have provided a reference: it’s normal for banks to distribute large loans through consortia, and it usually takes a few weeks to settle. What’s unusual isn’t the act of “selling,” but that this deal has been stuck for six months without being sold. This detail distinguishes between “normal business” and “market inability to absorb.”
Shiwen: What do those who take over this risk think?
Yongliang: Matthew Moniot, co-head of credit risk sharing at Man Group, told FT that the scale of these transactions “exceeds anything we’ve considered before.” Note who he is: he isn’t a bearish commentator; his job is to price such risks. The person who priced said there’s no model that can handle this level of risk, which is more significant than ten bearish reports combined.
Shiwen: So what are the banks really worried about?
Yongliang: They’re worried about concentration. There’s a saying circulating in the comments, echoed in the original text: banks won’t sell debt they’re confident about. In plain terms: whether they can handle too many assets in the same industry on their books is a concern, so they need to get the most risky part out first. “Reaching internal holding limits” refers to this: the concentration of AI risks has shifted from risk control reports to real business constraints.
Q2: Money has become expensive, and they’re selective about who buys
Shiwen: What does it mean for banks to tighten on those who borrow money?
Yongliang: The financing window is narrowing, and the tightening is uneven. According to Bloomberg, funds are being reallocated: they flow to data centers rented for investment-grade hyperscalers, such as Microsoft and Google’s leases, away from weak credit entities. The differentiation is clear in prices. CCIR bond data shows that bonds corresponding to CoreWeave leases are priced at 7.00% to 9.875%, while hyperscaler leases are only 5.70% to 7.875%. For the same data center projects, with different names on the leases, the cost of borrowing differs significantly.
Shiwen: What about the demand side?
Yongliang: It is also thinning. According to Bloomberg, bonds from Alphabet, Amazon, Meta, Microsoft, and Oracle are now sold at higher yields than similar-rated non-AI corporate bonds. The subscription multiples are more illustrative: around 5 times in February, less than 2 times by July. HN commentators also cited reports that in September, global AI-related debt financing was about $23 billion, less than half of August. However, some commentators warn that knowing how much was originally planned to be issued, whether buyers rejected it or issuers stopped issuing it, has completely different implications. I present both possibilities here.
Shiwen: Putting these together, what is the picture for an AI startup?
Yongliang: There is a judgment in the original text that I believe is the most important for startups: a smaller, less capitalized AI infrastructure company, if it wants to issue bonds next quarter, is competing with Oracle for the same tight pool of lenders, and without Oracle’s balance sheet and ready customer base. Even a company as large as Oracle takes six months to get loans from banks; what awaits a much smaller company at the same doorstep is clear from the price difference table.
Shiwen: Does this mean AI companies can’t borrow money anymore?
Yongliang: No, they can “borrow but at higher costs and with more selectivity.” Credit hasn’t closed, but the gap is narrowing, and the way it narrows is very elitist: those with investment-grade leases continue to get loans, while those relying on narratives and forward revenue commitments are re-priced. This is my observation, not statistics. Practical implication: if your financing plan includes debt instruments, raise your cost assumptions higher, move up your time window, and don’t bet that conditions will loosen next year.
Q3: GPUs are the hardest to trade among the collateral
Shiwen: Engineers in the comments section are concerned about another issue: what happens to the collateral if the loan isn’t repaid?
Yongliang: This is the most heart-wrenching part from a technical person’s perspective. Some commentators point out that data center GPUs are almost useless outside of AI workloads. Another added a harsh comment: they can’t even connect to standard PCs and servers; they are highly specialized, and if sold, they can only be traded in whole racks or whole servers, to buyers for the same purpose.
Shiwen: In other words, the liquidity of this collateral is extremely low.
Yongliang: Yes. The value of collateral depends not only on how much it’s worth but also on how quickly it can be converted into money. Houses can be lived in by anyone, and general-purpose servers can be used in almost every data center. But those rows of GPUs in AI data centers have a short list of buyers, only companies that bet on AI workloads. When the market reverses, potential buyers and you are the same people, becoming poor together in the same narrative. This is my inference, marked out.
Shiwen: What’s the use of this for tech people to evaluate an AI company?
Yongliang: A specific action: when reviewing the balance sheet of an AI infrastructure company, apply a liquidity discount to the computing power assets. For example, with ten billion in assets, cash, general data centers, and GPU clusters are three different things. The third type is the hardest to convert into cash when money is most needed. Engineers understand hardware: while others see financial figures, you can tell behind them are dedicated cards that can’t be plugged into standard racks.
Q4: The Ghost of 2008, and the Analogies Pressed
Shiwen: What was the most mainstream opinion among over 230 comments?
Yongliang: The 2008 analogy. Many commenters directly compared “bundling and reselling after bank underwriting” to the mechanism of the 2008 financial crisis. Some recalled that they sensed the risk of subprime mortgages being used as down payments in the summer of 2007, and they shorted on a small scale before the real collapse. His conclusion is that prices celebrate when good, but collapse when bad. Others repeatedly cited this sentence: Banks don’t sell debts they are confident about. I quoted these exactly and added: This is an analogy, not a prediction. The value of an analogy is to suggest checkpoints, not to pronounce the outcome.
