Morgan Stanley capped withdrawals at its $7 billion private credit fund for a third straight quarter after investors asked to pull 11.4% of shares. In the same week, the AI buildout’s funding need peaked — OpenAI projected US$278 billion of negative free cash flow through 2030, and central banks raised rates. The repricing arrives through the funding channel before the equity tape.
Two stories that are not usually in the same sentence: the cost of capital turned up, and the AI buildout’s borrowing requirement peaked at the same moment.
| Date | Event | Level / Size |
|---|---|---|
| Sep 14 | Speculators turned net-long the yen for the first time since February on BOJ hike bets, as rate-hike expectations spread across G7 central banks (BT, Reuters)[1][2] | net long |
| Sep 15 | The US 10-year Treasury yield topped 5% for the first time since 2023 as inflation fears mounted (BT)[3] | >5.0% |
| Sep 16 | The Federal Reserve raised rates 25bp to a 3.75%-4.00% target range — its first increase since 2023 — on a unanimous 12–0 vote; the projections showed 16 of 18 officials expecting at least one more hike this year (CNBC, Reuters)[4][5] | +25bp, 12–0 |
| Sep 16 | OpenAI was reported to be weighing a funding round at a US$1.2 trillion valuation ahead of an IPO (FT, via BT)[6] | US$1.2T |
| Sep 18 | The Bank of Japan raised its policy rate 25bp to 1.25% — the highest since April 1995 — on a 7–2 vote, pivoting to a pre-emptive inflation fight (Reuters, CNBC, Guardian)[7][8] | 1.25% |
| Sep 18 | Redemption requests at BlackRock’s US$23.1 billion HPS Corporate Lending Fund came in at roughly 11.5% of shares, against 13.3% the prior quarter, per a regulatory filing (BT, WSJ)[9] | 11.5% requested |
| Sep 18 | Morgan Stanley’s North Haven Private Income Fund capped quarterly withdrawals at 5% after investors asked to pull 11.4% of shares, largely unchanged from the prior quarter (shareholder letter, via BT, Bloomberg)[10][11] | capped at 5% |
| Sep 18–19 | Family offices were reported to be increasing allocations to AI and data-centre investments as the capex bill grows (TechCrunch)[12]; OpenAI was reported to project US$278 billion of negative free cash flow from 2026 to 2030 (FT, via BT)[13] | US$278bn FCF |
The obvious link is the rates side: a Fed hike, a 10-year at 5%, a BOJ at a 31-year high. The less obvious one is that the same asset class being asked to finance the buildout is the one where investors are queuing to leave.
Private credit funds offer quarterly liquidity with a cap — typically 5% of shares per quarter. The underlying assets are direct loans that do not trade. When requests exceed the cap, the fund repurchases what it can and the remainder rolls into the next window. It is a queue, and this week the queue got longer in public.
1. The Morgan Stanley letter is the clearest document of the week. Requests covered 11.4% of shares; the fund bought back 5%. Completed repurchases across three periods through September will total about US$479 million. The detail that matters is composition: nearly two-thirds of the latest requests came from investors whose earlier withdrawal attempts had already been limited in the prior two repurchase offers. That is not a one-off redemption. That is a backlog compounding at 5% a quarter against demand running at twice that rate.[10][11]
Source: Business Times — “Morgan Stanley caps private credit exits again as 11% want out” (20 Sep 2026).
2. The gate is not the story. The gate is the disclosure. A private credit fund’s net asset value is an appraisal, not a transaction. No buyer and seller agreed on the price of a middle-market loan this quarter — an agent valued the book, and that valuation determined what the queue is being paid against. Public equities clear continuously, so a disagreement shows up as a price. Here, a disagreement shows up as a queue length. When two-thirds of the queue is repeat demand, the market is telling you something the NAV has not yet been marked to reflect.
3. This is not new, and that is the point. In March, BlackRock limited withdrawals at its HPS Corporate Lending Fund after requests reached 9.3% of net asset value, paying out US$620 million at the 5% threshold. Blackstone raised its usual 5% cap to 7% on an US$82 billion fund and put US$400 million of firm and employee money in so that every request could be met. Six months on, requests at the BlackRock fund are still above 11% and Morgan Stanley is capping for a third consecutive quarter. The mechanism did not resolve the pressure. It rescheduled it.[14][15]
4. Why this belongs in a memo about AI. Illiquid credit is where the buildout’s financing went: beyond the bond markets, into private placements, project structures and directly originated loans that never print a price. The financing of the buildout and the exit door of private credit are the same asset class — same managers, same capital pool, same structural mismatch between long-dated assets and quarterly redemption rights. Rate increases make both ends worse at once: refinancing gets more expensive, and the discount rate applied to an appraised book goes up.
