表外融資的 AI 時代: SPV 如何為資料中心熱潮 融資 卻隱藏槓桿風險

Published:
隨手記
by Ernest Chiang

Post Title Image (Photo by Ray Hennessy on Unsplash)

✳️ tl;dr

  • Meta 透過 SPV 結構完成近 $300 億美元融資
  • 建設路易斯安那州 Hyperion 資料中心 (數據中心)，創下史上最大私募股權交易記錄 1
  • Hyperion 資料中心佔地 400 萬平方英尺，
  • 全面運作時耗電 5 gigawatts，相當於 400 萬戶美國家庭用電量
  • Meta 僅保留 20% 股權，卻維持完全營運控制權，
  • 這種「控制但不合併」的會計技巧讓 $270 億美元債務不出現在資產負債表上 2
  • 股權僅佔總融資的 8.5% ($25 億/$295 億)
  • 保險公司透過私募信貸大量投資此類專案，
  • 但面臨資產負債不匹配風險，在經濟下滑時可能被迫清算投資 3
  • 歷史教訓: 1990 年代電信公司鋪設 8000 萬英里光纖，
  • 泡沫破裂四年後仍有 85%-95% 未使用，被稱為「暗光纖」4
  • Meta、Amazon、Google、Microsoft 承諾今年資本支出達創紀錄的 $3200 億美元,大部分用於 AI 基礎設施,但 Meta 10-K 坦承:「無法保證 AI 的使用將增強我們的產品或服務」5
  • 電力基礎設施成為瓶頸之一
  • Grid Strategies 估計到 2030 年資料中心需新增 60 gigawatts 電力，相當於義大利全國峰值需求 6
  • 冷卻技術也是關鍵之一: 從氣冷到直接液冷再到浸入式冷卻，影響長期營運成本 7
  • Morgan Stanley 作為獨家承銷商，同時為多個類似專案提供融資諮詢 1
  • 債券以 144A 格式私募發行，利差高出國債 2.25% 1
  • 技術管理者可以做什麼？ (1) 量化 AI 投資的實際 ROI 時間表 (2) 評估電力供應鏈風險 (3) 保持適當的財務槓桿比例

8 9

✳️ 知識圖譜

(更多關於知識圖譜…)

%%{init: {'theme':'default'}}%%
graph LR
    Meta[Meta Platforms]:::instance
    SPV[Special Purpose Vehicle]:::concept
    BlueOwl[Blue Owl Capital]:::instance
    PIMCO[PIMCO]:::instance
    MorganStanley[Morgan Stanley]:::instance
    
    Hyperion[Hyperion Data Center]:::instance
    AIInfra[AI Infrastructure]:::concept
    DataCenter[Data Center]:::concept
    GPU[GPU Clusters]:::concept
    
    PrivateEquity[Private Equity]:::concept
    DebtFinancing[Debt Financing]:::concept
    EquityFinancing[Equity Financing]:::concept
    OffBalanceSheet[Off-Balance-Sheet Financing]:::concept
    
    CreditRisk[Credit Risk]:::concept
    Leverage[Financial Leverage]:::concept
    AssetLiability[Asset-Liability Mismatch]:::concept
    Opaque[Opaque Risk]:::concept
    
    InsuranceCo[Insurance Companies]:::concept
    RatingAgency[Rating Agencies]:::concept
    InvestmentGrade[Investment Grade Bonds]:::concept
    
    Power[Power Infrastructure]:::concept
    Entergy[Entergy]:::instance
    Gigawatt[5 Gigawatt Consumption]:::instance
    
    DotCom[Dot-com Bubble]:::instance
    Telecom[Telecom Fiber Buildout]:::instance
    GFC[Global Financial Crisis]:::instance
    
    Meta -->|retains 20% in| SPV
    BlueOwl -->|co-owns| SPV
    PIMCO -->|provides debt to| SPV
    MorganStanley -->|arranges financing for| SPV
    
    SPV -->|owns| Hyperion
    Meta -->|operates as tenant of| Hyperion
    Hyperion -->|supports| AIInfra
    AIInfra -->|requires| GPU
    
    PrivateEquity -->|structures| DebtFinancing
    PrivateEquity -->|structures| EquityFinancing
    DebtFinancing -->|funds| SPV
    EquityFinancing -->|funds| SPV
    
    SPV -->|enables| OffBalanceSheet
    OffBalanceSheet -->|hides| Leverage
    Leverage -->|amplifies| CreditRisk
    
    SPV -->|creates| Opaque
    Opaque -->|leads to| CreditRisk
    Leverage -->|causes| AssetLiability
    
    InsuranceCo -->|invests in| PrivateEquity
    InsuranceCo -->|faces| AssetLiability
    RatingAgency -->|assigns| InvestmentGrade
    InvestmentGrade -->|applies to| SPV
    
