Autoresearch: if AI-capex derates, which private-credit book is most exposed?

Generated by /autoresearch on 2026-08-17. Synthesized across 3 rounds from 12 successful web fetches (1 failed), anchored by Grokipedia Private_credit. See Provenance. Treat as raw material — review before promoting into a project or thread. Context: vault/projects/stock-market

Priors capture skipped — unattended weekday shift; user instructed skip.

X pass enabled. 0 high-signal posts surfaced (0 via direct fetch, 0 via search snippets). Three site:x.com discovery queries returned no permalinks. X MCP needsAuth in this environment. See Provenance.

Summary

Layer 1 of this stack (the 2026-08-17 CSP CapEx spike-vs-floor source on the parent branch) narrowed 2026–27 hyperscaler spend toward not a spike. This pass asks the residual: if AI-capex still derates, which public private-credit / BDC / alt-manager book takes the first asset-quality losses? Grokipedia Private credit is the vocabulary primer only.

The honest answer from Q2 2026 primaries is: do not name a most-exposed listed BDC for Feeder A (datacenter / neocloud / GPU-SPV loan losses). The large AI-infrastructure financings that would mark first in a capex derate sit in different vehicles than the candidate tickers (ARCC, BXSL, OBDC, OWL, OTF, GBDC, FSK, BX). Blue Owl's Meta Hyperion equity is in Digital Infrastructure / Real Assets funds, with PIMCO-issued ~$27B investment-grade project bonds as the debt (Bisnow). Blackstone's visible datacenter book is BREIT / QTS equity (BREIT Q1 2026 update); the one named GPU-loan anchor is Blackstone Credit & Insurance, not BXSL (CoreWeave 8-K exhibit 99.1). Ares' datacenter platform is Ada Infrastructure (equity development), not ARCC (Ares Q2 2026 call).

Public BDC Q2 prints update software / middle-market asset quality (the June Feeder-B ranking). They do not disclose a ranked AI-infra loan sleeve. No fetched source reports a realized neocloud / datacenter-SPV payment default. The August tell is pricing: CoreWeave's $2.6B DDTL 5.5 closed at SOFR+550 with maintenance covenants after a 100–125 bp blowout (CoreWeave IR; TechTimes).

Findings

Theme 1 — The conversion trigger is still unmet: no fetched ranking of AI-infra loan exposure by public BDC

The question's gap is "exposure concentration by manager" on the asset-quality / loan-loss path — datacenter, neocloud, GPU-SPV paper — not the FOMC redemption-outflow thesis and not the software-disruption (Feeder B) ranking already filed in June.

Grokipedia's Private credit page is a useful primer (direct lending vs banks; BDCs as the retail window; named platforms Blackstone, KKR, Apollo, Ares, Blue Owl). It does not quantify AI-infrastructure loan books by vehicle. Treat it as vocabulary, not a 2026 exposure table.

The Drift's 2026-07-20 BDC weekly is the load-bearing negative finding for the listed names: Blue Owl is the majority equity partner on Meta's Louisiana project, but that is a platform-level fact, not proof that OBDC owns the project debt or equity. The same piece says the verified source set does not establish direct ARCC, BXSL, or BBDC ownership of the Meta financings (The Drift, 2026-07-20). The OBDC Q2 deep-dive repeats the boundary: "The broader Blue Owl platform can participate in data-center, power, equipment, real-estate, or asset-backed transactions without those exposures ever appearing inside OBDC" (The Drift, 2026-08-06).

Do not invent a "most exposed" public-BDC name from platform brand. OWL / BX / ARES parent fee exposure to digital-infra AUM is a different chain (FRE, not loan marks).

Theme 2 — Where the AI-infra book actually sits (named vehicles, not ticker inference)

Blue Owl Real Assets / Digital Infrastructure — not OBDC. Meta and Blue Owl formed a JV to develop the Hyperion campus in Richland Parish, Louisiana: funds managed by Blue Owl own 80%, Meta 20%. Blue Owl contributed around $7B; Meta contributed land and assets under construction. Meta will lease the completed facilities on an initial four-year term with extension options, plus a residual-value guarantee for the first 16 years. A portion of the capital is PIMCO-issued ~$27B investment-grade bonds backed by the project's assets; the securitization priced above par (Bisnow). On the OWL Q2 2026 call, management said the data-center footprint spans more than 140 data centers owned or under construction globally, with 15.3 gigawatts of leased and owned capacity (OWL Q2 2026 transcript). Stack Infrastructure Issuer, LLC — an indirect wholly-owned subsidiary of Blue Owl Digital Infrastructure Trust — issued $695M Series 2026-1 Class A-2 notes on 2026-03-03 (SEC 8-K, accession 0002069692-26-000012 — search-extracted; not re-fetched as HTML body). That is a Real Assets / digital-infra trust vehicle, not the listed BDC.

