Two Records, One Balance Sheet: Private Credit Is Now Lending Against Both Sides of the AI Trade

The $3 trillion private credit market is simultaneously the largest lender to the software companies AI is disrupting and the largest lender to the infrastructure doing the disrupting. Those two exposures do not offset. Today the equity market priced a record high; the loan market priced 125 basis points wider. One of them is wrong.

Deep dive · Wednesday, August 5, 2026 · prices as of 16:00 UTC (midday US session) · @dailyanalysts

1. The finding, stated plainly

On a day when the Dow printed a fresh all-time intraday high and the S&P sat within a whisker of its first-ever close above 7,700, the entire private-credit complex was down 2% to 4.3% — and it happened one day after Apollo reported record fee-related and spread-related earnings. That is not a financials move. XLF was flat (-0.05%). This was a targeted repricing of one business model.

The reason is visible in the primary loan market rather than the equity market. In the last week of July, CoreWeave had to widen the spread on a $2.6 billion GPU facility by 100–125bp, accept a 97-cent issue price, and — for the first time in a decade and a half of covenant-lite convention — grant a 1.35x maintenance debt-service-coverage covenant before investors would fund it (deal terms and syndication detail). At least four AI borrowers sweetened terms in the same week. Proofpoint reworked a $5 billion refinancing to roughly a 9.3% yield after a lender revolt. Meanwhile Blackstone is shopping a second $36 billion package for Anthropic's Google TPU leases, two months after the first $35 billion deal closed.

The core insight: the cost of financing the AI buildout is rising at exactly the moment the asset yields inside private credit portfolios are falling. Average portfolio yields across the four largest public BDCs dropped roughly 60–110bp year over year. Rising liability costs for borrowers and falling asset yields for lenders is the definition of a compressing spread business — and the market is now discriminating violently between the lenders that have already reset their dividends to reality and the ones that have not.

2. What actually moved today

TickerPriceTodayLatest NAV/sharePrice/NAVFwd yieldNon-accruals (portfolio)
OBDC (Blue Owl Capital Corp)$10.98-1.96%$14.41 (Q1'26)0.76x (-23.8%)11.3%1.0% (from 1.1%)
ARCC (Ares Capital)$19.27-1.93%$19.35 (Q2'26)1.00x (-0.4%)10.0%2.4% (from 2.0%)
FSK (FS KKR)$10.98-4.27%$18.83 (Q1'26)0.58x (-41.7%)15.3% base4.2% (from 2.1%)
MAIN (Main Street)$56.00-2.01%$33.46 (Q1'26)1.67x (+67%)5.7%1.2% (from 1.7%)
GBDC (Golub)$12.82-1.99%~10%dividend cut $0.39→$0.33
BXSL (Blackstone Secured Lending)$23.31-1.98%
APO (Apollo)$127.84-4.02%52wk range $99.56–$153.2917.2x fwd P/E1.7%record FRE+SRE Q2
ARES$138.23-3.27%
KKR$104.32-3.59%
BX (Blackstone)$134.68-1.84%
XLF (financials, for contrast)$57.85-0.05%SPY $770.65 (-0.09%) · QQQ $720.48 (-0.47%) · IWM $300.26 (-0.48%)

NAV and non-accrual figures above are as reported by each BDC and summarised in this sector review of Q1/Q2 2026 BDC results. Portfolio yields: OBDC 11.1% (Q4'24) → 10.0% (Q1'26); ARCC 10.9% → 10.3% y/y; FSK 11.0% → 9.9%; MAIN 11.4% → 10.3%.

3. Why this is happening now: three mechanisms, each with a number

(a) The asset-yield mechanism — 100bp gone, and it lands entirely on the dividend

A BDC is a spread machine: borrow at a spread over its own cost of funds, lend at SOFR-plus. When the average loan coupon falls from 11.1% to 10.0%, on a portfolio levered roughly 1.1x, roughly 230bp of gross return on equity disappears before a single loan goes bad. That is why Blue Owl cut its base dividend from $0.37 to $0.31 — precisely to Q1 adjusted NII of $0.31 — and why FS KKR's adjusted NII fell 37% year over year to $0.41 and its base dividend has already been cut twice, to $0.42. Fitch flagged this exact sequence in its "deteriorating" sector outlook: falling rates plus elevated PIK compress core earnings, dividend cuts follow, and unsecured maturities peak in 2026.

