How the AI datacenter buildout is being financed. Neocloud debt, hyperscaler bond issuance, ABS structures, Stargate financing — and where the credit cycle gets uncomfortable.
This primer assumes a reader who knows the AI capex story at the level of “hyperscalers are spending a trillion dollars a year on data centers” but does not yet have the vocabulary for how that capex is actually being financed. By the end, the reader will be able to: distinguish the three financing layers that have emerged inside a single AI data center (the building, the power kit, and the GPUs are each financed differently); recognize the dominant private-credit and bank lenders; place current bond and term-loan pricing on a spread chart; and name the investable pockets at each tier of the credit stack.
The anchor question this primer answers is the one a fixed-income PM would ask: if the project’s other deliverables are right about token-demand surprise, the networking-interconnect bottleneck, the agentic-archetype capex tilt, and the unresolved chip-refresh-cycle question, what are the credit instruments that win or lose under each scenario, and what is on offer today?
Before naming a single instrument, the reader needs the framing that makes the rest of the primer coherent. The AI data center capex stack has split into three financing layers, each with its own clearing price, its own dominant lenders, and its own collateral logic. A single physical campus can carry IG-rated senior unsecured at the parent ($30B Meta bond 1), single-A securitized notes against the shell and tenant cash flows (Vantage’s €640M EUR ABS 2), and SOFR-floating term loans against the GPUs sitting in the racks (CoreWeave’s $8.5B DDTL 4.0 3) — all simultaneously, all priced at different points on the spread curve. The collateral has been split because the underlying economic lives are different: buildings depreciate over ~20 years, power infrastructure over ~25 years, GPUs over 3–6 years.
Reusable axes for the rest of the primer:
A worked vocabulary anchor: when a deal “moves from sub-IG to IG pricing,” what changed is almost never the GPUs. What changed is the tenant credit — the credit of the end-customer whose contract sits inside the deal’s bankruptcy-remote SPV. CoreWeave’s DDTL 4.0 at SOFR + 225 bps 3 and DDTL 5.0 at SOFR + 450 bps 4 are backed by physically identical GPU server pools; the 225-bp pickup is the price of sub-IG tenant credit, not different collateral. (“Tenant” here means the customer using the GPU compute under the long-term contract, not CoreWeave itself — CoreWeave is the operator-sponsor. The bankruptcy-remote SPV mechanic that makes this structure work is the entire subject of §7 below.) Same playbook at Applied Digital: ComputeCo 1 priced 9.250% in November 2025 5, ComputeCo 2 priced 6.750% four months later in March 2026 6 — a 250-bp tightening driven by a non-disclosed investment-grade hyperscaler tenant signing at Polaris Forge 2, not by GPUs becoming better collateral.
This three-tier framework is the unifying lens. Every named instrument in the rest of the primer fits inside one of these three boxes — and the bankruptcy-remote SPV mechanic in §7 is the structural engineering that makes Tier 2B and Tier 3 actually work.
§3 — Tier 1: the IG hyperscaler bond market. Between September 2025 and February 2026, the public hyperscalers plus Oracle ran the largest IG bond issuance window in tech-sector history — roughly $101B of on-balance-sheet debt directly attributable to AI capex, with Oracle issuing twice ($18B + $25B) and Meta, Amazon, and Alphabet returning to the bond market on a scale unprecedented for those names. Microsoft is the conspicuous outlier — funding through operating cash flow and ~$95B of finance-lease commitments instead.
Old flow. Oracle ran with modest investment-grade leverage, a stable A-rated balance sheet, and limited recourse to the bond market. The 2025 Stargate announcement and the $300B 15-year compute commitment to OpenAI changed the underlying need.
New flow. On September 24, 2025, Oracle priced an $18.0B multi-tranche IG deal across maturities 2030–2065 7, with the 5-year tranche at +75 bps to UST and the 40-year at +137 bps. The book reportedly peaked around $88B of demand — second-largest US IG deal of 2025. Moody’s had already moved Oracle’s outlook to negative in July 2025, flagging the expectation of continuing elevated leverage and increasingly negative free cash flow as Oracle ramped its AI infrastructure business [S302].
Then in February 2026, with the stock down ~50% from the September high, Oracle returned with a $50B financing package: $25B bonds across eight tranches 8, plus $20B of ATM (At-The-Market — sales of new shares directly into the open market through a broker) equity, plus a $5B mandatory convertible preferred 8. The bond book reportedly drew the largest order book in corporate-bond history (>$125B), though that order-book figure is from trade-press dealer commentary, not from the FWP (Free Writing Prospectus — the SEC filing that documents pricing terms but does not disclose subscription levels). Moody’s and S&P both moved Oracle to negative outlook on Baa2/BBB [S305]; Fitch held BBB-stable. Oracle committed for the first time to maintaining IG status as a formal policy and signaled no further bond issuance in 2026. September buyers are sitting on >$1B of mark-to-market losses.
What changed (quantified). Oracle’s debt-to-EBITDA moved from ~2.5× pre-AI ramp to well above 3× post-September 2025 issuance. The $5B mandatory convertible is a tell: when an A-rated issuer adds equity-linked paper inside an IG package, the rating agencies are signaling they want capital-structure repair, not more straight debt.