OBDC's own Q2 2026 release (quarter ended 2026-06-30) reports a $15.0B portfolio, 78.8% senior secured, 96.0% floating-rate, non-accruals 2.8% at cost / 0.8% at fair value, NAV $14.26, and no datacenter or AI-infra loan sleeve (OBDC 2026-08-05 release). YE2025 10-K language (search-extracted, not this pass's fetch) already put the largest industry at internet software and services, 11.1% of fair value and located digital infrastructure on Blue Owl's Real Assets platform, not the Credit / OBDC adviser.

Blackstone — BREIT/QTS equity and Credit & Insurance, not BXSL. BREIT's Q1 2026 update says data centers grew from 1% of the portfolio in 2020 to 23% today, with $5.8B into pre-leased data-center developments in 2025 and $2.4B in Q1 2026, via the QTS platform (BREIT Q1 2026 update). That is real-estate equity, not a BDC loan book. BXSL's Q2 2026 call: software 19% of fair value across 70 borrowers; largest disclosed industries are software / healthcare / professional services / insurance; no established ownership of the Meta project financings (BXSL Q2 2026 transcript; The Drift, 2026-07-20).

CoreWeave GPU-SPV — bank syndicate + Blackstone Credit & Insurance. The $8.5B DDTL 4.0 (closed 2026-03-30/31) is borrowed by CoreWeave Compute Acquisition Co. VIII, LLC, rated A3 / A (low) — "the first investment-grade rated financing secured by HPC infrastructure and an associated customer contract." Arrangers: MUFG, Morgan Stanley, Goldman Sachs, JPMorgan. Anchored by Blackstone Credit & Insurance, plus "global financial institutions, asset managers, and insurance investors" (CoreWeave exhibit 99.1). CCIR's document read shows the rating travels with the offtake: DDTL 4.0 (Meta take-or-pay) is A3 at SOFR+2.25%; DDTL 5.0 (two non-IG customers) is Ba2 at SOFR+4.50% with a full parent guarantee (CCIR). No fetched filing names ARCC, BXSL, OBDC, FSK, GBDC, or OTF as a CoreWeave lender.

Ares — Ada Infrastructure, not ARCC. On the Ares Management Q2 2026 call, Ada Infrastructure is a vertically integrated data-center development platform: "over 100 professionals… executing on 7 large data center campuses representing 22 individual data center investments with approximately 1 gigawatt of compute," plus a global digital-infrastructure fund still raising (Ares Q2 2026 transcript). That is an equity / FRE story. ARCC's Q2 2026 release and call discuss software LTV and AI-disruption risk inside the software book (<50 bps of ARCC fair value in "higher AI risk software"); they do not disclose a datacenter loan sleeve (ARCC 2026-07-29 release; ARCC Q2 2026 transcript).

Theme 3 — If Feeder A still fires, the first marks are IG project / GPU-SPV paper — and that paper has not defaulted

Layer 1's "not a spike" finding lowers the prior on a 2026–27 capex-derate. The residual risk is a later cut (FCF compression, power/grid gates) that strands financed assets.

What the August 2026 market actually printed is tighter underwriting, not a missed payment:

• CoreWeave closed a $2.6B DDTL 5.5 on 2026-08-10 at SOFR+5.50%, Ba2 / BB+, five-year maturity against ~three-year average customer contracts — lenders underwriting renewal risk. Year-to-date debt-and-equity capital >$30B (CoreWeave IR, 2026-08-10).
• TechTimes, citing Bloomberg syndication, says initial talk was SOFR+425–450 at 99; the deal blew out 100–125 bp to SOFR+550 at 97, YTM 10.44%, plus a 1.35x DSCR maintenance covenant and $112.5M minimum liquidity. At least four AI borrowers sweetened terms the same week. Moody's (as retold) flags Oracle and CoreWeave as the most immediate ratings pressure among six tracked tech names; Fitch's Q3 Global Risk Outlook flags an AI correction as one of two dominant short-term global-credit risks (TechTimes, 2026-08-03).
• Hyperion's ~$27B PIMCO bonds are described as investment-grade and priced above $1.10 (Bisnow). Meta's residual-value guarantee is the contractual backstop if the four-year lease is not renewed.

