The AI story stopped being purely an equity story sometime last autumn. Hyperscaler capex (~$400B in 2025, an estimated ~$700B in 2026) has blown past the free cash flow that funds it — Amazon is projected to run negative free cash flow up to $28B in 2026; Alphabet and Meta free cash flow is expected to fall ~90%. The gap is being bridged with debt, and increasingly with private credit: outstanding private-credit loans to AI-related borrowers have gone from near zero to over $200 billion in a few years, and Morgan Stanley projects private credit will supply another ~$800 billion of data-center financing within two years (Prosek/industry estimates put the through-2030 figure near $750B–$1T+). My core, non-consensus call: "AI credit" is not one exposure. It is four claims — hyperscaler duration, leased-datacenter residual, securitized tenant cash flow, and merchant compute demand — and the credit market is separating them in real time while the equity market still trades the theme as a monolith. Own the toll-booth (Blackstone). Avoid the concentrated fault line (Blue Owl, CoreWeave). Keep owning honest, market-priced public BDCs (ARCC, BXSL) over gated non-traded funds.
Three transaction templates now dominate, and each hides risk in a different place. (Sources: Quinn Emanuel litigation memo, Mar 2026; CNBC, Apr 2026.)
In October 2025, Meta closed the largest private-credit data-center deal in history for its Hyperion campus in Louisiana. An SPV — Beignet Investor LLC — was created, owned 80% by Blue Owl and 20% by Meta. It raised ~$30B: roughly $27B in loans (PIMCO, BlackRock, Apollo and others) plus $3B of Blue Owl equity. The SPV owns the data center and leases it back to Meta. Net effect: Meta borrowed ~$30B without a dollar of that debt appearing on its balance sheet. It records only its equity stake and lease obligations. Meta also gave a "residual value guarantee" of up to $28 billion — the commitment appears only in the footnotes of its annual report, with no liability booked. The Hyperion structure is ~91.5% debt ($27B debt / $2.5B equity). Days after closing, Meta raised another $30B in the public bond market.
Oracle sold $18B of bonds in a single day in September 2025 to fund OpenAI commitments, then layered on off-balance-sheet SPVs: a ~$13B Blue Owl/JPMorgan structure for the Abilene facility, a $38B syndicated facility for Texas/Wisconsin, and an $18B loan for New Mexico — four-year "mini-perm" loans priced ~250bps over Treasuries that must be refinanced through ABS/CMBS once (promised) tenant cash flows show up. Oracle's total debt now exceeds $130B with $248B in new lease commitments; its 5-year CDS has risen ~310% to a 16-year high. Bondholders already sued (Ohio Carpenters' Pension Plan v. Oracle, filed Jan 14, 2026).
The riskiest collateral of all is the chips themselves. On March 31, 2026 CoreWeave closed an $8.5B delayed-draw term loan (DDTL 4.0) — the first investment-grade-rated (A3 / A-low) GPU-backed financing, anchored by Blackstone Credit & Insurance, priced at SOFR+2.25% (floating) / ~5.9% (fixed), maturing March 2032. CoreWeave has raised ~$28B of debt and equity in 12 months. Its earlier $7.5B facility carried an ~11% rate and began amortizing in January 2026 — just as GPU collateral values were falling. That is the crux: hyperscalers depreciate GPUs over ~6 years, but datacenter GPUs run hot and last 1–3 years in practice. Loans outlive the collateral. Commentators call it the "GPU debt treadmill." CoreWeave also faces a securities class action (Masaitis v. CoreWeave, Jan 12, 2026).
This is the part worth reading twice. Per JunkBondInvestor's data pack (Jun 28), the spread dispersion inside "AI debt" runs from ~120bps to ~600bps — not because one thing is mispriced, but because the label covers four fundamentally different creditor claims:
| Claim | What you're really underwriting | Approx. spread | My risk read |
|---|---|---|---|
| Hyperscaler duration (MSFT/AMZN/GOOGL/META/ORCL bonds) | Long-dated IG corporate credit + capex/leverage trajectory | AA cluster tight; ORCL long end 214–245bps (BBB) | Lowest default risk; highest duration risk. Oracle is the widest and still repricing. |
| Leased-datacenter residual | Tenant quality, lease durability, power access, real-estate residual | Mid | Contained if tenant is a hyperscaler; the physical asset has value. |
| Securitized tenant cash flow (DC ABS/CMBS) | Contracted lease cash flows w/ tranche subordination | Senior AAA ~120–145bps | Best risk-adjusted claim in the stack; extension/refi risk on long tenors. |
| Merchant compute demand (CoreWeave, neoclouds) | GPU utilization, customer concentration, hardware residual, refi access | ~500–600bps, ~10% yield | The fault line. Chips depreciate faster than the loans; hardest to control in a default. |
"This is not one credit market yet. It is a set of claims being priced apart in real time." — JunkBondInvestor, Jun 28, 2026
My take: the equity market is doing the opposite of what the credit market is doing. Equity treats AI as a single momentum theme; credit is already carving it into layers. That divergence is an opportunity, because it means the beaten-down "private credit" equities are not all the same trade — and neither are the AI clouds.
