Private Credit Ran Out of Middle Market. Now It's Building Data Centers.

In Brief

The loud worry in private credit right now is middle-market default. Loose-covenant deals from 2021, borrowers paying in IOUs instead of cash, a major fund freezing redemptions back in March. That story isn't wrong. I just think the bigger one sits a layer above it. The same lending machine has moved into AI infrastructure, where the asset wears out in three to five years and the loan is really financing the next loan. It holds together only as long as a buyer for the next round keeps showing up.

A rotation that doesn't fit the asset

Most of the private-credit anxiety this year has been about the middle market. That isn't the most useful frame anymore.

A Hana Securities credit outlook this week put a label on the shift I keep coming back to: private credit's next chapter is data centers, not dental-clinic roll-ups. The machine that absorbed leveraged buyouts after 2008 and grew into a $2 trillion business has found a new borrower. The catch is that the borrower doesn't match the structure that was built around the old one.

A bridge or a toll road depreciates over thirty or fifty years and pays itself off with steady cash flow. A GPU rack depreciates in three to five. The shell of a data center lasts; the chips inside don't. So the loan isn't really paying off the asset. It's paying off the refinancing of the asset, on a treadmill that has to keep moving.

What a single deal looks like

Take Meta's Hyperion data center as the example. Meta owns 20%. Blue Owl owns 80%. The vehicle that holds it sits off Meta's books — EY flagged the question in its audit as something they spent real time on and weren't fully comfortable with. On top of that vehicle: senior debt, mezzanine, a construction loan. On top of that: Blue Owl's infrastructure fund, which has its own borrowings against its equity stake.

One underlying asset. Three layers of borrowing stacked on it. Each layer marked by a different counterparty using a different model.

Private credit hasn't extended duration here. It's extended the cycle itself.

The banks didn't actually leave

There's a story I hear all the time: after 2010, banks pulled back from leveraged lending and credit funds filled the gap. That's half right. The Kansas City Fed put a number on it. US bank lending to non-bank financial firms — the polite name for private credit funds, BDCs, and PE — is now around $1.4 trillion. Banks earn 29% return on that kind of lending versus 8% on direct corporate loans.

That spread bends behavior. The bank doesn't lend to the company anymore. It lends to the fund that lends to the company. Same exposure, one extra layer of indirection, less capital required against it.

You see the same shape in data center finance. The bank doesn't underwrite a GPU farm. It underwrites the credit fund that underwrites the entity that owns the GPU farm. Three layers between the bank's loan and the cash that's eventually supposed to pay it back.

How the system hides stress

This is the part I find most under-appreciated. In a normal credit cycle, rising rates produce defaults and defaults produce recognized losses. That feedback loop has been turned off.

When a borrower can't pay cash, the loan converts to accrued principal — the lender stops getting paid, but the loan grows on paper instead of going bad. Maturity dates get pushed out by mutual agreement. Funds raise cash against their own portfolios instead of selling positions. General partners roll old portfolios into new vehicles rather than mark them down.

These aren't workarounds. They're features. They exist so a cycle can be stretched in time rather than show up as a price.

The numbers that matter sit in Lincoln International's data. Loans where the borrower has switched to non-cash payment — meaning they couldn't pay — are around 6.4% of BDC portfolios. The officially reported "loan in trouble" figure stays at 1 to 2%. The gap is the part of the cycle that has been pushed to a future quarter.

Stack AI infrastructure on top of this and the structure depends on the machinery rather than just borrowing it. A three-year GPU has to be refinanced before its loan amortizes. The whole thing assumes the next round of capital shows up at terms not much worse than this one. That's a reasonable assumption when liquidity is plentiful and the regulatory backdrop stays friendly to this kind of bank exposure. Less reasonable when either of those stops being true.

The right question

The right question for the back half of this cycle isn't "will defaults rise?" It's "who buys the paper next?"

Two refinancing walls are coming at the same time. The 2021-22 vintage of direct lending and LBO debt is hitting its rollover window. The 2025-26 vintage of AI infrastructure debt is locked into a three-year refinance clock for as long as the technology stays interesting. Both rely on a buyer for the next layer.

In a normal year that buyer is the same machine that wrote the last layer: banks routing through non-bank lenders, private credit funds rolling into new vehicles, retail money flowing in through 401(k) wrappers and BDC-style funds. None of those channels are closed. For now, liquidity is still around and investor appetite for this kind of paper hasn't visibly stepped back — though both can shift faster than the refinance clock.

The real risk isn't a channel closing. It's that the channels all stay open, both refinancing walls land on schedule, and the market discovers the same dollar has been counted in three different funds.

That's not the kind of accident a single bad quarter cleans up. It needs a different buyer to absorb it, and the list of plausible buyers hasn't grown in five years.

Source Notes

• Hana Securities, 2H26 Overseas Credit Outlook (이영주, 2026-05-14) — primary framing for the shift from middle-market lending to AI infrastructure finance.
• Howard Marks, "Is It a Bubble?" (Oaktree) — corroborating view on bubble characteristics.
• Samsung Securities, BDC redemption note (2026-03-09) — BDC stress signals.
• KB Securities, US Market Pulse (2026-03-09) — context on the broader credit cycle.
• Kansas City Fed (2025) — bank lending to non-bank financial firms and the 29% versus 8% ROE comparison.

The Bottom Line

The loudest risk in private credit — middle-market default cycling — is the one most people are watching, and it's probably the smaller one. The quieter risk is that the asset class has restructured itself around an infrastructure that refinances on a three-year clock, with no obvious next pocket of capital to absorb it. It doesn't show up on a default chart because, by design, it doesn't show up anywhere until the refinance window arrives.

Worth at least asking who's on the other side of the next round.