Meta priced $30B of senior unsecured notes across six tranches on October 30, 2025 1 — the largest single tech bond before the Oracle Feb 2026 deal. Order book reportedly around $125B. Ratings Aa3/AA-. Spreads to UST: +50 / +70 / +78 / +88 / +98 / +110 bps across 5y/7y/10y/20y/30y/40y. The headline tranche fact: Meta’s face long-term debt jumped from $29B at FY24 year-end to $59B at FY25, doubling in 12 months. Meta went from net-cash to structurally levered.
The companion off-balance-sheet structure landed two weeks earlier — and is covered in §5 below — the Hyperion JV adds another ~$26B of PIMCO-led private debt on top of the on-balance-sheet $30B.
Amazon returned to the bond market for the first time in roughly three years with $14.96B across six tranches on November 17, 2025 9. Order book ~$80B. The 40-year tranche tightened to +85 bps from +115 bps initial price talk. Goldman, JPM, and Morgan Stanley led. The context the deal sits inside: AWS Q3 2025 capex was up 61% year-over-year to $34.2B per quarter, and analyst commentary publicly flagged that Amazon’s $84B cash position would not cover the ~$150B AI capex run-rate likely to print in FY26.
Alphabet broke from its historical near-zero-debt posture. $5.0B USD notes in May 2025 (2030–2065, four tranches) plus a debut €6.75B euro offering on April 29, 2025 (five tranches, 2029–2054, peak demand €31.5B), plus a second €6.5B euro offering in November 2025 (~16-year weighted-average maturity, 3.44% weighted-average coupon) 10. Combined ~$13B-equivalent of fresh debt at a company that had been minimal on the debt side outside legacy 2016/2020 notes.
Microsoft is the outlier. Microsoft issued no major conventional bonds for AI capex in 2025 [S308]. The company carries ~$78B of cash against ~$40B of total debt and remains net-cash. The AI-capex exposure shows up instead in ~$95B of finance leases not yet commenced (peaked at ~$108B at June 30, 2024) — overwhelmingly data centers under leases with operators like Brookfield (10.5 GW power), Three Mile Island PPA, and multiple sale-leasebacks. Microsoft is funding through operating cash flow and lease commitments, not bonded debt. If you are screening for hyperscaler AI-capex exposure in the bond market, Microsoft has the smallest opportunity set because the exposure isn’t traded as straight debt.
Equinix (EQIX) dominates 2025-2026 IG DC issuance. Notable tranches: €750M 3.250% 2029 + €750M 4.000% 2034 in May 2025 (Equinix Europe 2 Finco; Baa2/BBB/BBB+; +117.6 / +150.9 bps to Bund) [S225]; $1.25B 4.600% Senior Notes due 2030 in November 2025, priced +85 bps to UST [S226]; C$700M 4.000% Senior Notes due 2032 in November 2025 (Canadian-dollar tranche, first such for Equinix) [S227]; and a $1.5B combined offering on March 5, 2026 ($700M 4.400% 2031 from Singapore Finco + $800M 4.700% 2033 from Europe 2 Finco). Moody’s upgraded EQIX senior unsecured to Baa1 ahead of the March 2026 offering 11. The EQIX 4.6% 2030s trade at ~85.9 bps OAS (Option-Adjusted Spread — the spread over the Treasury curve adjusted for embedded optionality) per UNC credit research (December 2025), which is tight for a Baa1/BBB+ name and a useful benchmark for the IG DC complex.
Digital Realty (DLR) ran a green-bond-heavy 2025, including substantial euro issuance. Iron Mountain (IRM), the BB-/Ba3 outlier in the DC complex, trades at ~214 bps OAS — about 130 bps wide of pure-play IG DC REITs, reflecting the storage-business mix and the higher leverage profile.
§4 — Tier 2A: data center ABS and CMBS. The data center securitization market has institutionalized at scale. 2025 calendar-year issuance reached $27B 12 (KBRA via MBA NewsLink roundtable), up from a 2018–2022 baseline of ~$4B per year and a 2023–2024 run-rate of ~$10B per year. CMBS share of total structured-finance issuance rose to 42% in 2025 from 25% in 2018-2024. The institutional bid is durable; pricing has tightened to 150–200 bps over US Treasuries for single-A, ~5-year weighted-average-life paper.
Old flow. Pre-2021, the data center sector financed through bank loans and IG REIT bond issuance against the parent balance sheet. Securitization existed in cell tower trusts and some hyperscaler-adjacent financings, but the data center sector itself was barely present.
New flow. Blackstone’s BX 2021-VOLT in Q4 2021 was the first pure data center single-borrower CMBS, financing part of the QTS take-private. Aligned Data Centers’ August 2021 inaugural $1.35B ABS was the first-ever green DC ABS 13. CyrusOne’s July 2024 $687M inaugural SASB CMBS was the first DC SASB (Single-Asset Single-Borrower — a CMBS structure where the bond is backed by one borrower’s portfolio rather than a pool of unrelated commercial mortgages) since the 2008 financial crisis 14. Since then, the master-trust master-issuer structure (where one trust vehicle accepts periodic ABS additions backed by an expanding pool of customer contracts as the operator grows) has become the dominant template.