    Hyperion -->|demands| Power
    Entergy -->|supplies| Power
    Hyperion -->|consumes| Gigawatt
    
    DataCenter -->|parallels| Telecom
    Telecom -->|occurred during| DotCom
    CreditRisk -->|may trigger| GFC
    
    classDef concept fill:#FF8000,stroke:#CC6600,stroke-width:2px,color:#000
    classDef instance fill:#0080FF,stroke:#0066CC,stroke-width:2px,color:#FFF

%%{init: {'theme':'default'}}%%
graph TD
    Start[Meta needs $30B for Data Center]
    Decision1{Use Traditional Financing?}
    
    Traditional[Traditional Corporate Debt]
    Impact1[Adds debt to balance sheet]
    Impact2[Hurts credit rating]
    Impact3[Limits future borrowing]
    
    SPVRoute[Create SPV Structure]
    Setup1[Meta retains 20% equity]
    Setup2[Blue Owl co-owns 80%]
    Setup3[PIMCO provides $27B debt]
    Setup4[Private equity provides $2.5B equity]
    
    Benefits1[Debt off Meta balance sheet]
    Benefits2[Maintains credit rating]
    Benefits3[Meta still controls operations]
    
    Risks1[Opaque risk hidden]
    Risks2[High leverage 90% debt]
    Risks3[Insurance companies exposed]
    
    Outcome[Hyperion Data Center Built]
    
    Start --> Decision1
    Decision1 -->|Yes| Traditional
    Decision1 -->|No| SPVRoute
    
    Traditional --> Impact1
    Impact1 --> Impact2
    Impact2 --> Impact3
    
    SPVRoute --> Setup1
    SPVRoute --> Setup2
    Setup1 --> Setup3
    Setup2 --> Setup4
    
    Setup3 --> Benefits1
    Setup4 --> Benefits2
    Benefits1 --> Benefits3
    
    Benefits1 -.->|Hidden| Risks1
    Setup3 -.->|90% debt| Risks2
    Setup4 -.->|Funded by| Risks3
    
    Benefits3 --> Outcome
    Risks1 -.-> Outcome
    Risks2 -.-> Outcome
    Risks3 -.-> Outcome
    
    style Start fill:#FFD700,color:#FFF
    style SPVRoute fill:#90EE90,color:#FFF
    style Traditional fill:#FFB6C1,color:#FFF
    style Outcome fill:#87CEEB,color:#FFF
    style Risks1 fill:#FF6347,color:#FFF
    style Risks2 fill:#FF6347,color:#FFF
    style Risks3 fill:#FF6347,color:#FFF

%%{init: {'theme':'default'}}%%
graph TB
    Total[Total Financing: $29.5B]
    
    Equity[Equity: $2.5B - 8.5%]
    Debt[Debt: $27B - 91.5%]
    
    MetaEquity[Meta Equity: $0.5B - 20% ownership]
    BlueOwlEquity[Blue Owl Equity: $2B - 80% ownership]
    
    PIMCODebt[PIMCO Senior Debt: $27B+]
    OtherDebt[Other Investors: allocation]
    
    Terms[Bond Terms]
    Term1[Maturity: 2049 - 24 years]
    Term2[Fully Amortizing]
    Term3[Spread: +225 bps vs Treasury]
    Term4[Rating: A+ by S&P]
    Term5[Format: 144A private placement]
    
    Total --> Equity
    Total --> Debt
    
    Equity --> MetaEquity
    Equity --> BlueOwlEquity
    
    Debt --> PIMCODebt
    Debt --> OtherDebt
    
    Debt --> Terms
    Terms --> Term1
    Terms --> Term2
    Terms --> Term3
    Terms --> Term4
    Terms --> Term5
    
    style Total fill:#4169E1,color:#FFF
    style Equity fill:#32CD32,color:#FFF
    style Debt fill:#FF6347,color:#FFF
    style MetaEquity fill:#87CEEB,color:#FFF
    style BlueOwlEquity fill:#87CEEB,color:#FFF
    style Terms fill:#FFD700,color:#FFF

%%{init: {'theme':'default'}}%%
graph LR
    Tech[Tech Companies Need AI Infra]
    Capital[Massive Capital Requirement]
    
    SPVCreate[Create SPV Structure]
    OffBalance[Keep Debt Off Balance Sheet]
    
    PE[Private Equity Seeks Yield]
    Insurance[Insurance Co Provides Capital]
    
    HighYield[Lend at Higher Rates]
    LowRisk[To Investment Grade Companies]
    
    Hidden[Risk Appears Low]
    Actual[But Actual Risk is High]
    