The numbers on AI-related borrowing are now large enough to sit alongside the capex totals rather than beneath them.
1. Roughly half a trillion dollars already issued this year. Goldman Sachs Research estimates nearly US$500 billion of AI-related debt issuance so far in 2026, across the whole ecosystem — chipmakers, developers, utilities and the hyperscalers themselves. The hyperscalers were only about 40% of that. Their own issuance ran to US$194 billion globally year-to-date, against US$108 billion for all of 2025. Goldman’s estimate for the year: roughly one-third of hyperscaler capital expenditure debt-financed, or about US$250 billion of direct supply. Vanguard’s range for total 2026 AI-related issuance runs from about US$300 billion to US$570 billion depending on where the ecosystem boundary is drawn.[16][17]
2. Project finance is where the off-balance-sheet capacity sits. Of the roughly US$950 billion issued in the project finance market in 2025, about US$170 billion was data-centre related — a 57% increase on the prior year, according to IJGlobal.[18] The structure is familiar from infrastructure: a developer or special-purpose vehicle owns the facility, the hyperscaler signs a long lease, and lenders are repaid from the lease rather than the tenant’s balance sheet. That keeps leverage off the ratings-agency scorecard and puts it into a pool of creditors with their own liquidity terms.
3. Private credit is now a measured part of the AI balance sheet. The Bank for International Settlements counts more than US$200 billion of outstanding private credit loans to AI-related companies, with a range of US$300 billion to US$600 billion by 2030. The BIS is explicit about the direction of travel: AI firms, traditionally reliant on internal cash flows and equity, now face higher leverage, “which could amplify shocks and affect the health of financial intermediaries.”[18][19]
4. The funding leg turned more expensive this week, not less. The Fed’s 12–0 hike to 3.75%-4.00%, with 16 of 18 officials projecting another increase by year-end and futures pricing roughly a 90% probability of one, resets the discount rate on every appraised book. The BOJ at 1.25% removes the last of the world’s free funding. The 10-year at 5% sets the alternative return that a 5%-gated, appraisal-marked fund has to compete against — and 30-year mortgage rates above 7% show how quickly that flows into the real economy.
Put the two halves together and the memo writes itself: the buildout’s funding requirement peaked in the same week that the price of funding rose, in a channel that cannot clear its exits at the stated marks. That is a mechanical problem, not a sentiment one — and mechanical problems surface in the credit data before they surface in the equity tape.
The transmission channel is not “AI stocks fall.” It is “the cost of marking an illiquid book rises while the exit stays capped” — and the portfolio implications follow from there.
1. In this channel, credit leads equity. Gating frequency is a free, quarterly, public signal of mark integrity. A fund that pays out at 5% while requests run at 11% is telling you its stated valuation has not been tested by a transaction. If the request rate stays elevated for a third or fourth quarter, the market’s opinion of the marks is on the record — and equity investors in the same theme will be reading it after the fact. This memo’s predecessor on the capex reckoning argued that borrowed money meets real returns eventually; this is the plumbing underneath that argument.
2. Anything marked to appraisal should be sized as if the exit is a queue. For a Singapore-based investor, private credit and data-centre exposure arrives through listed vehicles, platform feeders and credit funds — instruments that report quarterly. The practical adjustment is not to avoid the asset class; it is to stop treating the reported NAV as a liquidation value. Assume a 5% quarterly exit and ask whether the position can be held through four quarters of a gate.
3. Singapore’s banks are on the other side of the same hike. The Fed move was read locally as a margin tailwind, with UOB flagged as the likely biggest beneficiary of higher rates.[20] That trade has a credit-quality leg: the same higher rates that widen net interest margins increase debt-service costs for leveraged borrowers, including the ones behind data-centre and direct-lending structures. Bank margins improve on the way in; provisions show up on the way out. Both parts belong in the same analysis.