What changed (quantified). Issuance jumped from ~$4B/year baseline (2018–2022) to $10B/year (2023–2024) to $27B in 2025 12 — a ~7× increase from the pre-AI baseline in three years. JLL projects 2026 issuance could reach ~$50B 15.
The pricing standard for a Compass / Switch / Vantage / DataBank / Cologix / TierPoint / Centersquare deal in 2025-2026 is 150–200 bps over USTs, single-A typical, 5-year WAL 24. This is roughly 50–70 bps wide of whole-business securitizations (which print 115–130 bps over USTs), reflecting newer asset class plus collateral specificity. Deal sizes have scaled: the largest CMBS deals now reach $4B, with $1–3B common.
§5 — Tier 2B: project finance, private credit, and the off-balance-sheet hyperscaler innovation. A new category emerged in 2025: large hyperscaler-adjacent private-credit transactions structured as joint ventures or fund vehicles, with the hyperscaler as long-dated tenant rather than borrower. The single most important deal — Meta’s Hyperion campus financing — established the template that defines the asset class in 2026.
A live correction up front: media chatter in summer 2025 framed Meta’s “Hyperion” Louisiana megacampus financing as an Apollo-led deal. The verified Oct 21, 2025 structure is a Blue Owl + Meta joint venture: Blue Owl-managed funds own 80% of the JV; Meta retains 20% 25. Total development cost ~$27B [S201]. Blue Owl contributed roughly $7B cash at JV formation; Meta received a one-time $3B distribution after contributing land and construction-in-progress. The debt — about $26B — was funded via private placement led by PIMCO and select other bond investors, with Morgan Stanley as sole bookrunner 25. Apollo, KKR, Brookfield, and Carlyle were in the final round of negotiations and lost [S239]. Trade press peg the debt at 24-year tenor, ~6.58% coupon, A+ rated, though those exact terms come from Bloomberg’s paywalled deep-dive rather than from a SEC-filed 144A prospectus visible to outside readers.
Two structural innovations make this the template. First, an off-balance-sheet JV with a short operating lease: Meta’s initial lease term is just four years with extension options, which keeps the debt off Meta’s books while Blue Owl carries the asset. The campus targets 2 GW by 2030, expandable to 5 GW, across roughly 4 million square feet 25. Second — and the load-bearing innovation — a 16-year residual value guarantee from Meta to the investors 25: if Meta terminates the lease early and the asset is worth less than a threshold, Meta reimburses investors for the shortfall. The RVG is what attracted PIMCO at investment-grade pricing despite the GPU-obsolescence risk underneath. The PIMCO appearance is the signal that traditional fixed-income money — not just specialty private credit — has moved into AI-DC project finance.
Stargate ($500B headline announced January 21, 2025) is a brand and a coordination vehicle, not a single balance-sheet entity 26. Equity ownership of the umbrella: SoftBank 40% / OpenAI 40% / Oracle ~$7B / MGX ~$7B, with SoftBank responsible for raising debt. Actual buildout happens at site-specific vehicles. The canonical site-specific financed deal is Abilene, Texas: 1.2 GW, 8 buildings, structured as a $15B Crusoe / Blue Owl Capital / Primary Digital Infrastructure JV, with JPMorgan-led construction loans of $9.6B total ($2.3B Phase 1 January 2025 + $7.1B Phase 2 May 22, 2025) 272829. Oracle signed a 15-year lease on the entire campus; Oracle sub-leases the GPUs to OpenAI; OpenAI’s payments to Oracle are ultimately backstopped by SoftBank capital. This is the canonical Stargate template, and it is closer to a traditional power-project finance with credit-tenant-lease characteristics than to a securitization.
| Sponsor | Major AI-DC commitments (2024–2026) | Capital deployed |
|---|---|---|
| Blue Owl Capital | Meta-Hyperion JV (80% ownership of $27B project); Crusoe-Abilene JV ($15B all-in); STACK Infrastructure, EdgeConneX in BODI portfolio | ~$40–45B+ AI-DC committed; BODI ~$14.2B AUM at YE2024 |
| PIMCO | Lead bond investor in Meta-Hyperion JV private placement | ~$26B in this single deal |
| Apollo | Valor-Compute Infrastructure $3.5B for xAI GB200 TNL; Stream Data Centers majority stake (Aug 2025); STACK Europe carve-out from Blue Owl (Apr 2025) | $40B+ cumulative next-gen infra since 2022 32 |
| Brookfield | Crusoe $750M senior facility; broader DC infrastructure book | $750M named + portfolio |
| KKR | $50B ECP partnership; CyrusOne ownership; 190 MW Bosque County CyrusOne JV | $50B ECP commitment, ~$4B Bosque deal |
| Macquarie | Applied Digital $5B preferred-equity facility + $100M dev loan | $5.1B+ committed AI-DC |
| JPMorgan | $9.6B+ construction loans into Stargate-Abilene Crusoe/Blue Owl JV; $1B senior secured for Lambda 40 | $10B+ committed AI-DC project finance |
Blue Owl is the AI-DC private-credit leader by dollars committed.