    Overbuild[Signal to Overbuild]
    ThinCushion[Only 8.5% Equity Cushion]
    
    Mismatch[Asset-Liability Mismatch]
    Concentration[Concentration in AI Data Centers]
    
    Bubble[Potential Bubble Formation]
    Trigger[If AI ROI Disappoints]
    Crisis[Cascade of Defaults]
    
    Tech --> Capital
    Capital --> SPVCreate
    SPVCreate --> OffBalance
    
    PE --> Insurance
    Insurance --> HighYield
    HighYield --> LowRisk
    
    LowRisk --> Hidden
    Hidden -.-> Actual
    
    Capital --> Overbuild
    SPVCreate --> ThinCushion
    
    Insurance --> Mismatch
    PE --> Concentration
    
    Overbuild --> Bubble
    ThinCushion --> Bubble
    Mismatch --> Bubble
    Concentration --> Bubble
    
    Bubble -.->|If| Trigger
    Trigger -.->|Leads to| Crisis
    
    style Tech fill:#87CEEB,color:#FFF
    style SPVCreate fill:#90EE90,color:#FFF
    style Actual fill:#FF6347,color:#FFF
    style Bubble fill:#FFA500,color:#FFF
    style Crisis fill:#8B0000,color:#FFF

%%{init: {'theme':'default'}}%%
graph TD
    Timeline[Historical Comparison]
    
    Era1[Dot-com Era 1998-2001]
    Telecom1[Telecom Companies]
    Telecom2[Massive Fiber Buildout]
    Telecom3[Overinvestment in Infrastructure]
    Telecom4[Speculation on Internet Growth]
    
    Collapse1[2001-2002 Telecom Crash]
    Loss1[Billions in Write-offs]
    Bankrupt1[Multiple Bankruptcies]
    
    Era2[AI Era 2023-2025]
    AI1[Tech Giants and Hyperscalers]
    AI2[Massive Data Center Buildout]
    AI3[Overinvestment in AI Infrastructure]
    AI4[Speculation on AI Adoption]
    
    Current[Current Status 2025]
    Spend1[AI Spending > Telecom Peak]
    Spend2[Thin Equity Cushions]
    Spend3[Off-Balance Sheet Structures]
    
    Question[Future Unknown]
    Path1[AI ROI Materializes]
    Path2[AI ROI Disappoints]
    
    Success[Infrastructure Pays Off]
    Failure[Potential Crisis]
    
    Timeline --> Era1
    Timeline --> Era2
    
    Era1 --> Telecom1
    Telecom1 --> Telecom2
    Telecom2 --> Telecom3
    Telecom3 --> Telecom4
    Telecom4 --> Collapse1
    Collapse1 --> Loss1
    Collapse1 --> Bankrupt1
    
    Era2 --> AI1
    AI1 --> AI2
    AI2 --> AI3
    AI3 --> AI4
    AI4 --> Current
    
    Current --> Spend1
    Current --> Spend2
    Current --> Spend3
    
    Current --> Question
    Question -.->|Scenario A| Path1
    Question -.->|Scenario B| Path2
    
    Path1 --> Success
    Path2 --> Failure
    
    style Era1 fill:#DDA0DD,color:#FFF
    style Collapse1 fill:#8B0000,color:#FFF
    style Era2 fill:#87CEEB,color:#FFF
    style Question fill:#FFD700,color:#FFF
    style Success fill:#90EE90,color:#FFF
    style Failure fill:#FF6347,color:#FFF

✳️ 延伸閱讀

  1. https://www.reuters.com/legal/transactional/meta-set-clinch-nearly-30-billion-financing-deal-louisiana-data-center-site-2025-10-16/ ↩︎ ↩︎ ↩︎

  2. https://paulkedrosky.com/weekend-reading-plus-spvs-meta-and-fiber-buildout-2-0/ ↩︎

  3. https://gfmag.com/insurance/private-equitys-growing-role-in-insurance-rewards-and-risks/ ↩︎

  4. https://fortune.com/2025/09/28/ai-dot-com-bubble-parallels-history-explained-companies-revenue-infrastructure/ ↩︎

  5. https://www.axios.com/2025/09/25/nvidia-openai-data-center-boom ↩︎

  6. https://www.cnbc.com/2025/10/17/ai-data-center-openai-gas-nuclear-renewable-utility.html ↩︎

  7. https://www.deloitte.com/us/en/insights/industry/power-and-utilities/data-center-infrastructure-artificial-intelligence.html ↩︎

  8. https://www.fabricatedknowledge.com/p/lessons-from-history-the-rise-and ↩︎

  9. https://www.datacenterfrontier.com/hyperscale/article/55131851/in-ai-arms-race-data-centers-are-the-table-stakes-for-hyperscale-players ↩︎