The November 12 record
A $22 Billion Trap With No Exit Button
The most useful number printed this week is not the 16% yield on CoreWeave's 2032 notes. It is 61 cents on the dollar — the average price Bloomberg found for roughly $22 billion of US bank preferred shares that banks sold into the pandemic-era rate collapse. Those issues have given up 8.5% in total return this year and nearly 40% since they were sold, and the buyers cannot leave: the coupons never reset, the banks have no incentive to refinance them, and every call option has so far been skipped.
Preferred shares marked down
$22B
US bank fixed-for-life preferred stock sold in 2020 and 2021, Bloomberg, Nov 12, 2026
Average price
61¢
On the dollar, average indicated price across the group, Nov 12, 2026
Total return this year
-8.5%
Fixed-for-life bank preferreds, 2026 year to date
Decline since sale
~-40%
Pandemic-vintage issues measured from the original issue price
| Instrument | Size / coupon | Exit for the holder | Where it trades |
|---|---|---|---|
| US bank fixed-for-life preferreds sold in 2020 and 2021 | ~$22B group; sub-5% coupons | No call; banks have no incentive to refinance | 61¢ on the dollar (average) |
| Morgan Stanley fixed-for-life issue | $1.3B, with a January call option | Call priced as economically impossible | 61.4¢ |
| U.S. Bancorp 3.75% series | Lowest coupon in the cohort | Call skipped | 55¢ |
| Bank preferreds sold in 2026 | 6.69% average coupon | Fresh issue; the trap is not built yet | Structure out of favor with buyers |
- Every issuer has so far skipped the repayment option that would let investors exit without a loss.
- Fixed-for-life coupons never change, which makes the paper behave like long-duration debt: it falls when long yields rise.
- With the 10-year Treasury back above 5%, banks have no economic reason to refinance sub-5% paper, so the call never comes.
- Regulatory requirements prevent banks from adding clauses that would give investors an incentive to call the securities, so a call happens only when it is free for the lender.
- These were sold with a retail-friendly $25,000 par through banks' own preferred-share offerings, which is why the pain is spread across thousands of holders rather than a handful of institutions.
The holder's mark
CoreWeave's 24.7% Rally Repairs the Issuer, Not the Holder
CoreWeave answered the bond market on November 10 with a strong third quarter: revenue of $3.66 billion, up 168%, a $132.5 billion backlog, and guidance for $37 billion to $39 billion of 2026 capital spending. The stock rose 24.7% on November 11 and the company said it expects to be profitable in 2028. Everything in that release is an issuer's story. The holder's story sits in the other half of the same balance sheet, and no equity rally touches it.
| Metric | Q3 2026 |
|---|---|
| Revenue | $3.66B, up 168% year over year |
| Revenue backlog | $132.5B |
| Adjusted operating income | $379M |
| Interest expense, net | $846M |
| Net loss | $709M |
| Borrowings | $46.9B ($41.6B recourse, $5.3B non-recourse) |
| Lease obligations | $19.3B |
| Stockholders' equity | $4.0B |
| 2026 capital expenditure guidance | $37B–$39B |
CoreWeave's quarterly interest bill has more than doubled in four quarters
Interest expense, net, as reported in CoreWeave's quarterly results. Q4 2025 through Q3 2026.
Unit: $M
Q4 2025
388
Q1 2026
536
Q2 2026
640
Q3 2026
846
- The holder's mark is set by funding cost, not by the issuer's equity. Levered holders of long-duration paper fund with floating-rate borrowings, so a Fed hike raises their costs whatever happens to the credit.
- There is no call on either side of this trade. For the preferreds, the call is a bank refinancing decision it will not make; for CoreWeave, the call is the equity, and $4.0B of book equity sits behind $66.2B of borrowings and leases.
- Duration, not default, is what moved. AI corporate bonds trade about 148 basis points above Treasuries against 77 basis points for the rest of the investment-grade index, with the 10-year above 5% and money markets pricing a 14% chance of a 50-basis-point hike.
- A rally in the issuer's equity lowers the cost of the next equity issue; it does not touch the coupon on the $46.9B already outstanding.
Who actually holds it
The Same Levered Buyers Sit on Both Sides of This Trade
Here is the part of the chain that rarely gets written down. The bank preferreds are not themselves AI credit, and nothing in them says the AI buildout is a bubble. What they reveal is the balance sheet of the institutions that finance AI debt: leveraged funds, closed-end funds and private-credit vehicles whose marks are set by the long end of the curve rather than by any single borrower's fundamentals. AI credit is migrating onto those books because banks have successfully originated far more of it than they want to hold.
AI in the investment-grade index
5.14%
Share of the index as of late October 2026, up from 2.73% at the start of 2025
AI credit issuance expected next year
$685B
Bank of America estimate for 2027, about 31% above 2026 and almost entirely private credit
Hedge fund assets
$13T
IMF estimate, up from $4T in 2013
Borrowing at typical preferred funds
25–33%
Of total assets, mostly floating rate, for preferred closed-end funds
- Prime brokerage is booming on the AI buildout: the five largest US investment banks generated $67 billion of equities revenue in the first nine months of 2026, about $7 billion more than the whole of 2025.