Shiwen: What did the opponents say?
Yongliang: The cooling camp’s arguments also hold water. First, this is normal banking business: underwriting loans and selling them packaged to those who want that risk. The act itself doesn’t constitute evidence of guilt. Second, the key is the sale price: the debts in 2007 were truly toxic and dragged down the economy; now bond prices are deteriorating, but not to the level of panic. Third, some remind us this is a common basic probability fallacy: Selling debts is banking’s routine, and the market has been pre-trained with the sentiment that “data centers equal bad news,” so anything seen is linked to crisis.
Shiwen: After listening to both camps, what’s your verdict?
Yongliang: The cooling camp’s judgment on the “mechanism” is correct, and the analogy camp’s vigilance on the “price” is correct. Distributing debts is not rare; what’s rare is not being able to split six months, and some lenders prefer to sell at a discount. This price signal is real. I think the wording in the original text is the most accurate: The AI boom is built on the assumption that credit will continue to be available under reasonable conditions; a loan of 38 billion sold for six months is the clearest evidence that this assumption is being tested in real time. It’s a “test” not a “disproof.” We’re seeing it weighed, but we haven’t seen the result yet. Whoever concludes “bubble burst” or “needless worry” at this stage is not giving analysis, but certainty.
Q5: Three sentences for entrepreneurs, tech people, and investors
Shiwen: How should this be applied to specific individuals?
Yongliang: Three sentences for each category—all based on experience, not statistics.
For entrepreneurs: During the narrowing of financing windows, cash flow should take precedence over valuation stories. Debt financing should be discussed now, or you can choose to borrow less and spend later. Banks are approaching internal limit caps, and this usually takes two to three quarters to affect small companies. This is your window to adjust plans, not to wait and see.
For tech people: Add “asset specificity” to your company evaluation. A AI company’s moat is half in model and engineering capabilities, and the other half in what it spends to buy. If it has a general capability team, it can pivot; if it has entire racks of dedicated GPUs, it can only sell them at a discount to competitors. This rule can be used when evaluating employers, options, and career prospects.
To investors: Focus on three verifiable metrics, not emotions. First, whether the bond price differential between ultra-large clouds and low-credit entities is widening or narrowing; CCIR numbers will be updated continuously. Second, whether the subscription multiples for new issues can return to more than 2 times. Third, the pricing of “major risk transfer” transactions—what prices professional buyers like Man Group is willing to pay for the most risky layer, which is the market’s true valuation of AI tail risks.
Shiwen: The last question: What is this matter most likely misread as?
Yongliang: It is most likely misread as “banks are bearish on AI.” No. Banks are still taking insurance, but they don’t want to bear it alone. What they are expressing is not that “AI is bad,” but rather that “AI exposures are too large to handle alone.” Technical direction and investment risk are two different things: the internet direction is correct, but buying at the wrong price in 2000 still led to a decade of loss. This distinction determines whether you should continue to hone your AI skills or simultaneously conduct stress tests on your cash flow. My answer is to do both.
Conclusion
Half a year ago, banks rushed to take insurance, and half a year later, they sold at discounted prices. No accidents occurred in between, only the weight of risk was carefully measured once. This issue does not focus on whether the “AI bubble has burst”—who answers that question now is just guessing; the focus is on the window: the financing window is narrowing, and it’s narrowing layer by layer. Those with strong credit leases continue to use it, while those who borrowed through narratives are being re-priced. Assume it is being tested in real time; during the testing period, the most valuable thing is never judgment, but advance indicators.
Depth: Three areas to follow
- The pricing of risk transfer transactions. “Major risk transfer” has become a common practice for large banks. How high the risk compensation required by the receivers will set the true risk price of AI debt. The next transaction terms are worth paying attention to.
- The secondary market for specialized collateral. If GPU clusters are indeed sold at discounted prices, the transaction price per rack is first-hand data on AI asset liquidity. Currently, there are no public transaction records, and this gap itself is information.
- Follow-up adjustments after September’s halving of issuance. $23 billion is a monthly figure, and the question of “planned volume vs actual volume” has not been answered yet. Only when October and November data come out can we determine whether the issuer has stopped or demand has closed.
Information Sources
- StartupFortune “Banks are now trying to offload the AI data center debt they just underwrote” (2026-10-14, blog aggregation site; this article is not considered primary source): Key facts cited from FT (major risk transfer by banks, internal exposure limits, Moniot’s statement), Bloomberg (38 billion in construction debt not distributed for six months, capital rotation, order book reduced 5 times to less than 2 times), CCIR (CoreWeave at 7.00%-9.875% vs. hyperscalers at 5.70%-7.875%)
- Hacker News story 50092664: Ranked on 10-14; over 340▲ by 10-15 morning, with more than 230 comments; sampled 32 comments, full text omitted without citing usernames
- Around $23 billion in global AI debt financing in September: Reported by HN commenter, text retains “plans unknown” with skepticism
- “Hypotheses are being tested in real time”: Original wording of the source article, strictly followed in this article
- Guest opinions are based on personal experience and estimates, not statistical conclusions