未来五年,美国数据中心容量将从 25 吉瓦增长至 70 吉瓦,这是全球范围内总耗资约 5 万亿美元的建设浪潮的一部分。
资金将从何而来?
数据中心是作为房地产项目开发的,包含部分股权,但大部分资金来自债务——通常占比 70% 或更高。假设我们能够按计划建成所有这些数据中心,信贷市场是否有足够的债务资金可供融资?
为了理解这一规模,我将 4 万亿美元的新增 AI 债务与全球主要信贷市场的规模进行了比较。这一 AI 建设浪潮相当于美国公司债券市场扩张 34%。
在这一规模下,数据中心债务将使未偿商业票据市场扩大两倍,超过全球私人信贷市场的规模,并相当于美国市政债券市场的 91%。
几十年来,规模达 4.4 万亿美元的市政债券市场一直为美国道路、桥梁、供水系统和机场等实体基础设施建设提供资金。这也引发了一个问题:寻求经济增长的地方政府是否会像为发电厂融资那样,利用市政债券为部分数据中心提供资金。
所有这些债务都需要靠利润来偿还:到 2030 年,AI 在软件、模型 token 和企业自动化方面的年收入必须超过 1.2 万亿至 1.5 万亿美元。
目前,所有云服务提供商和模型实验室的 AI 数据中心年化收入估计在 1000 亿至 2000 亿美元之间。
从目前约 1500 亿美元增长到 1.35 万亿美元,需要在未来五年内实现 55% 的复合年增长率(CAGR)。相比之下,超大规模云厂商目前的年增长率在 37% 到 82% 之间(AWS 为 37%,Azure 为 43%,Google Cloud 为 82%),但增长正在加速。8
作为参照,目前全球企业软件市场总量约为 1.4 万亿美元,而 2030 年全球 IT 支出预计约为 9 万亿美元。9
为 AI 基础设施热潮提供融资,已不再是风险投资或企业盈利的故事。这是一场宏观经济层面的信贷事件,其规模将可与金融史上最大规模的债务扩张相媲美。
-
摩根大通资产管理、Western Asset 和 PIMCO 的研究对数据中心容量扩张及到 2030 年总计 5 万亿美元的资本支出进行了估算。↩︎
-
哥伦比亚商学院房地产学教授 Stijn Van Nieuwerburgh 及 CREFC 对数据中心项目融资的分析发现,项目层面的杠杆率通常带有 65% 至 75% 的债务(在合成合资 SPV 中,如 Meta 的 Beignet 工具,这一比例可高达 90%),而传统企业杠杆率则为 40%。↩︎
-
作为一名风险投资人,我对债券市场抱有一种天真的看法。
-
商业票据是短期企业债务,通常在 270 天内到期,公司用它来支付工资和日常运营开支。相比之下,公司债券是期限为数年或更长的长期债务,用于为资本项目融资。全球私募信贷管理资产涵盖直接贷款、夹层融资和困境信贷策略。↩︎
-
市政债券是州与地方政府为公路、桥梁、供水系统和机场等公共基础设施融资而发行的债务。
-
按当前市场利率 6.5% 至 7.5% 偿还 4 万亿美元债务,仅年度利息支出就需要 2600 亿至 3000 亿美元。按投资级 3 倍利息覆盖率计算,基础设施需要约 8000 亿至 9000 亿美元的年度营业利润才能满足贷款方要求。假设云与 AI 的毛利率为 60% 至 70%,这意味着 AI 年收入需达到 1.2 万亿至 1.5 万亿美元。
-
根据超大规模云厂商截至 2026 年年中的披露:Microsoft 报告的 AI 收入年化运行率超过 130 亿美元,AWS 报告的 AI 与定制芯片年化运行率超过 500 亿美元,同时 Google Cloud、Oracle Cloud 及领先的基础模型实验室的 AI 基础设施收入也在快速扩张。
-
https://tomtunguz.com/aws-answers-the-cloud-race/ 报告了当前云厂商增长率:AWS 为 37%,Azure 为 43%,Google Cloud 为 82%。
-
Gartner 全球 IT 支出预测显示,企业软件支出预计在 2026 年达到 1.4 万亿美元,全球 IT 总支出则有望在 2030 年复合增长至 9 万亿美元。
Over the next five years, US data center capacity will grow from 25 gigawatts to 70 gigawatts, part of a global buildout costing roughly $5t.1
Where will the money come from?
