The AI Neocloud + Stargate Debt & Capacity-Commitment Circularity

Forensic credit map — CoreWeave (CRWV), Oracle (ORCL), Stargate JV, and the neocloud complex. Prepared 2026-06-05. Every figure is cited inline. Where a figure is a third-party estimate, a press leak, or a non-binding "letter of intent," it is flagged explicitly.

Executive thesis

The AI buildout is being financed disproportionately by debt and capacity commitments, not by cash flow or equity. Morgan Stanley estimates ~$3T of hyperscaler/data-center capex through 2028, with internal cash funding only ~half and a ~$1.5T financing gap to be filled by credit markets (Energynow/Bloomberg). The collateral is GPUs, whose useful life Moody's pegs at 4-6 years versus 10-15 year leases — an asset/liability duration mismatch (Yahoo/Moody's).

The keystone circular structure: NVIDIA invests equity in / backstops neoclouds → neoclouds borrow against the GPUs they buy from NVIDIA → they sell that capacity to OpenAI → NVIDIA separately funds OpenAI, which also commits ~$300B to Oracle, which is itself debt-financing the chips it rents at ~14% gross margin. The same dollar of NVIDIA capital and the same GPU appear as revenue, collateral, and demand at multiple nodes.

1. CoreWeave (CRWV)

IPO (March 2025)

• Priced 37.5M Class A shares at $40.00 on 2025-03-27 (below the $47-$55 marketed range), raising ~$1.5B — the biggest US tech IPO since 2021. Trading began 2025-03-28 on Nasdaq under CRWV. (CoreWeave PR; CNBC)

Revenue & customer concentration

• FY2024 revenue $1.9B, net loss $863M (levelheadedinvesting). FY2025 revenue $5,131M, net loss $(1,167)M (CoreWeave Q4/FY25 results).
• Microsoft = 62% of FY2024 revenue; two customers = 77% (mlq.ai; ainvest). Microsoft = 67% of FY2025 revenue per the FY2025 10-K, with mgmt expecting it to fall below 50% as OpenAI/Meta ramp (CRWV 10-K FY2025; indexbox).

OpenAI contracts (cumulative ~$22.4B)

• Mar 2025: initial deal up to $11.9B (CoreWeave).
• May 2025: expansion up to $4B.
• Sep 2025: further $6.5B expansion → cumulative ~$22.4B (CoreWeave PR; Bloomberg).

Backlog / RPO

• Revenue backlog $30B (Q2'25) → $55.6B (Q3'25, +271% YoY) → $66.8B (12/31/2025) (CRWV Q3'25 8-K; Fortune; FY25 results).
• Q3'25: no single customer >~35% of backlog; >60% investment-grade. Q3 added Meta ~$14.2B and Poolside (Nasdaq).

Debt — total and GPU-collateralized facilities

• Total debt 12/31/2025 = $21,373M ($6,708M current + $14,665M non-current); cash ~$3.1B; capex FY2025 = $10,309M (CoreWeave FY25 results; CNBC Q4). Up from >$8B at YE2024 (levelheadedinvesting).
• DDTL 1.0 — Aug 2023: $2.3B GPU-collateralized facility (Blackstone, Magnetar) (ABF Journal).
• DDTL 2.0 — May 2024: $7.5B facility led by Blackstone & Magnetar (w/ Coatue), secured against the NVIDIA GPU fleet (Blackstone PR; CNBC).
• DDTL 4.0 — closed ~Jan 2026: $8.5B delayed-draw term loan, rated A3 (Moody's) / A (low) (DBRS) — the first investment-grade GPU-backed financing, secured by HPC infrastructure + an associated customer contract (CoreWeave PR).
• Aggregate GPU-backed debt estimated ~$10B (Yahoo/Chanos).

NVIDIA ties (equity + backstop)

• Equity: NVIDIA held ~24.3M shares (~$3.96B, its largest equity holding, ~91% of its portfolio) at end of Q2'25 (Motley Fool). Added $2.0B at $87.20/sh (22,935,780 shares) completed 2026-01-23 (CRWV 8-K).
• Backstop / take-or-pay: NVIDIA obligated to buy unsold CoreWeave capacity up to $6.3B through April 2032 — agreement signed 2023, disclosed in SEC filing Sep 2025 (Coindesk; Yahoo).

2. Oracle (ORCL)

The ~$300B OpenAI contract

• ~$300B over 5 years, ~$30B/yr, 4.5 GW incremental Stargate capacity, commencing 2027 (Built In; OpenAI; DCD).

RPO surge

• RPO $317B (Q4 FY25) → $455B (Q1 FY26, +359% YoY); cloud RPO +~500%; ~33% to convert within 12 months. Catz: RPO likely >$500B (Oracle Q1 FY26 PR; DCD).
• OCI revenue guided $18B (FY26) → $32B → $73B → $114B → $144B over following years (Oracle PR).

Capex & debt financing

• $18B notes issued 2025-09-26 (maturities 2030-2065, coupons 4.450%-6.100%) to pre-fund the AI buildout (Oracle 424B2; DCD).
• ~$91.3B senior unsecured borrowings outstanding as of 2025-08-31 (Oracle 10-Q FY26 Q1).