The June "hyperscaler backstop" objection on the question page is strengthened, not retired: the largest named AI-infra debt prints are IG or IG-adjacent because of offtake, not because a BDC underwrote speculative neocloud cash-flow. A derate that breaks take-or-pay / residual-value contracts would hit those SPVs and their insurance / PIMCO / Blackstone Credit & Insurance / bank-syndicate holders first. That set is not mapped onto the listed BDC tickers in any fetched source.

The FSB HTML page timed out this pass (https://www.fsb.org/2026/05/fsb-warns-on-private-credit-vulnerabilities/). Search snippets still describe the May 6 2026 FSB report as warning that sector concentration (technology, healthcare, services) and data-centre financing could produce "sizeable" losses; do not treat the snippet as a fetched primary. The May 2026 FSB PDF was already noted (not fetched) on the June 16 source.

Theme 4 — Q2 2026 public-BDC prints update Feeder B / general credit, not Feeder A concentration

These numbers are asset-quality updates. They do not convert the Feeder-A gap. Do not read them as a "most exposed to AI-capex derate" ranking.

Read-through for the June pair, without converting Feeder A:

• FSK still has the worst non-accrual level even after a sequential repair. That is general credit repair, not a disclosed AI-infra sleeve.
• OBDC covered the reset dividend; cost non-accruals rose; the portfolio is still shrinking. Still no AI-infra loan % .
• ARCC remains the cleanest large listed BDC on the prints above; management's AI comment is about software disruption, and they put high-AI-risk software at <50 bps.
• BXSL non-accruals improved because two names left after restructuring (Medallia the bulk). Software 19% is Feeder B.
• OTF remains the software-concentration name (~70%). Digital-infra comments are forward-looking origination, not a current exposure table.
• OWL (parent) is the platform with the named Hyperion / Stack / 15.3 GW book. That is Real Assets AUM / fee income, plus whatever marks sit inside Digital Infrastructure funds — not OBDC NAV.

Theme 5 — Two contagion paths must stay unbundled

Feeder A (this question, still open on the listed tickers). AI-capex derate → marks on datacenter / GPU-SPV / neocloud credit. Layer 1 makes the trigger less likely on the 2026–27 clock. If it happens, the named holders in this pass are PIMCO/IG project bonds, Blue Owl Digital Infrastructure / Stack / Hyperion equity funds, Blackstone Credit & Insurance (CoreWeave DDTL 4.0), BREIT/QTS (equity), bank syndicates, insurers. Public BDCs are not shown as the first-loss book. OWL/BX/ARES parents have fee exposure if those funds stop growing or take impairments — that is not the same as ARCC/BXSL/OBDC/FSK loan losses.

Feeder B (already the June dominant leg; this pass only refreshes prints). AI disruption of software/services borrowers inside BDC books. Q2 2026 does not overturn the June relative ranking (OTF still ~70% software; FSK still the worst NA print; ARCC still the diversified core). It also does not let anyone promote Feeder B into a Feeder-A "most exposed to capex derate" claim.

Not this question: FOMC / 5% redemption-cap outflows (fomc-private-credit-outflows-alt-managers).

Contradictions and open questions

• Most-exposed public BDC for Feeder A: still unnamed. No fetched 10-Q, earnings release, or analyst note in this pass publishes a ranked AI-infrastructure / datacenter / neocloud loan % for ARCC, BXSL, OBDC, OTF, GBDC, or FSK. Naming one would be an invention. The conversion trigger on the question page remains unmet for Feeder A.
• Platform vs vehicle. OWL/BX/ARES look most exposed at the brand level. The vehicles that hold the paper are Digital Infra funds, BREIT/QTS, Blackstone Credit & Insurance, Ada. A later 10-Q line that puts Hyperion, Stack, or CoreWeave paper inside a listed BDC would change this.
• OTF digital-infra "opportunity." Management wants to add IG-counterparty digital-infra debt. That would create Feeder-A exposure that is not in the current book %. Watch the next sector table, not the rhetoric.
• Blackstone Credit & Insurance ≠ BXSL ≠ BCRED ≠ BREIT. CoreWeave's press release names the Credit & Insurance platform as DDTL 4.0 anchor. It does not disclose ticket size or which fund. Do not collapse that into BXSL or the BX common.
• No realized SPV default. Pricing stress (CoreWeave blowout, covenants returning) is not a credit event. The first missed GPU-SPV or datacenter-project payment is still the watch item.
• FSB primary not re-fetched. HTML timed out; PDF is off the *.gov whitelist. The May 2026 FSB "sizeable losses" language stays at snippet / prior-source status.
• GBDC calendar lag. The latest fetched GBDC print is fiscal Q2 ended 2026-03-31. A later GBDC quarter was not fetched.
• Layer 1 vs this question's step 1. The question page still calls the AI-capex cycle "peaked and at risk of derating." Layer 1's Q2/FY26 guides contradict a 2026–27 guided derate. This pass does not rewrite that mechanism; it treats Feeder A as conditional.