~$45B of high-yield AI-infra paper comes due across 2030–2031 — most of it issued in 2025 when spreads were tight — and it is concentrated in exactly the merchant-compute/neocloud names where the creditor claim is most exposed to the demand cycle. The structural question becomes a timing question at that wall. Nobody modeling the equity upside is modeling that refi.
Watch what Blackstone does, not what everyone says. In the past week BX sold its Northern Virginia data centers to Digital Realty for $3.5B and its QTS unit abandoned a 2,100-acre Virginia campus — i.e., it is quietly taking chips off the data-center equity table — while simultaneously anchoring CoreWeave's senior-secured debt. Translation: the most sophisticated allocator in the world would rather hold the senior claim than the equity in this phase. That is exactly how you should be positioned.
| Idea | Direction / conviction | Entry | Target | Invalidation | Horizon |
|---|---|---|---|---|---|
| BX (Blackstone) — the toll booth | BUY on weakness · MODERATE/HIGH | $110–122 (last ~$122.81) | T1 $145, T2 $165 | Weekly close <$100 | 6–12 mo |
| OWL (Blue Owl) — the fault line | AVOID / SPEC short vs BX · SPEC | Fade bounces $9.0–10.5 (last ~$9.05) | $7.0 (pair: long BX / short OWL) | 2 straight quarters of falling redemptions, or weekly close >$12 | 1–3 mo |
| CRWV (CoreWeave) — merchant compute | AVOID / SPEC short on bounces · SPEC | Short $86–95 bounces (last ~$81.77) | $60, stretch $50 | Daily close >$100 OR major new hyperscaler take-or-pay contract | 1–3 mo |
| ARCC (Ares Capital BDC) | BUY · HIGH (reaffirmed from 7/2) | $17.75–18.75 (last ~$18.75) | T1 $20.25, T2 $22 | Weekly close <$16.75 OR non-accruals >3.5% OR div cut | 1–3 mo |
| BXSL (Blackstone Secured Lending) | BUY · HIGH (reaffirmed) | $23.25–24.40 (last ~$23.78) | T1 $26.25, T2 $27.50 | Daily close <$22 | 1–3 mo |
| Scenario | Prob. | Trigger | What works |
|---|---|---|---|
| Base — orderly repricing | 55% | AI revenue keeps scaling; rates drift lower; hyperscaler bonds fine; neocloud spreads stay wide but roll. Disputes rise, no systemic break. | BX, ARCC, BXSL, senior DC ABS. Merchant-compute equity (CRWV) underperforms. |
| Bull — the boom keeps funding itself | 25% | AI monetization accelerates, refi window stays open, GPU life proves longer, redemptions at OWL peak and reverse. | Everything, incl. OWL/CRWV squeeze — hence the small size and hard invalidations on the shorts. |
| Bear — the treadmill stops | 20% | A neocloud default triggers cross-defaults; GPU collateral marks collapse; 2030–31 refi wall pulls forward; OWL/non-traded redemption run forces asset sales; contagion into bank-NBFI channel. | Cash, hyperscaler IG (short duration), long BX-vs-OWL, short CRWV. The residual-value-guarantee footnotes become front-page liabilities. |
From the July 2 private-credit deep dive: ARCC (bought $17.75–18.75) sits ~$18.75 — working; BXSL ~$23.78 — working; OBDC (downgraded High→Spec, watch PIK >10% at Aug 5 print) — holding the line; FSK avoid — validated. The public-BDC-over-gated-fund framework is intact and today's data (Blue Owl's $4.7B redemptions, BlackRock TCP's 19% writedown) reinforces it.
The single most useful thing I can tell you today: stop trading AI credit as one thing. The equity market's habit of hammering Blackstone, Apollo and Ares on the same headline that hits Blue Owl is a gift — those are not the same exposure. Own the diversified senior-secured toll booth (BX) and the honest public BDCs (ARCC, BXSL); avoid the concentrated redemption-mismatch name (OWL) and the tip of the riskiest claim (CRWV). Watch the 2030–31 refi wall, GPU-collateral marks, and the next round of BDC filings (OBDC/FSK, early August) for the PIK >10% tell. And remember the smart-money tell: the biggest allocator alive is quietly selling data-center equity while keeping the senior debt. Position the same way.
This is analysis and opinion for informational purposes only — not investment advice. The author's directional views are clearly marked as opinion; underlying facts are sourced from the linked primary and secondary materials. U.S. equity markets were closed July 3, 2026 (Independence Day observed); prices reflect the most recent available quotes. Key sources: Quinn Emanuel, CNBC, JunkBondInvestor, CoreWeave IR, Tomasz Tunguz, Business Insider, FSB Report on Vulnerabilities in Private Credit (May 2026).