§6 — Tier 3: neocloud and AI-cloud debt — GPU-collateral term loans, senior unsecured high-yield, and convertibles. This is the tier where rates are widest, where collateral logic is most novel, and where the project’s other theses bite most directly. CoreWeave invented the IG-rated GPU-backed term loan in 2026; Applied Digital ran the same playbook in project-finance bonds; Nebius and IREN ran around the senior-debt market entirely via convertibles; xAI is in the wilderness at SOFR+700.
CoreWeave’s total principal of debt as of March 31, 2026 was $25.149B (net $24.859B after $290M original-issue discount) [S101], up from $21.615B at year-end 2025. The capital structure as of Q1 2026 10-Q:
The DDTL 4.0/5.0 pair is the central architectural innovation. Both facilities are backed by physically identical GPU pools, both run through bankruptcy-remote SPVs (Special Purpose Vehicles — legal entities created to isolate the financial risk of a specific pool of assets from the parent company’s balance sheet), both feature first-priority liens on the GPUs and the customer contracts. The 225-bp spread differential is entirely tenant credit: DDTL 4.0’s SPV is contracted to an investment-grade hyperscaler; DDTL 5.0’s SPV has “two large, non-investment-grade customer contracts” backing it.
APLD ran the same playbook for senior secured project-finance bonds. November 13, 2025: APLD ComputeCo LLC priced $2.35B at 9.250% senior secured notes due 2030, issued at 97% of par 5, funding Polaris Forge 1 (Ellendale ND) and repaying the SMBC Senior Loan. March 4, 2026: APLD ComputeCo 2 LLC priced $2.15B at 6.750% senior secured notes due 2031, issued at 98% of par 6 — a 250-bp coupon tightening in 4 months. Funds Polaris Forge 2 (Harwood ND, 200 MW). The tightening came from (a) Macquarie’s $5B preferred-equity facility backing the platform 36 and (b) the underlying Polaris Forge 2 campus contracting with a non-disclosed investment-grade hyperscaler tenant. May 4, 2026: APLD layered on a $300M senior secured 364-day bridge facility from Goldman Sachs at SOFR + 275 bps 45 for a third Polaris Forge 1 building — bridge implies ~7.6% all-in.
The APLD curve from 9.25% (Nov 2025) to 6.75% (Mar 2026) to ~7.6% bridge (May 2026) is the cleanest worked example of how an IG hyperscaler signature collapses cost of debt by ~250 bps. This is the playbook every neocloud and colo will run.
The neoclouds with rallying equity skip senior debt entirely and issue dilution-cheap convertibles.
The convertible coupons of 1.00% to 2.625% across NBIS and IREN are roughly one-eighth of the senior unsecured coupons CoreWeave is paying at 9.0–9.75%. The choice is dilution vs coupon: name with rallying equity and a high conversion premium runs the convertible route; name with already-issued equity at lower prices runs the senior debt route.
xAI is the high-yield outlier. Mid-2025: $5B debt + $5B equity, both led by Morgan Stanley (closed July 1, 2025) 48. Reported terms (Bloomberg coverage, paywalled in our local snapshot): floating-rate term loan B at SOFR + 700 bps priced at 97 OID, plus fixed-rate loan + secured bonds at ~12%. “Best efforts” deal — Morgan Stanley did not commit its balance sheet (a deliberate change from the 2022 Twitter financing). Subscription reportedly only ~1.5× vs. ~2.5–3× typical for high-yield paper.
Private neoclouds with bank relationships are accessing $750M–$1B syndicated senior secured facilities at undisclosed pricing inside the broker-dealer/bank channel. May 7, 2026: Lambda Labs closed a $1.0B syndicated senior secured credit facility, JPMorgan lead arranger 40, upsizing the August 2025 $275M facility by ~4×. Crusoe carries the $750M Brookfield facility plus the $300M February 2026 AMD-collateralized debt round (lower confidence — Sacra/Tracxn aggregations only) [S123]. Voltage Park / Together AI / RunPod have no material public debt disclosed [S125] — equity-funded.
| Layer of the Tier 3 stack | May 2026 indicative pricing | Anchoring example |
|---|---|---|
| IG GPU-backed term loan (IG hyperscaler tenant, SPV) | SOFR + 225 bps (~7.1% all-in) | CoreWeave DDTL 4.0 3; APLD Goldman bridge SOFR + 275 45 |
| Sub-IG GPU-backed term loan, publicly syndicated | SOFR + 450 bps (~9.3% all-in) | CoreWeave DDTL 5.0 4 |
| Senior secured project-finance bond, IG hyperscaler tenant | 6.75% fixed | APLD ComputeCo 2 2031s 6 |
| Senior secured project-finance bond, mixed tenant | 9.25% fixed (97 par) | APLD ComputeCo 1 2030s 5 |
| Senior unsecured high-yield (sub-IG neocloud) — secondary | 9.04% YTM (Z-spread 534 bps) | CRWV 9.25% 2030s [S103] |
| Senior unsecured high-yield (sub-IG neocloud) — new-issue | 9.75% coupon | CRWV April 2026 2031s 42 |
| Senior unsecured convertibles (sub-IG, mid-equity premium) | 1.0–2.625% coupon, 25–32.5% conversion premium | NBIS 2031/2033s 4646; IREN 47 |
| xAI-tier high-yield (single-name AI lab, best-efforts) | SOFR + 700 bps floating + ~12% fixed | xAI Morgan Stanley $5B July 2025 48 |
The gap between IG GPU paper at SOFR+225 and xAI single-name at SOFR+700 is roughly 475 bps. That is the price of full-faith hyperscaler tenant vs. single-name AI lab counterparty risk, holding the underlying GPU collateral roughly constant.