Provenance

Rounds run: 3 of 3

Sub-questions by round:

Round 1 (broad survey):

1. What Q2 2026 disclosures quantify AI-infrastructure / datacenter / neocloud loan exposure by named public BDC or alt-manager (ARCC, BXSL, OBDC, OWL, OTF, GBDC, FSK, BX)?
2. Have any named datacenter-SPV / neocloud / GPU-as-a-service credit events (missed payment, restructuring) been reported since June 2026?
3. How do Q2 2026 BDC earnings update asset-quality metrics without collapsing Feeder A into Feeder B?
4. Are AI-infra loans mostly IG hyperscaler / take-or-pay, or speculative neocloud?
5. What do OFR / Fed / FSB 2026 updates say about private-credit concentration in AI-capex financing?

Round 2 (drill-down):

1. Which named vehicles (not platforms) hold Blue Owl Stack / Meta Hyperion / CoreWeave GPU-SPV debt — and is any of that in public BDCs? — targeting the platform-vs-vehicle gap
2. Does Blackstone disclose a comparable platform-vs-BDC split for AI-infra? — targeting BX vs BXSL vs BREIT
3. Retry FSB / OFR primary on data-centre private-credit losses — targeting the failed FSB fetch

Round 3 (resolve remaining uncertainty):

1. Confirm Blackstone Credit & Insurance vs BXSL for CoreWeave — targeting the one named GPU-loan anchor
2. Confirm Ares Ada vs ARCC — targeting the last large platform
3. Does any public BDC sector table show a datacenter sleeve? — targeting the "most exposed" naming bar

Anchor source (Grokipedia, fetched before round 1):

• Private credit — 25,017 chars extracted (capped) — primer on market size (~$2T early 2026), BDC window, named platforms; no AI-infra loan table by manager.

X sources (--include-x; x-fetch skill + _lib/x.py):

• 0 items surfaced (0 direct fetch, 0 search snippets)
• Three site:x.com queries (AI datacenter private credit BDC; private credit AI datacenter loans; named tickers + datacenter) returned no permalinks
• X MCP server needsAuth in this environment — same failure mode as the 2026-08-17 CSP CapEx pass

URLs fetched (12 successful, 1 failed):

Round 1:

• OBDC Q2 2026 deep-dive — specialist analysis — platform ≠ OBDC; Q2 income vs shrinking book
• CoreWeave $2.6B DDTL 5.5 close — official — SOFR+550, Ba2/BB+, 5y loan vs ~3y contracts, >$30B YTD capital
• TechTimes on the 125 bp blowout — news — covenants returning; Moody's/Fitch warnings as retold
• CCIR GPU SPV — specialist / filing read — rating travels with offtake; DDTL 4.0 A3 vs 5.0 Ba2
• [Failed: https://www.fsb.org/2026/05/fsb-warns-on-private-credit-vulnerabilities/] — timeout

Round 2:

• OBDC 2026-08-05 results — official — NA 2.8%/0.8%; no AI-infra sleeve
• The Drift, 2026-07-20 — specialist — Meta Louisiana is Blue Owl platform, not OBDC/ARCC/BXSL

Round 3:

• Bisnow on Meta–Blue Owl Hyperion JV — trade press — 80/20 JV, ~$7B Blue Owl equity, PIMCO ~$27B IG bonds, 4-year lease + 16-year residual-value guarantee
• CoreWeave DDTL 4.0 exhibit 99.1 — official / SEC — A3/A(low); anchored by Blackstone Credit & Insurance
• BREIT Q1 2026 update — official — data centers 1%→23%; $5.8B 2025 + $2.4B Q1 2026 via QTS

Also used as full-text search extracts (not a second WebFetch; cited in Findings):

• ARCC 2026-07-29 release — official
• ARCC Q2 2026 transcript — transcript reprint
• FSK 2026-08-06 release — official
• BXSL Q2 2026 transcript — transcript reprint
• OTF 2026-08-05 release — official
• OTF Q2 2026 transcript — transcript reprint
• OWL Q2 2026 transcript — transcript reprint
• Ares Q2 2026 transcript — transcript reprint
• GBDC fiscal-Q2 2026 transcript — transcript reprint (older quarter)

Tools used: WebSearch, WebFetch, grokipedia-fetch (skill / _lib/grokipedia.py), x-fetch discovery via site:x.com WebSearch (X MCP needsAuth). Generated: 2026-08-17 13:45 UTC