Autoresearch: private-credit/BDC stress — AI-disruption-of-borrowers vs datacenter-capex-derate

Generated by /autoresearch on 2026-06-19. Synthesized across 2 rounds (early-exit; round-2 evidence resolved the question) from 5 web pages, no Grokipedia anchor (fast-moving topic). Treat as raw material — review before promoting. Context: vault/projects/stock-market

Summary

The question was whether the private-credit contagion from AI runs primarily through datacenter-capex-derate financing losses (Feeder A) or AI disruption of the software/services borrowers whose loans sit in BDC books (Feeder B). Mid-June evidence tilts hard toward Feeder B. On the capex side, hyperscaler commitments are rising (~$650B+ in 2026, +36% YoY) and the institutional datacenter buildout is largely intact — the viral "30–50% of 2026 capacity cancelled" figure reconciles with SemiAnalysis's "~1% revision" because the binding constraint is power, not capital: projects are power-gated and pushed to 2028, and the long tail of never-real announcements is being culled — not a financing-loss event. On the borrower side, software is the single largest BDC sector (~20–26%, ~35% incl. adjacent) and agentic-AI seat displacement is eroding the per-seat subscription collateral faster than traditional credit models assume, with a visible "shadow drawdown gap" (software stocks −30%, marks ~flat; ~$25B of software loans trade <80¢ vs par marks). Net: the contagion is real but it's a borrower-quality / software-disruption story, the same forcing function as agentic-ai-seat-erosion-to-saas-rerate routed through credit, and the cleanest expression is the dispersion pair: long the diversified, well-covered lenders (ARCC, BXSL) / short the concentrated, redemption-pressured ones (FSK, OBDC/OWL, GBDC, OTF).

Findings

(a) Datacenter capex is holding — power-gated, not financing-cancelled (weakens Feeder A)

Hyperscaler capex commitments are accelerating, not retreating: Alphabet/Amazon/Meta/Microsoft are expected to spend >$650B on AI infrastructure in 2026, and CreditSights projects the top-five hyperscalers alone at ~$602B, +36% over 2025 (BloombergNEF, Futurum). The viral claim that "30–50% of 2026 US datacenter capacity will be delayed/cancelled" (tech-insider) reconciles with SemiAnalysis's rebuttal — NA colocation moved <5% and hyperscaler self-build ~1% over six months (SemiAnalysis) — once you see that the binding constraint is power infrastructure (transformers, switchgear, batteries, >4-year grid-connection waits), not capital or chips. Delays push capacity to 2028 and cull never-real announcements; they are not a wave of stranded-asset financing losses. Implication for the contagion thesis: the capex-derate financing-loss leg (Feeder A) is weaker than the headlines imply — datacenter SPV/neocloud debt is mostly serving real, power-constrained demand with the next buyer in line, not vanishing.

(b) AI disruption of software borrowers is the dominant private-credit risk (confirms Feeder B)

Software is the single largest BDC sector — ~19.7% average / 19.4% median for non-traded BDCs at YE2025, rising to ~35.6% including software-adjacent (healthcare-tech, IT/professional services) (With Intelligence). The structural argument (treated as root cause, not cyclical): agentic AI eliminates the human seats that per-seat SaaS subscriptions — the loan collateral — are priced on, and seat displacement manifests at contract-renewal cadence (annual/biennial), so credit losses surface in 12–18 months, not the 24–36 the models assume (AL Capital). The market signal is concrete: software stocks fell ~30% (Oct-2025→Feb-2026) while private-credit marks stayed ~flat — a "shadow drawdown gap" — and a record ~$25B of software loans now trade below 80¢ on the dollar vs near-par BDC marks. Q1-2026 stress is already visible: 54 BDCs reported new non-accruals, $1.4B newly / $5.1B total, software ~25% of portfolios with AI flagged as the developing credit risk, and 2026 software debt maturities of $12.7B, +73% YoY (SOLVE). UBS's severe scenario projects 13–15% default rates (vs 2–2.5% historical) if AI disruption proceeds unabated. This is the same forcing function as agentic-ai-seat-erosion-to-saas-rerate, expressed in credit.