§7 — The structural mechanic underneath what we just saw. We just walked Tier 2B (Meta-Hyperion JV; Stargate-Abilene; Apollo-Valor / xAI) and Tier 3 (CoreWeave DDTL 4.0 / 5.0; Applied Digital ComputeCo 1 / 2; Nebius and IREN convertibles). Every one of those deals — across two tiers, eight named sponsors, and rating-agency assessments spanning A3 down to Ba2 — runs through some variant of the same legal-financial engineering: a bankruptcy-remote SPV. The mechanic that turned CoreWeave’s sub-IG corporate credit into A3 / A(low)-rated GPU paper, and that turned Blue Owl’s role as Meta’s joint-venture partner into ~$26B of PIMCO-buyable A+ rated notes, is what this section unpacks. The reader who learns it once stops being confused about “who is the tenant” forever.
A bankruptcy-remote SPV — Special Purpose Vehicle — is a separate legal entity created to isolate the financial risk of specific assets from the parent operator’s balance sheet. The operator (CoreWeave, Applied Digital, the Blue Owl-Meta JV) transfers a specific pool of GPUs and the customer contracts attached to them into the SPV. The SPV then borrows against those assets. The SPV’s lenders are paid directly from the customer-contract receivables, with senior liens on the GPUs as additional security.
The critical legal feature: if the parent operator goes bankrupt, the SPV does not. The SPV’s assets and creditors are walled off from the parent’s bankruptcy estate. Lenders to the SPV have first claim on its cash flows and collateral, ahead of the parent’s bondholders, equity holders, or unsecured creditors. This is what “bankruptcy-remote” means in legal practice — courts have repeatedly upheld these structures in REIT subsidiaries, aircraft equipment trusts, and cell-tower trusts. The doctrine is decades old; the application to GPU collateral is what’s new.
OPERATOR (sponsor) SPV LENDERS
───────────────── ──────────────
CoreWeave / APLD / Blue Owl-Meta JV Blackstone Credit /
│ PIMCO / etc.
│ transfers assets ▲
▼ into SPV │
┌───────────────────────────────────────┐ │ debt
│ BANKRUPTCY-REMOTE SPV │ │ service
│ ┌─────────────────────────────┐ │ │
│ │ Asset 1: GPUs (physical) │ │ ─────┘
│ │ Asset 2: Customer contract │ │
│ │ (receivable) │ │ ◄────┐
│ │ Asset 3: First-priority lien│ │ │ senior lien
│ │ for SPV lenders │ │ │ on both
│ └─────────────────────────────┘ │ │ assets
│ │ │
│ Cash IN: customer contract payments │ │
│ Cash OUT: SPV debt service ─────────┼──────┘
└───────────────────────────────────────┘
▲
│ contracted monthly payments
│ (the "rent" in this analogy)
│
CUSTOMER (the "tenant")
──────────────────────
The entity USING the GPU compute
under the long-term contract
e.g. Microsoft, Meta, OpenAI, AnthropicThe reader who confuses “tenant” with “operator” makes the same mistake every first-time observer makes. The vocabulary comes from real-estate finance, where the tenant is whoever pays the rent. In a GPU-financing SPV, the tenant is whoever pays the customer-contract fees — the end-user of the compute, not the operator who built and runs the cluster.
The lenders to a bankruptcy-remote SPV are protected from operator default by construction — that’s the whole point of the structure. What they cannot escape is tenant default: if the tenant stops paying the contract, the SPV’s cash flow stops, debt service halts, and the lenders foreclose on the GPUs (which then have to be redeployed to another customer in a possibly impaired market). The rating agencies underwrite this risk by looking at the tenant’s credit, the contract’s structural protections (length, take-or-pay clauses, residual value guarantees), and the GPUs’ liquidation value if the SPV has to redeploy. Operator credit is irrelevant at this layer — that’s the entire purpose of using an SPV in the first place.
CoreWeave is the operator-sponsor of both SPVs. CoreWeave’s parent-company corporate credit is sub-IG (Fitch BB-, secondary YTM 9.04% on the 2030 senior unsecured [S103]). But the two DDTL SPVs themselves carry different ratings driven by different tenants:
| DDTL 4.0 | DDTL 5.0 | |
|---|---|---|
| Operator-sponsor | CoreWeave | CoreWeave |
| Physical collateral | GPU servers | GPU servers (same operator pool, fungible) |
| Tenant (customer) | Single IG hyperscaler — undisclosed; likely Microsoft per CoreWeave’s disclosed customer concentration | “Two large non-investment-grade customer contracts” per CoreWeave PR — likely AI labs (OpenAI, Anthropic, others) |
| SPV size | $8.5B | $3.1B |
| SPV rating | A3 / A(low) 3 | Ba2 / BB+ 4 |
| Spread | SOFR + 225 bps (~7.1% all-in) | SOFR + 450 bps (~9.3% all-in) |
| Spread differential | (benchmark) | +225 bps for sub-IG tenant credit |
Same operator. Same physical GPUs, fungibly drawn from CoreWeave’s fleet. Same legal structure. The 225-bp spread differential is entirely the lenders’ price of accepting sub-IG customer-contract receivables instead of IG customer-contract receivables. Not GPU obsolescence; not collateral mix; not legal seniority. Tenant credit, all the way down.
