The neocloud refinancing wall (2026-2028)

The AI-debt block (spec-ai-datacenter-debt-spv) mapped who lends; this one maps when it comes due - and it collides with the Fed hike (spec-fed-hike-2026).

CoreWeave - the documented case

From Q2 2026 filings: total debt ~$35.1B vs ~$5.0B equity + ~$6.9B cash. Contractual principal repayments:

• $4.4B rest-of-2026 - $6.2B in 2027 (the key hurdle; largest single year before 2030) - $4.4B in 2028 - $2.4B (2029) - $3.2B (2030) - $14.9B thereafter.
• Quarterly interest expense ~$640M (2.4x YoY). Rates span converts ~2% / senior notes ~10% / oldest delayed-draw facility ~15% / a new $2.6B facility at SOFR+550bps (2031).
• Ba3 (Moody's) / B+ (S&P) - firmly high-yield. HSBC flagged a ~$8.9B 2026 liquidity shortfall (analyst estimate). Capex guidance $35-39B/yr keeps the treadmill running.

Oracle - bigger, staggered, still climbing

Total borrowings ~$129.5B (May 2026, up from ~$92.6B). Raised ~$43B debt + $5B equity in FY2026; plans ~$40B in FY2027 (incl a $20B ATM equity issue); says no new debt in CY2026. Q1-FY2027 interest expense $1.428B, +55% YoY. Maturities are staggered (some decades out) - but the year-by-year table isn't reliably public here, so treat per-year figures as to-verify in the 10-K.

Why it's a "wall"

Neocloud + hyperscaler-SPV maturities concentrate into a tightening 2026-2028 window - a refinancing test: roll at higher rates, or contract. Three things make it bite at once:

1. The Fed hike raises refi cost exactly as the wall arrives (floating facilities reprice; new issues price off higher Treasuries).
2. Collateral is depreciating (GPUs) while the debt is short + expensive - the duration-mismatch trap made concrete.
3. Revenue is concentrated in a few, often unrated anchors (e.g. Anthropic ~40% of one CoreWeave loan) - hard-to-price counterparty risk.
4. Private-credit funds hold much of it - so a failed roll transmits into Blackstone/Blue Owl/Apollo/Pimco/BlackRock portfolios, not just the borrowers'.

Honest limits

CoreWeave's year-by-year schedule + interest expense are fact from filings; Oracle's aggregate debt/interest are fact but its per-year maturity table is graded (verify in 10-K); analyst shortfall/rate figures are reported estimates; the "test/trap/contagion" framing is interpretation. Structural overlay - no financial-core edges added.

Sources: CoreWeave Q2-2026 disclosures; Oracle FY2026 10-K + Q1-FY2027 10-Q; Motley Fool; Capacity; DCD; HSBC; FT. Cross-refs: CoreWeave, Oracle, Nebius, SINK_bondmarket, Fed_Funds_Rate, AI_Datacenters, PrivateCredit_Funds.

← Research index · structured data: spec-neocloud-refinancing-wall.json · spec-neocloud-refinancing-wall.md