- Goldman Sachs earned more than $19 billion of that, up 57% year over year, and its equities financing revenue alone jumped 96% in the third quarter; Morgan Stanley took roughly $17 billion and JPMorgan about $15 billion.
- The IMF estimates hedge funds hold $13 trillion of assets, up from $4 trillion in 2013; bank clients' net leverage sat at a one-year low in October, with technology stocks now roughly a quarter of their holdings.
- July is the precedent: the hedge fund Situational Awareness collapsed during the summer semiconductor selloff after losses on heavily leveraged AI trades.
- Preferred closed-end funds are the cleanest illustration of the squeeze: they typically borrow 25% to a third of assets, mostly at floating rates, so a Fed hike lands twice, once on portfolio prices and once on the funding line.
Bank revenue is being earned from the leveraged AI bid
Equities revenue, first nine months of 2026. The five largest US investment banks generated $67B; Goldman Sachs, Morgan Stanley and JPMorgan accounted for $52B of it.
Unit: $B
Goldman Sachs
19
Morgan Stanley
17
JPMorgan
15
The transmission belt
Banks Are Successfully Lending and Failing to Find a Holder
The mechanism that moves AI risk onto those holders is mundane and well documented. Banks originated more AI construction debt than they wanted to keep, so they went looking for buyers. JPMorgan and Mitsubishi UFJ Financial Group arranged roughly $38 billion of senior secured construction facilities for two Vantage Data Centers campuses linked to Oracle, priced about 250 basis points over benchmark. Spreading the paper took months and involved more than two dozen banks and investors, and some lenders were still trying to place less than $1 billion well into 2026.
| Facility | Amount | Terms |
|---|---|---|
| Texas data center campus (Vantage / Oracle-linked) | $23.25B | Senior secured construction facility |
| Wisconsin data center campus (Vantage / Oracle-linked) | $14.75B | Senior secured construction facility |
| Total arranged | ~$38B | JPMorgan and Mitsubishi UFJ as arrangers; priced about 250bp over benchmark |
| Still unplaced with lenders | Under $1B | Unsold well into 2026 |
- Traditional bank balance sheets would not hold the paper, so private credit funds and non-bank investors became the buyers, demanding higher yields and tighter covenants and shifting pricing power away from the underwriters.
- Those buyers take construction risk: completion, permitting, power availability and equipment costs sit with them until the campuses operate and the leases begin.
- The banks' own capital market is closing at the same time. The 2026 vintage of fixed-for-life bank preferreds averages a 6.69% coupon and that structure has fallen out of favor, while JPMorgan is redeeming $2.0 billion of its 3.65% fixed-rate reset preferred stock.
- AI lending grows risk-weighted assets, so banks need capital precisely when the buyers of bank capital are impaired, and those same buyers are now the ones holding the AI paper.
Horizons
The Fed Path Is the Only Exit Left
Two horizons matter, and they resolve in opposite directions. In days to quarters, nothing in AI credit changes unless the discount rate changes: the 10-year Treasury topped 5% on November 12 and money markets priced a 14% chance of a 50-basis-point hike after hotter-than-expected October inflation, with the two-year yield up as much as 22 basis points to 4.78%. Over one to three years, the buildout needs roughly $440 billion of hyperscaler debt in 2027 on Goldman's count plus about $300 billion of project-level paper, and Bank of America expects total AI-related credit issuance to peak near $685 billion next year, almost all of it private credit. That pipeline only clears if the leveraged bid stays open, and the leveraged bid is what the rate shock has impaired.
- Watch the new-issue concession on any CoreWeave bond. The 2032 notes yield above 16% against a 9.625% coupon, and five-year default swaps on the company spiked above 1,289 basis points before easing to about 1,155.
- Watch the clearing coupon on new bank preferred stock. The 6.69% average on 2026 issues is today's price of bank capital; a further rise would say that AI lending is being funded by expensive capital.
- Watch bank earnings for which side of the trade the arrangers are on: financing and placement revenue tells you whether AI paper is being sold or warehoused.
- Watch default-swap pricing at the AI complex. Protection costs rose76% at Nvidia and 49% at Broadcom over the past month, against roughly 44% at Amazon, which shows the stress moving from neoclouds toward the balance sheets that fund them.
- Concentration is the vulnerability, not leverage in the abstract. The same ~$38 billion of Vantage and Oracle facilities and CoreWeave's $46.9 billion of borrowings sit with a small set of holders whose marks all move with one rate.
- If the bid stays impaired, absorption fails and capital spending slows before any default happens; Goldman estimates a state-level moratorium could put 45% of planned US capacity at risk, so policy can arrive faster than capital.
- For banks, the slow variable is the cost of capital: a closed preferred market raises the price of every dollar of AI risk-weighted assets they carry, while Oracle already runs total debt of $155.9 billion against $67.2 billion of equity.