(c) The per-name dispersion pair — long ARCC/BXSL, short FSK/OBDC/OWL/GBDC/OTF

The May-2026 BDC safety ranking sorts the names cleanly (AL Capital):

• Long / safest:
  • ARCC (Ares) — "best in class": 67 consecutive stable dividends, ~1.5% non-accruals, $6B liquidity buffer, 1.10× leverage, ~7% discount to NAV, 10.3% yield. Largest absolute software book ($6.6B across 135+ companies) but most diversified (Investing.com, stockanalysis).
  • BXSL (Blackstone Secured Lending) — 100% NII coverage ($0.77 = $0.77), lowest leverage (1.27×); ~26% software and a non-accrual uptick (3.1% FV per AL Capital; one source cited 4.7% at cost — basis difference), but conservatively structured. The "high-quality" leg, slightly noisier than ARCC.
• Short / riskiest:
  • FSK (FS KKR) — critical: NAV −9.9% to $18.83, non-accruals 8.1% cost / 4.2% FV, dividend cut 40% ($0.70→$0.42), Moody's Ba1 (junk), JPMorgan cut its credit facility $648M.
  • OBDC / OWL (Blue Owl) — OBDC downgraded to "Watch": NAV down 5 consecutive quarters to $14.41, dividend cut 16% to $0.31, adjusted NII missed by ~11%. Parent OWL down ~66% from peak; the redemption epicenter — OTIC 40.7%, OCIC 21.9% redemption requests (record), ~$4.2B unfulfilled (With Intelligence).
  • GBDC (Golub) — ~26% software, elevated non-accruals, 15% dividend cut taken.
  • OTF — pure-tech BDC (carryover from prior wiki) — highest structural software beta.

The trade is relative, not directional: redemption pressure + concentrated software + thin dividend coverage on the short leg vs diversification + coverage + dry powder on the long leg. The edge is the mark-lag: BDC software marks still cluster near par while the secondary and the equities have already moved, so the dispersion has further to run as marks catch down.

Contradictions and open questions

• BXSL non-accrual basis: 3.1% (FV, AL Capital) vs 4.7% (cost, round-1 search). Cost-vs-fair-value basis explains most of it, but BXSL is the noisiest member of the "safe" leg — watch its next mark.
• How fast do marks catch down? The whole pair depends on BDC marks converging toward the ~80¢ secondary. Fitch earlier framed redemptions as sentiment- not fundamental-driven; if managers extend-and-pretend, the dispersion can stay wide longer than the catalyst window.
• Does power-gating eventually become Feeder A? If grid constraints persist into 2027, some over-levered neocloud/SPV builders financed at 2025 assumptions could still default on timing (carry without revenue) even though aggregate demand is intact — a slower, smaller Feeder A, not zero.
• ARCC's absolute software book: largest in dollars; diversification is the bull case, but a system-wide SaaS default wave would still touch it — the long leg is "less bad," not immune.

Provenance

Rounds run: 2 (early-exit — round-2 per-name evidence resolved the long/short pair; no productive round-3 sub-question).

Sub-questions by round:

Round 1 (broad survey):

1. Is AI datacenter capex being cancelled/delayed or holding in mid-2026?
2. What are Q1-2026 BDC software loan marks / non-accruals (OBDC, OTF, FSK, GBDC)?

Round 2 (drill-down):

1. Per-name dividend/NAV/exposure dispersion — OBDC vs ARCC and the safest-vs-riskiest ranking — targeting the tradeable long/short pair.
2. Software-exposure %, non-accrual rates, and the AI-disruption-as-root-cause verdict — targeting which feeder dominates.

Anchor source: no Grokipedia entry fetched (fast-moving credit-markets topic; primary value is analyst/filing data).

URLs fetched (5 successful, 0 failed):

Round 1:

• BloombergNEF — AI data center build advances at full speed — industry — hyperscaler capex magnitude + power-constraint framing.
• SemiAnalysis — Stop Saying Half of 2026 US Datacenter Capacity Is Canceled — analyst — the ~1% revision rebuttal (also filed as a feed clipping today).
• SOLVE — BDC filings reveal new non-accruals in 1Q — data — $1.4B new / $5.1B total non-accruals; $12.7B software maturities +73%.

Round 2:

• With Intelligence — What is actually going on in BDC portfolios? — analyst — software exposure %, PIK dispersion, Blue Owl redemptions.
• AL Capital Advisory — Private Credit 2026: BDC crisis, default outlook & safest BDCs — analyst — the safest/riskiest ranking + per-name non-accruals + UBS severe scenario.

Tools used: WebSearch, WebFetch. Generated: 2026-06-19.