4. The IPO window is the pressure valve, and it is being asked to do a lot. OpenAI weighing a round at a US$1.2 trillion valuation ahead of a listing is the equity market being invited to absorb the buildout’s funding need. Family offices were reported this week to be increasing AI allocations as the capex bill grows. If that equity demand holds, the pressure stays contained. If it stalls, the funding leg has to carry a requirement it was never sized for — which is precisely when appraisal-based marks get tested in public. This is the risk the capital share memo described from the household side: the gains accrue to whoever owns the capital, and so do the losses.
5. The carry regime has already changed underneath this. A BOJ at 1.25%, net-long yen positioning for the first time since February, and a 10-year Treasury above 5%. Together, these mean the cheap-financing assumption that funded leveraged structures of every kind — including private credit — is now a variable rather than a constant. The thirty-year line memo tracked that break when it arrived in JGBs; this week it showed up as the cost of every appraised book.
The case against this reading is strong enough that it deserves to be stated plainly, because the overlap between the two halves of the story is a channel, not a shared portfolio.
1. The funds that gate are not necessarily the funds that finance data centres. North Haven and the HPS Corporate Lending Fund are retail-facing, redeemable vehicles. Much data-centre debt is held by insurers and pensions in private placements, bought to hold for 15 to 20 years to match long liabilities, with no redemption terms at all. Those books do not have a queue, and a queue at the retail door is not evidence that AI loan marks are wrong. The honest version of this memo’s thesis is that both sit inside the same asset class and the same managers’ capital pools — not that they are the same loans.
2. The AI credit itself is not distressed paper. Hyperscalers are investment-grade issuers with profitable existing businesses, and much of the project finance is backed by long leases from them. The BIS flags rising leverage from a low base, not a solvency problem. The 2025 precedent is informative: data-centre loan volume rose 57% in a year and spreads absorbed it. Demand for the paper has not disappeared — family offices were reported this week to be seeking it out.
3. The gate is a feature of the design, and the design is small. Private credit cannot fund 20-year assets with daily liquidity; the quarterly cap is what allows patient capital to do patient work. The specific fund under scrutiny is about US$7 billion against a private credit market estimated at roughly US$1.8 trillion to US$2 trillion.[21] Requests of 11% at two funds is a signal worth reading — it is not yet a systemic funding event. Treating it as one would have been wrong in March as well, when the same headlines appeared and the market kept functioning.
4. Higher rates can coexist with a resilient buildout. The Fed hiked on a 12–0 vote into an economy its own projections describe as sound.[4] If growth holds, AI-linked cash flows keep meeting debt service and the funding leg simply reprices like everything else. The bear case requires a credit event, and a credit event requires either revenue disappointment or a refinancing wall that cannot be met — neither of which is in the data yet. A reasonable reader can take this week as rate normalisation, not as the beginning of a mark-down cycle.
The signals here are unusually cheap to read: they are quarterly, public, and quantitative. Four items matter over the next three months.
1. Next quarter’s request rate at the three gating funds. Morgan Stanley’s North Haven, BlackRock’s HPS Corporate Lending Fund and Blackstone’s flagship fund disclose requests against the cap every quarter. Requests falling back below 5% would say the marks held and the queue cleared. A third consecutive quarter above 10% would mean the market has marked the book lower than the fund has — and the queue is where it is being expressed.
2. Data-centre debt spreads and new-issue concessions. Goldman noted concessions widening to as much as 20 basis points on one large hyperscaler deal during a period of indigestion, against 2–3 basis points in easier conditions.[16] That spread is the price of the funding leg. Widening means the buildout’s marginal project needs more equity or does not get built; tightening means the market has reabsorbed the supply.
3. Whether the Fed delivers the second hike. Sixteen of 18 officials project another increase by year-end and futures price roughly a 90% probability. Each increment raises the discount rate applied to appraised books and the refinancing cost of everything floating.[5] A pause would relieve the funding channel first — which is the point: this story is decided by rates more than by AI adoption.
4. Whether the IPO window stays open. OpenAI’s reported US$1.2 trillion round is the equity market’s chance to carry the funding requirement.[6] If the round and the listings behind it are absorbed, the credit pressure stays a queue rather than a markdown. If the window closes, the difference between an appraised NAV and a clearing price stops being academic — and the funds that reported 11% request rates will report what the loans are actually worth.
The title is the mechanism. A fund that caps redemptions cannot demonstrate what its assets are worth, because the test is the transaction it is refusing. That does not make the marks wrong. It means nobody currently knows, and the queue is the only price available.