核心摘要 · English Summary
Alphabet, Amazon, Meta, Microsoft and Oracle are using corporate guarantees and special-purpose vehicles to fund the AI build-out while keeping roughly $300bn of exposure off their balance sheets. Google's promise to cover data-centre lease payments rose from $16.9bn to $43.8bn in six months, yet less than 2% appears as an actual liability. Wall Street gets cheaper funding; the parent keeps its investment-grade rating. The risk only becomes real if a tenant defaults or AI demand disappoints.
核心摘要 · 中文版
据《金融时报》报道,Alphabet、亚马逊、Meta、微软与甲骨文正通过"企业担保 + 特殊目的载体"为 AI 基础设施融资,把约 3000 亿美元敞口留在资产负债表之外。谷歌为数据中心租赁付款提供的担保在半年内从 169 亿美元升至 438 亿美元,但真正计入负债的部分不足 2%。摩根士丹利统计七家云与芯片巨头的表外承诺与信用支持已超 3.1 万亿美元。这套结构让华尔街拿到更便宜的资金、让母公司保住投资级评级——只有租户违约或 AI 需求不及预期时,风险才会"显形"。
完整原文 · Full Text
Big Tech is leaning on corporate guarantees and off-balance-sheet vehicles to keep roughly $300 billion of AI infrastructure exposure out of its financial statements, part of nearly $3 trillion in commitments investors can barely see.
Google's promise to cover data center lease payments if a tenant defaults jumped from $16.9 billion to $43.8 billion in just six months this year, according to the Financial Times. Less than 2% of it shows up as an actual liability on Alphabet's balance sheet. The rest sits in the footnotes, structured as guarantees rather than debt — even though Alphabet would still be on the hook for the full amount if something went wrong.
That mechanism sits at the centre of the FT report, which describes how Alphabet, Amazon, Meta, Microsoft and Oracle are using corporate guarantees and off-balance-sheet debt vehicles to fund the AI build-out while keeping the bulk of it off their books. The report puts the hidden exposure at roughly $300 billion. It comes on top of a separate Wall Street Journal analysis that found nine major tech and chip firms — those five plus Nvidia, Broadcom, SpaceX and AMD — carry close to $3 trillion in AI-related off-balance-sheet commitments.
Here is how the Google structure works. Fluidstack, an AI cloud provider, leases and builds data centers stocked with Google's TPU chips, then rents the capacity to Anthropic. Google guarantees Fluidstack's lease payments, which lets lenders finance the build-out at lower rates because Alphabet is effectively standing behind it. Alphabet books the guarantee as a credit derivative and records only its estimate of probable losses, which came to $815 million against a maximum exposure of $43.8 billion. Another $24.1 billion in guarantees is already committed and awaiting final terms.
A guarantee is a contingent liability, not a loan, so accounting rules let companies report only their modelled probability of loss rather than the full amount they have promised to cover. That keeps reported debt low and protects investment-grade ratings. It also keeps equity investors — already nervous about how much cash the AI race is burning — from getting spooked.
Meta's Hyperion data center in Louisiana shows the same playbook from a different angle. Meta struck a $30 billion deal to build the facility through a special purpose vehicle called Beignet Investor, majority-owned by Blue Owl Capital, which put up equity alongside roughly $27 billion in loans from Pimco, BlackRock and Apollo. Meta holds a 20% stake and leases the capacity back. Because Blue Owl's fund controls the entity, the $27 billion in debt never touches Meta's own balance sheet. Meta also split its leases into four-year segments specifically so rating agencies would not classify them as long-term debt.
None of this is illegal. Guarantees and SPVs, dressed up in residual-value structures, are permitted under current GAAP, and companies disclose the lot in footnotes and commitment tables that any analyst can read. But disclosure buried in a footnote is not the same as a number on the balance sheet that shows up in a debt-to-equity ratio or a credit model. That gap is exactly what is fuelling the AI-bubble debate among investors right now.
The real test comes if AI demand growth slows. Guarantees only stay cheap on paper as long as the underlying tenant — Anthropic, OpenAI, whoever — keeps paying its lease and the data center keeps generating revenue that justifies its cost. That is the whole bet.
If usage disappoints, or a tenant defaults, the parent company no longer gets to treat the obligation as contingent. Alphabet's $815 million provision would need to grow fast. The same goes for Meta's exposure through Beignet Investor, and every other SPV now scattered across Big Tech.
Regulators have taken notice. A letter sent to Treasury Secretary Scott Bessent earlier this year flagged concerns about AI-related debt structures, and the scale of the commitments — now measured in the trillions across the sector — makes the question of GAAP adequacy harder to wave away. For now, the balance sheets look fine. The footnotes tell a different story.