Meta is the tenant: it signs a 4-year operating lease on the Louisiana campus, with a 16-year residual-value guarantee 25. Blue Owl Capital is the operator-sponsor: it owns 80% of the JV that owns the campus; Meta retains 20% 25. The JV itself is the bankruptcy-remote SPV. PIMCO and other bond investors lent ~$26B against the JV’s cash flows at IG pricing (trade press: 24-year tenor, ~6.58% coupon, A+ rated 25).
The bond rating reflects Meta’s credit, not Blue Owl’s. If Meta defaults on the lease, PIMCO and co-investors hold collateral with a 16-year RVG floor, which is what made the deal IG in the first place. If Blue Owl as JV operator defaults, the JV continues paying lenders from Meta’s lease receivables — Blue Owl’s own corporate credit is irrelevant. The deal is, in effect, a synthetic Meta IG bond with GPU-and-shell collateral on top.
The bankruptcy-remote SPV is decades-old plumbing in adjacent asset classes: - REITs — the operating subsidiaries are bankruptcy-remote from the parent so that mortgage lenders can underwrite individual properties. - Aircraft equipment trusts — Enhanced Equipment Trust Certificates (EETCs) since the 1990s let airlines (often sub-IG) finance fleets at IG pricing because the planes sit in bankruptcy-remote trusts with senior lien rights. - Cell-tower trusts — Aligned, Vantage, and the broader tower REITs use master-trust structures to issue IG-rated ABS against contracted tenant cash flows from sub-IG mobile carriers.
What’s new in 2026 is the application of this template to GPU collateral with long-dated hyperscaler tenant contracts. That combination produces the first IG-rated GPU paper in history (CoreWeave DDTL 4.0, March 31, 2026) and the first $26B IG private placement against AI-DC collateral (Meta-Hyperion, October 21, 2025). It is also why the IG-rated GPU-paper market is a per-deal-customized asset class rather than a fungible benchmark — each SPV’s rating depends on its specific tenant, and that tenant’s credit varies enormously across deals (Microsoft Aaa/AAA to xAI sub-IG to a private AI lab with no public rating at all).
The next time you see a Tier 2B or Tier 3 headline (“Apollo finances $X for Y AI lab,” “Blue Owl backs $Z for hyperscaler tenant W”), the question to ask is not “what’s the operator’s credit?” but “whose contract is in the SPV, and what’s their credit?” That’s the question the rating agencies are asking; it’s the question the spread is pricing.
The descriptive sections above name what exists. This section maps the actionable pockets to the project’s other deliverables: the BOM memo on token-demand and archetype-allocation, the Networking & Interconnect Primer on the $154B optical TAM and bottleneck-progression, and the unresolved question on chip refresh cycle that surfaced in our earlier conversation around the Burry / Research Affiliates depreciation thesis.
The BOM memo projects that the agentic-AI archetype will pull incremental capex above the eight-source consensus, with 48.4% of the five-year cumulative pool ($362.8B of $750.0B at Base × Concentrated knobs) flowing into Compute Silicon. The investment-grade hyperscaler bond complex is the cleanest long expression of that view — Meta at +50 to +110 bps 1, Amazon at ~+85 bps 9, Alphabet at ~3.44% weighted-average euro coupon 10 are pricing near historical IG tights despite issuing at unprecedented sizes. If the BOM memo is right about token demand and the deliverability gap (Janus Henderson’s 72-GW shortfall, BCG’s 50–80 GW US shortfall), the hyperscaler IG complex tightens further as cash-flow visibility lengthens. The trade is straightforward: buy the long end of Meta, Amazon, Alphabet, and Equinix; underweight Oracle relative to the IG complex given the negative outlook and the formally-stated debt-issuance pause [S305].
The second-order winner is the project-finance bond stack with IG hyperscaler tenant. APLD ComputeCo 2 6.75% 2031s 6 is a worked example: the IG tenant signature collapses cost of debt by ~250 bps and creates investment-grade-tenant exposure at sub-IG-sponsor pricing. As more sub-IG colocation and neocloud operators sign IG hyperscaler tenants on Polaris Forge 2-style deals, this becomes the dominant new-issue product in 2026-2027. The pool will be substantial: JLL projects $50B+ in 2026 issuance 15, much of which will flow through structures resembling APLD ComputeCo 2 rather than traditional master-trust ABS.