(b) The shadow-default mechanism — the headline 2% is not the number that matters

Private credit's headline payment default rate is around 2%. The number that matters is "bad PIK" — interest deferred mid-loan because the borrower cannot pay cash, as opposed to PIK elected at origination. That figure reached 6.4% of loans in Lincoln International's database as of Q4 2025, up from 2.5% in Q4 2021, with covenant defaults at 3.2%; Morgan Stanley's credit team puts direct-lending default rates at roughly 5.6% today with a path to 8% (see the CAIA compilation of BIS, Lincoln, Morgan Stanley and FSB data). Consequence: stated NAVs are marked against a 2% world while secondary buyers of gated fund interests transact against a 6% world. The gap closes at maturity events, not at quarter-ends — and 23 of 32 rated BDCs have $12.7 billion of unsecured debt maturing in 2026, a 73% increase over the prior year.

(c) The concentration mechanism — the same book, twice

Loans to SaaS companies grew from about $8 billion in 2015 to over $500 billion, 19% of all direct loans, by end-2025; BDCs put more than 15% of their books there, and at least 250 software loans worth over $9 billion were classified under other industries in BDC disclosures. At the same time the Financial Stability Board found that AI accounted for more than a third of private credit deals in 2025, up from 17% over the prior five years. So the same managers hold the disrupted (horizontal application software originated at ~60x earnings in 2021–22, refinancing into an ~18x market) and the disruptor (GPU and data-centre paper). A correction in AI infrastructure valuations does not hedge the software book — it adds to it.

4. The part almost nobody is pricing: who actually holds the AI risk

Read the structure of the Anthropic financings and the risk-transfer becomes obvious. A bankruptcy-remote SPV borrows, buys Google Ironwood TPUs, and leases compute to Anthropic, which pays rent rather than debt service. The proposed second package is roughly $6bn A1 notes / $25bn A2 notes / $4.5bn B notes, and the A tranches carry Broadcom's residual value support agreement — if Anthropic stops paying and the chips fetch less than the senior claim, Broadcom covers the shortfall (full structural breakdown).

My opinion, and the single most underpriced item in this entire complex: across the closed $35bn deal and the proposed $36bn deal, roughly $60 billion of senior notes are made pension-eligible by Broadcom's guarantee of used-TPU resale value — for an asset class with no established secondary market. That is a contingent liability that appears in no leverage ratio, on a stock that is breaking out on triple-digit AI revenue growth. Broadcom is monetising the same silicon twice: once by selling it, once by insuring its residual value. I am not short AVGO — the cash generation is real — but I want that notional disclosed, and the market has not asked.

Two further consequences most commentary skips:

And the macro tell arrived this morning: private payrolls rose just 44,000 in July versus 75,000 expected, with 36,000 of that in health care and outright declines in trade/transport (-8k) and mining (-6k) (ADP report). Weak labour demand plus "worse than pandemic era" input-cost commentary in the manufacturing survey is the stagflationary mix that hurts levered mid-market borrowers first — they are the ones with floating-rate debt and no pricing power. Bank of America's CEO, asked today about the Situational Awareness hedge fund near-collapse, said: "These are all warning shots… Valuations get out, leverage in the system gets there. You have to be careful" — and that the tendency is to "tighten the underwriting standards, just a hair" (CNBC). Prime brokers tightening "a hair" is how private credit funding costs rise without a single headline.

5. Valuation: what today's prices actually imply

This is where the market has stopped doing arithmetic. Take OBDC. NAV is $14.41 with the portfolio levered roughly 1.1x, so assets are about 2.15x equity. The $3.43 per-share discount therefore implies a portfolio-wide loss of roughly 11% of assets.

Scenario for OBDC's bookAsset lossImplied NAVvs $10.98 price
Observed non-accruals (1.0%) at 40% severity0.4%$14.28+30%
Shadow distress (6.4% bad-PIK) at 40% severity2.6%$13.61+24%
Morgan Stanley bear case (8% default) at 40% severity3.2%$13.42+22%
What the price implies~11%$10.980%

Now put that next to ARCC, which trades at par to NAV with non-accruals of 2.4% — 2.4x OBDC's — and NAV that fell from $19.90 to $19.35 year over year, with Truist cutting its target to $21 and KBW to $20 after the Q2 print. The market is paying a ~24-point relative premium for the BDC with the worse credit book. That is the mispricing, and it is the trade.