语言学习价值 · Vocabulary
长难句拆解 · Sentence Anatomy
| 原句 | Guarantees and SPVs, dressed up in residual-value structures, are permitted under current GAAP, and companies disclose the lot in footnotes and commitment tables that any analyst can read. |
|---|---|
| 结构拆解 | ① 主语Guarantees and SPVs(并列名词作主语) ② 插入定语dressed up in residual-value structures(过去分词短语作非限制性后置定语,相当于 which are dressed up…,"被包装成剩余价值结构") ③ 谓语 1are permitted under current GAAP(被动语态 + 依据状语) ④ 并列连词and 连接第二个独立分句 ⑤ 谓语 2 + 宾语companies disclose the lot(the lot = 全部这些东西,英式口语) ⑥ 定语从句that any analyst can read 修饰 tables。全句用"合法但不够透明"的对照语气,为下一句的转折(footnote ≠ balance sheet)铺路。 |
| 参考译文 | 担保与特殊目的载体这类被包装成"剩余价值结构"的安排,在当前美国会计准则下都是允许的,公司也会把这些内容全部披露在附注和承诺表里,任何分析师都能读到。 |
全文翻译 · Full Translation
以下按原文段落逐段对译,便于中英对照精读。
科技巨头正借助企业担保与表外载体,把大约 3000 亿美元的 AI 基础设施敞口挡在财务报表之外——而这只是它们将近 3 万亿美元、投资者几乎看不见的承诺中的一部分。
据《金融时报》报道,谷歌承诺在租户违约时替其支付数据中心租赁款,这一承诺规模仅今年半年内就从 169 亿美元跳升至 438 亿美元。其中不到 2% 会作为实际负债出现在 Alphabet 的资产负债表上。剩下的都躺在附注里,以"担保"而非"债务"的形式存在——尽管一旦出事,Alphabet 依然要为全额买单。
这一机制正是《金融时报》报道的核心:它描述了 Alphabet、亚马逊、Meta、微软和甲骨文如何用企业担保与表外债务载体为 AI 建设融资,同时把绝大部分留在账外。报道估算这部分隐性敞口约为 3000 亿美元。这还不算《华尔街日报》另一项分析的结果——包括上述五家以及英伟达、博通、SpaceX、AMD 在内的九家科技与芯片公司,合计持有接近 3 万亿美元的 AI 相关表外承诺。
谷歌这套结构是这样运作的:AI 云服务商 Fluidstack 租地并建设装满谷歌 TPU 芯片的数据中心,再把算力租给 Anthropic。谷歌为 Fluidstack 的租赁付款提供担保,由于 Alphabet 实际上站在它身后,贷款方愿意以更低利率为这个项目提供资金。Alphabet 把这项担保记为信用衍生品,只确认其估计的可能损失——在最大敞口 438 亿美元之下,这一计提为 8.15 亿美元。另有 241 亿美元担保已作出承诺、等待最终条款。
担保是"或有负债"而不是贷款,因此会计准则允许公司只披露其模型测算的损失概率,而不必列出承诺覆盖的全额。这让报告出来的债务保持低位,也保住了投资级评级;同时,还避免了本就担心 AI 竞赛烧钱速度的股票投资者受到惊吓。
Meta 位于路易斯安那州的 Hyperion 数据中心从另一个角度展示了同一套手法。Meta 谈成一笔 300 亿美元的交易,通过名为 Beignet Investor 的特殊目的载体建设该设施;该载体由 Blue Owl Capital 多数持股,Blue Owl 出资时还搭配了 Pimco、贝莱德和阿波罗提供的约 270 亿美元贷款。Meta 持有 20% 股权,并把算力租回来使用。由于 Blue Owl 的基金控制着这一主体,这 270 亿美元债务从不触及 Meta 自己的资产负债表。Meta 还把租约刻意切成四年一段,正是为了让评级机构不把它归类为长期债务。
这些做法都不违法。担保与特殊目的载体,只要包装成剩余价值结构,现行美国会计准则全部允许,公司也确实把这些内容完整披露在附注和承诺表里,任何分析师都能查到。但"埋在附注里的披露"与"出现在债务股本比或信用模型里的一个数字",完全是两回事。这道落差,正是当下投资者争论 AI 泡沫的核心。
真正的考验出现在 AI 需求增速放缓之时。担保之所以在纸面上保持便宜,前提是底层租户——Anthropic 也好、OpenAI 也好——持续付租,而数据中心持续产生足以证明其成本合理的收入。这就是全部的赌注。
一旦使用量令人失望,或者租户违约,母公司就再也无权把这笔义务当作"或有"处理。Alphabet 那 8.15 亿美元的计提将需要迅速扩大。Meta 通过 Beignet Investor 承担的敞口、以及其他散布在科技巨头之间的每一个特殊目的载体,都是同样的道理。
监管机构已经注意到了。今年早些时候寄给美国财政部长斯科特·贝森特的一封信,就对 AI 相关债务结构提出关切;而眼下整个行业的承诺规模已以万亿美元计,会计准则是否够用这个问题越来越难以搪塞过去。目前看,资产负债表一切正常。附注讲的却是另一个故事。