The Networking primer argues that the next leg of the bottleneck-progression sits in optical interconnect: Goldman’s $154B optical-networking TAM by 2028 (9× the 2026 figure of $15B), with CPO capturing $91B of that prize at 29% scale-out penetration. The first-order impact on the debt market is modest — networking content is a small share of total DC capex (per the BOM memo, Net+IC runs 6–10% of capex across most archetypes) — but the second-order effect is real: deals like Crusoe-Abilene with 1.2 GW of GB200/Rubin-era compute will require the optical layer rebuild that the networking primer describes. If networking costs run materially higher than current DC budgets assume, ABS deal sizes scale up; per-MW project-finance loans size up; and the lender base broadens. Equinix is the cleanest debt-side beneficiary: the IG DC REIT carries the colocation interconnect business that grows with the optical-density buildout, and its bond complex is the most defensive IG DC name.
The networking primer’s identification of Marvell-Celestial AI (and the broader CPO transition for accelerators) as a 2028+ TAM expansion has no direct first-order debt exposure today. The debt-side play is indirect: if Celestial AI’s $1B-by-2028 run-rate materializes inside Marvell, Marvell’s IG credit profile improves; if the broader CPO build forces the AI-DC capex stack to absorb the high-end CPO content the Networking primer sizes for 2026–2028, the IG hyperscaler bonds and DC REIT bonds funding that capex all see incremental issuance need.
The BOM memo’s central finding is that incremental capex tilts to the Agentic AI archetype (Compute Silicon 50.5% of BOM; Memory 17.0%) and away from Legacy Enterprise (Compute Silicon 25.5%; LSE 37.2%). Translated to the debt market: deals funding agentic-archetype facilities (CoreWeave, Nebius, Applied Digital ComputeCo 2 at Polaris Forge 2) carry more obsolescence risk per dollar of asset than deals funding Legacy Enterprise refresh (the shell is the asset; depreciates over 20 years). The CoreWeave DDTL 5.0 at SOFR+450 4 is more exposed to agentic-archetype-driven Compute Silicon turnover than the EQIX 4.6% 2030s at OAS+85.
This is also the structural reason for the Tier 1 vs Tier 3 spread — the Tier 1 hyperscaler bonds have shell + power + GPU exposure blended at the parent level, while Tier 3 GPU-backed term loans isolate the highest-obsolescence asset class. As agentic-mix tilts incremental capex toward Compute Silicon, expect more Tier 3 deals to emerge from non-CoreWeave names, with similar IG-vs-sub-IG bifurcation depending on tenant credit. Watch for the next CoreWeave-style DDTL 4.0 analog from a different sponsor — it would meaningfully expand the IG-rated GPU-backed paper float beyond CoreWeave’s single-name dominance.
Three of the project’s tracked architectural alternatives — Cerebras (wafer-scale; CS-3 SKUs; six datacenters announced), Google TPU v7, AWS Trainium 3, plus the broader MI300X/MI400 series, Microsoft Maia, Meta MTIA — challenge the implicit assumption underneath every Tier 3 deal that NVIDIA-GPU residual value is the relevant collateral price.
The cleanest exposure to this risk is the Apollo-Valor / xAI structure 32. NVIDIA is the anchor LP in VCI, which acquired $5.4B of GB200 GPUs to lease to xAI. If a non-NVIDIA architecture (wafer-scale, custom XPU, second-source GPU) substantially erodes GB200 residual value over the lease tenor, the triple-net-lease structure breaks: VCI investors are exposed to the residual-value gap, and the circular co-investment with NVIDIA does not insulate them. The same logic applies to CoreWeave’s DDTLs (sized to depreciable cost of GPU servers) and to the senior-secured ComputeCo bonds at APLD. Wafer-scale architectures don’t fit the existing collateral templates — a single Cerebras CS-3 system at ~$3M list with 23kW power draw is not directly substitutable in a borrowing-base formula sized to 70,000-GPU clusters.
The opposite read also holds: if NVIDIA continues to dominate (as the architecture-agnostic capex story in the BOM memo implicitly assumes), the existing collateral templates work, and the Tier 3 instruments mature into a settled asset class with tighter pricing over time. The IG GPU-paper-supported by RVG-type structures (Meta-Hyperion 16-year RVG; APLD’s IG hyperscaler tenant) is the most defensive Tier 3 exposure to architectural disruption — the residual value is contracted, not market-priced.
The unresolved thread from earlier in this project’s conversation: hyperscalers extended GPU useful life from 3–5 years in 2020 to 5–6 years now. If they’re forced to reverse course (because of architectural disruption, accounting standards push, or actual residual-value impairment realized through secondary-market GPU prices), the impact runs through three channels.
First, hyperscaler operating cash flow compresses (faster depreciation = lower reported net income for the same capex). This is a corporate-credit issue, not an asset-credit issue. Tier 1 IG bonds widen at the parent. Oracle’s existing negative outlook becomes a notch downgrade; Meta and Amazon could see watchlist actions. This is the channel where IG hyperscaler bond spreads have actual loss-given-default exposure.
Second, GPU-collateral DDTLs see borrowing-base compression. CoreWeave’s DDTL 4.0 and DDTL 5.0 are sized to a percentage of the depreciable cost of GPU servers. If hyperscaler depreciation schedules shorten — which signals lower expected residual value across the GPU asset class — the borrowing bases shrink, the SPV’s debt capacity shrinks, and refinancing the existing $5.8B drawn between DDTL 4.0/5.0 becomes structurally harder. This is the channel where the Tier 3 IG GPU paper carries the most counterintuitive risk: it is IG-rated because the tenant is IG, not because the GPUs are.