FSK is the opposite error in the opposite direction: 0.58x NAV looks like deep value until you notice non-accruals doubled to 4.2%, NAV fell 19% in twelve months ($23.37 → $18.83), adjusted NII fell 37%, and the base dividend has been cut twice. At 0.58x, the market is extrapolating two more years of exactly what just happened. It is probably close to right.

6. The trades

HIGH CONVICTION Pair: LONG OBDC / SHORT ARCC (relative-value, positive carry)
HIGH CONVICTION AVOID / UNDERWEIGHT FSK — the 15%+ base yield is a countdown, not a coupon
SPECULATIVE Fade rallies in the AI-SPV originators: APO (and ARES) into strength
WATCH Prefer AI infrastructure credit over AI infrastructure equity

7. Three scenarios, with trigger levels

Bull — 30%

The Strait of Hormuz deal closes, oil stays down, the 10-year drifts back under 4.30% and credit spreads tighten. Marks stabilise (remember OBDC's NAV decline was attributed to spread widening, not credit losses), Q2 non-accruals come in flat sector-wide, and discounts compress. Trigger: Oracle 5-year CDS back below 150bp and OBDC Q2 non-accruals ≤1.2%. Outcome: OBDC $13.00–13.75, APO back to $150, FSK bounces to $13 (the lowest-quality name rallies hardest — do not confuse that with vindication).

Base — 50%

The grind. Asset yields keep bleeding ~25bp a quarter, two or three more dividend cuts land (FSK again, GBDC again, PSEC), NAVs erode 1–3% per quarter, and discounts persist because the maturity wall keeps forcing recognition. Trigger: another BDC cut announced before September 30. Outcome: the long OBDC leg returns 11% in dividends plus modest discount narrowing; the alt managers range-trade $110–140; ARCC de-rates to a 5–8% discount as its dividend coverage thins. The pair works in this scenario, which is the point of running it as a pair.

Bear — 20%

An AI credit accident. Triggers, any one of: CoreWeave trips or renegotiates its brand-new 1.35x DSCR covenant; a second agency puts Oracle at low-BBB (Morgan Stanley's credit desk explicitly flags two low-BBB ratings as possible by year-end, and a fallen-angel path as a medium-term risk); or the $36bn Anthropic package fails to syndicate at the terms floated. Outcome: the public BDC complex re-tests and breaks the ~17% index discount toward 25–30%, OBDC $9.00 (its 52-week low is $10.52, set April 2), FSK $7, APO $95. In that world the correct expression is not "buy the dip" — it is short-duration Treasury bills and the VIX hedge we already have on.

8. Cross-asset consequences

9. Prior calls, marked

The tactical volatility hedge we put on August 4 (September VIX calls / 1–2% OTM SPY puts with cash tilted to T-bills, VIX ~16.5) is working in thesis if not yet in price — today's 44k ADP print against a 75k consensus raises the stakes on Friday's July jobs report, which is precisely the event it was bought for. Keep it. The chip-equipment underweight remains downgraded to WATCH after last week's short-covering squeeze; no change today.

10. Bottom line

Private credit is not one trade, it is two: a spread business whose asset yields are falling ~100bp a year, and a concentrated bet on AI infrastructure whose borrowing costs are rising 100–125bp a deal. The market is pricing the first risk indiscriminately and the second risk barely at all. The way to be paid for that is not to buy or sell "private credit" — it is to own the lender that has already told the truth about its earnings power (OBDC, 0.76x NAV, 1.0% non-accruals, dividend reset to NII) against the lender the market still treats as unimpeachable (ARCC, 1.00x NAV, 2.4% non-accruals, thinning coverage), avoid the value trap (FSK at 0.58x is cheap for reasons), fade the SPV originators into strength (APO $135–142), and — where you can access it — take the covenant-protected paper at 10.44% over the equity of the same borrower.

Three things to do tomorrow: (1) put OBDC's Q2 non-accrual line on your calendar — it is the invalidation for the whole thesis; (2) set an alert on Oracle 5-year CDS at 250bp; (3) mark August 11 for CoreWeave's interest expense against the $650–730m guide. Everything else in this market is commentary.