Thin/negative AI cloud margins

• Internal docs (via The Information, Oct 2025): in the 3 months to Aug 2025, ~$900M of Nvidia-GPU rental revenue produced only ~$125M gross profit = ~14% gross margin (vs ~70% legacy software). Average AI-cloud deal margin reportedly ~16%; struggling to clear 25% on 1-2 year-old chips (CNBC; DCD; The Information).

3. Stargate JV

• Announced 2025-01-21: company to invest $500B over 4 years; $100B initial (OpenAI).
• Equity (committed): SoftBank and OpenAI $19B each (40% each); Oracle and MGX $7B each (S&P Global; Wikipedia/Stargate LLC). SoftBank = financial lead; OpenAI = operational lead; Masayoshi Son chairman.
• Announced vs committed: $500B/10 GW is a target by 2029; the headline includes already-announced projects. By Sep 2025, OpenAI/Oracle/SoftBank said five new sites + Abilene push the program to ~7 GW and >$400B over three years (OpenAI).
• Sites & SPV financing: Flagship Abilene, TX built by Crusoe / Primary Digital Infrastructure / Blue Owl Capital JV, leased to Oracle (Wikipedia). JPMorgan lent $2.3B (2025-05-22) to the Abilene project (Wikipedia). Vantage Data Centers taking on ~$38B of debt for two sites (TX, WI) (wheresyoured.at).

4. Other neoclouds & circular NVIDIA ties

• Crusoe: ~$425M GPU-backed debt (Yahoo/Chanos); NVIDIA participated in $1.4B Series E (Oct 2025) at $10B valuation (TechCrunch).
• Lambda: ~$500M GPU-backed debt; NVIDIA backstop $1.5B capacity deal; NVIDIA in $480M Series D (Yahoo; TechCrunch).
• Nscale: NVIDIA $433M SAFE + £500M commitment (Oct 2025); soared to $14.6B valuation (TechCrunch; DCK).
• Nebius: Microsoft contract up to $19.4B (Sep 2025); NVIDIA in $700M raise (Dec 2024) and a further $2B NVIDIA investment (SEC 6-K; NVIDIA newsroom).
• NVIDIA also separately backs $860M of a partner data center's lease obligations (DCD).

5. Systemic: how much is debt-financed + GPU-depreciation risk

• Morgan Stanley: ~$3T data-center capex through 2028; cash funds ~half; ~$1.5T financing gap to credit markets across secured/unsecured/structured/securitized (Energynow; Morgan Stanley).
• Moody's: $662B of not-yet-commenced hyperscaler lease commitments (Amazon, Meta, Alphabet, Microsoft, Oracle) = 113% of their combined adjusted debt; $969B total undiscounted future lease commitments at YE2025. AI semis useful life 4-6 yrs vs 10-15 yr leases (Yahoo/Moody's).
• GPU ABS: First GPU-backed ABS (~$1.1B shelf) in early 2025, AAA notes inside ~110 bps; issuance forecast ~$8B (2025) → ~$25B (2028) (airealist; Medium).
• Short-seller flag: Jim Chanos warns of "debt defaults" in the GPU-collateral debt market (Yahoo).

Debt table

Edge (relationship) table

Unverifiable / soft data

• Vantage's "~$38B" project debt is sourced to commentary (wheresyoured.at) citing reporting, not a primary filing.
• Lambda/Crusoe GPU-backed debt amounts (~$500M / ~$425M) are from a Chanos-referencing aggregation, not the borrowers' filings.
• NVIDIA→OpenAI "$100B" was a letter of intent; only ~$30B was reportedly finalized (Feb 2026, part of OpenAI's ~$110B round) — figure still firming.
• Stargate's $500B is a multi-year target; only ~$100B/initial $52B of equity is committed; the rest is announced, not funded.
• The Information's Oracle margin docs (14%/16%) are leaked internal documents, not audited disclosures.

The duration-mismatch layer (added 2026-06-11)

The debt table above understates the problem, because it counts the financing but not the useful life of what it bought. Oracle's FY2026 8-K shows the pivot in stark form: free cash flow −$23.7B, ~$50B capex, >$108B debt (+$30B raised), and $248B of additional 15–19-year datacenter leases, substantially off-balance-sheet — all against GPUs whose economic life is ~2–3 years. That gap is the subject of the fin-ai-depreciation-debttrap block and the depreciation_trap Z3 proof (D1–D4): when asset life < financing tenor, there is an interval where the asset is worthless but the debt is still owed, and the "useful life" hyperscalers book (5–6 yr) is itself a self-marked number inflating today's profit (~$176B industry overstatement 2026–28, per Burry). So the CoreWeave/Oracle leverage here is not just concentrated (67% one customer; >half the $523B backlog on OpenAI) — the real assets backing it amortize slower in the accounts than they age out in reality. CNBC's framing — "building yesterday's data centers with tomorrow's debt" — is the duration mismatch in one line.

← Research index · structured data: fin-coreweave-oracle.json · fin-coreweave-oracle.md