- For issuers of AI paper, the milestone to watch is not revenue but the spread at which new debt clears. Anything above the 16% currently demanded from CoreWeave resets the entire buildout's math.
Bottom line
The Convexity Sits With the Holders, Not the Issuers
Fact: roughly $22 billion of US bank preferred shares sold in 2020 and 2021 trade at an average 61 cents, are down 8.5% in total return this year, and carry no usable call. Inference: the same long-duration, no-exit structure now defines the AI funding chain, where CoreWeave's borrowings of $46.9 billion carry a yield above 16% and the marginal buyers are levered intermediaries whose marks are set by funding cost rather than by any issuer's fundamentals. Speculation: if those buyers' books keep tightening, AI paper has to clear at wider spreads than the $685 billion of issuance everyone is planning for, and the first casualty will be project finance rather than the hyperscalers.
Really few places to hide in fixed income credit this year. US HY up 0.8% and mostly everything else in the red (except floating rate credit). Bank fixed-for-life preferreds are at -8.5%. Yikes.
The practical consequence is that AI credit stress should be read off the holder's funding line, not off the issuer's spread. For issuers, the question is what a new bond clears at. For holders, the question is what their own borrowing costs and net asset value discounts are doing while they wait for a rate cut that will not come on demand. And for anyone buying the buildout's equity because the operating numbers are spectacular, the relevant figure is not the $132.5 billion backlog but the $66.2 billion of borrowings and leases standing behind $4.0 billion of book equity.
Where the exposure actually sits
- Third-quarter revenue of $3.66B and a $132.5B backlog prove demand, but $846M of quarterly interest expense on $46.9B of borrowings means the credit sets the cost of capital, not the operating story.
- The stock rose 24.7% on Nov 11 while the 2032 notes still yield above 16%; the rally repairs access to equity, not the coupon on $46.9B of existing debt.
- Near term: the new-issue concession on the next CoreWeave bond is the real verdict on the capital structure.
- One to three years: the 2028 profitability path depends on AI credit markets staying open at double-digit yields.
- Ratings fees scale with new AI issuance, and Bank of America expects AI-related credit issuance to peak near $685B next year, about 31% above 2026.
- Moody's reported $2.02B of revenue and $734M of net income in the third quarter of 2026, and every new CoreWeave, Oracle or data-center securitization adds to the same line.
- One to three years: the toll is collected on issuance whether the AI credit cycle ends in boom or in downgrades, which makes it the cleanest way to own the buildout without owning the risk.
- An enterprise value of roughly $3.8B against a $2.6B market capitalization implies about $1.3B of borrowings ranking ahead of shareholders.
- The portfolio is the bank paper trading at 61 cents on the dollar, so the same mark that traps the retail holder runs through net asset value and the distribution rate.
- It feels a Fed hike twice, once on portfolio prices and once on floating-rate borrowings, and a further rate rise would widen the discount to net asset value from its current 3.36%.
- Watch the next distribution coverage disclosure: below100% coverage with rising borrowing costs is what forces a cut.
- Equities financing revenue jumped 96% in the third quarter of 2026 and nine-month equities revenue topped $19B, up 57%, because the leveraged AI bid is a direct revenue line.
- That revenue is cyclical in the wrong order: it peaks when leverage peaks, and July's collapse of a heavily levered AI fund showed how quickly it can unwind.
- Goldman also keeps legacy sub-5% preferred stock outstanding at 61 cents, which is the cheapest capital in the system, while issuing new fixed-rate reset preferred at 6.5%.
- Third-quarter 2026 revenue of $76.2B and net income of $14.9B, with the bank still redeeming $2.0B of its 3.65% preferred stock on its own terms.
- It helped arrange roughly $38B of Oracle-linked construction facilities that took months to place, with under $1B still unsold into 2026, which is balance-sheet drag that has not cleared.
- One to three years: if the AI bid narrows, the arranger keeps the assets and the distribution business slows at the same time.
- It trades its own $1.3B fixed-for-life preferred issue at 61.4 cents, the clearest single read on what the market will pay for bank capital that cannot be called.
- As an arranger and preferred issuer, it earns fees on a $38B distribution it cannot fully place and issues new capital into a market averaging a 6.69% coupon.
- If the preferred market reopens below 6.69%, bank capital gets cheaper and the buildout accelerates; if it stays shut, risk-weighted asset growth becomes more expensive to fund.
- It co-arranged roughly $38B of Oracle-linked data center construction facilities, the clearest case of a non-US bank warehousing US AI credit.
- Placement took months with more than two dozen banks involved, and lenders were still placing under $1B into 2026.
- Near term the exposure is balance sheet rather than credit: the paper is investment grade at issue but stays on the bank's books until it is sold.
- Blue Owl agreed to help fund a $15B OpenAI data center project, the template for how private credit absorbs what banks originate.
- Banks now treat private credit funds like Blue Owl as the buyer of last resort for AI construction risk, paid in higher yield for taking completion and power risk.
- One to three years: if the leveraged bid stays impaired, deployment and fundraising slow before management fees do, and the $440B 2027 hyperscaler pipeline stops clearing at current spreads.