Data centers are built as real estate projects with some equity, but the majority debt : typically 70% or more2. Assuming we achieve our plans to build all these data centers, is there enough debt available in the credit markets to finance it?3
To understand the magnitude, I compared the $4t of new AI debt to the sizes of the world’s primary credit markets. The AI buildout represents a 34% expansion of the US corporate bond market.
At this scale, data center debt triples the outstanding commercial paper market, grows larger than the global private credit market, & equals 91% of the US municipal bond market.4
For decades, the $4.4t municipal bond market has financed the physical buildout of American roads, bridges, water systems, & airports. It also raises the question of whether municipalities seeking economic growth will use municipal bonds to fund some of these data centers, much like power plants.5
All of this debt needs to be serviced from profits : annual AI revenue must exceed $1.2t to $1.5t by 2030 across software, tokens, & enterprise automation.6
Today, annualized AI data center revenue across all cloud providers & model labs is estimated at $100b to $200b.7
Reaching $1.35t from roughly $150b today requires a 55% compound annual growth rate (CAGR) over the next five years. By comparison, hyperscalers currently grow between 37% & 82% annually (AWS at 37%, Azure at 43%, & Google Cloud at 82%) ; but the growth is accelerating.8
For perspective, the global enterprise software market totals roughly $1.4t today, out of an estimated $9t in worldwide IT spending in 2030.9
Financing the AI infrastructure boom is no longer a venture capital or corporate earnings story. It is a macroeconomic credit event that will rival the largest debt expansions in financial history.
-
J.P. Morgan Asset Management, Western Asset, & PIMCO research estimates on data center capacity expansion & $5t in total capital expenditure through 2030. ↩︎
-
Columbia Business School real estate professor Stijn Van Nieuwerburgh & CREFC analysis on data center project finance find facility-level leverage routinely carries 65% to 75% debt (& up to 90% in synthetic joint venture SPVs like Meta’s Beignet vehicle), compared to traditional 40% corporate leverage. ↩︎
-
As a venture capitalist, I have a naive view of the bond market. ↩︎
-
Commercial paper is short-term corporate debt, typically maturing in under 270 days, that companies use to fund payroll & day-to-day operations. Corporate bonds, by contrast, are long-term debt with maturities of several years or more, used to finance capital projects. Global private credit assets under management across direct lending, mezzanine, & distressed credit strategies. ↩︎
-
Municipal bonds are debt issued by state & local governments to finance public infrastructure like roads, bridges, water systems, & airports. ↩︎
-
Servicing $4t in debt at prevailing market rates between 6.5% & 7.5% requires $260b to $300b in annual interest expense alone. At an investment-grade interest coverage ratio of 3x, the infrastructure requires roughly $800b to $900b in annual operating profit to satisfy lenders. Assuming cloud & AI gross margins of 60% to 70%, that implies $1.2t to $1.5t in annual AI revenue. ↩︎
-
Based on hyperscaler disclosures through mid-2026: Microsoft reported an AI revenue run rate surpassing $13b, AWS reported an AI & custom silicon run rate exceeding $50b, alongside rapidly scaling AI infrastructure revenue across Google Cloud, Oracle Cloud, & leading foundation model labs. ↩︎
-
https://tomtunguz.com/aws-answers-the-cloud-race/ reports current cloud growth rates: AWS at 37%, Azure at 43%, & Google Cloud at 82%. ↩︎
-
Gartner Worldwide IT Spending Forecast projects enterprise software spending reaching $1.4t in 2026, with overall worldwide IT spending compounding toward $9t in 2030. ↩︎