Third, the Meta-Hyperion-style RVG (residual value guarantee) becomes either the rescue mechanism or the source of impairment. If hyperscalers offer 15–20 year RVGs (as Meta did to PIMCO), they essentially absorb the chip-refresh-cycle risk that the broader market would otherwise have to underwrite. The credit transfers from “AI-DC residual value risk” to “Meta corporate credit.” This is good for IG investors in the Hyperion paper (and any future Microsoft / Amazon / Alphabet RVG-backed deal that copies the template) but bad for the hyperscaler issuing the RVG — Meta is now contingent-liable for residual value across a 16-year window inside an industry where the asset depreciates over 3–6 years. If you’re long Meta bonds, watch the next 4–6 Hyperion-style structures very carefully — each one adds 16-year contingent liability that may not flow through standard IG analyses today.
Buy / overweight. (1) IG hyperscaler bonds — Meta and Amazon at the long end as the cleanest beta to incremental token demand. (2) IG DC REIT bonds — Equinix the safest expression given Baa1 upgrade and tight OAS. (3) Senior secured project-finance bonds with IG hyperscaler tenant on long contract — APLD ComputeCo 2 6.75% 2031s as the worked example, and any subsequent deal at sub-IG sponsor + IG tenant signing. (4) Master-trust ABS at scaled operators — DataBank, Switch, Compass — institutional bid is durable and pricing has tightened to 150–200 bps over USTs.
Avoid / underweight. (1) Sub-IG senior unsecured at the neocloud parent — CoreWeave 9.25% 2030s and 9.75% 2031s carry issuer-credit risk that is widening (75 bps spread move from July 2025 to April 2026) even as the secured layer tightens. (2) Apollo-Valor and other single-name AI-lab structures with circular vendor financing — concentrated counterparty risk plus architectural-disruption exposure plus the lease-tenor / asset-life mismatch that even an RVG cannot fully bridge. (3) xAI debt at SOFR+700 — best-efforts deal, weak subscription, single-name single-product counterparty.
Relative value worth watching. Long IG DC REIT (EQIX) versus short sub-IG neocloud senior unsecured (CRWV 2030s) as a hedge for “AI capex is real, but the borrower mix matters more than the asset mix.” Long IG hyperscaler bonds versus short the broad investment-grade bond index as a sector-overweight expression of the token-demand thesis. Long the Compass / DataBank / Switch ABS complex versus short the broader CMBS index as a “DC asset class is durably bid” expression.
Counter-thesis (synthesized from the chip-refresh / Research Affiliates / Burry literature and from Stargate skeptics). The current credit-spread structure assumes (a) hyperscaler IG tenant signatures fully insulate the GPU obsolescence risk underneath; (b) the IG GPU-backed DDTL market scales beyond CoreWeave to a broader sponsor base; (c) the Stargate, Hyperion, and Apollo-Valor templates are durable rather than 2025-2026-vintage anomalies. The counter case: the chip-refresh cycle is structurally shorter than 5-6 years (closer to the Research Affiliates 3-year estimate), and when the market wakes up to it, the IG GPU paper bid disappears, the RVG-backed structures see Meta-class hyperscalers absorbing latent contingent liability that the rating agencies have not priced in, and the single-name AI-lab debt at SOFR+700 prices the right risk while the IG complex prices the wrong one. Apollo’s January 2026 ($3.5B to xAI via Valor) is the canonical at-risk template: NVIDIA as anchor LP, single AI-lab tenant, $5.4B of GB200 GPUs depreciating against an uncertain residual value, in a lease structure that requires xAI to keep paying for compute it may not need at the same intensity in 2028. The deeper counter-thesis writeup is the chip-refresh / depreciation thread from this project’s earlier conversation; the SemiAnalysis Microsoft AI Strategy piece (May 15, 2026) and the Fortune / Research Affiliates piece (April 15, 2026) carry the load-bearing facts. The right credit posture if you take this counter-thesis seriously is: avoid Tier 3 entirely, hedge Tier 1 exposure via SOX or NVDA equity hedges (the real risk transfers to the chip-vendor’s residual-value exposure), and concentrate Tier 2A ABS exposure in master-trust deals whose collateral is enterprise tenant base (Switch, Compass) rather than hyperscaler-anchored (Vantage New Albany).
Every numerical claim, quoted passage, and direct attribution in this primer is sourced from one of 85 research findings, organized as:
Each finding carries the claim, the exact value with units, the source publication date, the entity’s latest event date, a confidence grade, and verbatim source quotes where applicable. The research packet was audited line-by-line for citation integrity: roughly four-fifths of findings passed clean on the first pass, and the remaining findings were either repaired or flagged where the underlying primary is paywalled or access-blocked but corroborates against widely-reported coverage.
Sister deliverables: the Networking & Interconnect Primer — the $154B optical TAM, CPO transition, bottleneck-progression framing — and the AI Data Center CapEx by Archetype IC Memo — the Gap Allocation Thesis, the 5-archetype taxonomy, and the 48.4% Compute Silicon allocation under Base × Concentrated.