英伟达（Nvidia）历经三十年市值突破1万亿美元，仅用九个月便跃升至2万亿，随后在不到两年内跨越5万亿门槛，当前估值高达5.4万亿美元。随着明年销售额预计翻倍并在2029年有望冲击1万亿美元年营收，黄仁勋正通过激进的金融工程全方位拉动硬件需求。8月中旬，英伟达为俄亥俄州大型数据中心提供了高达1,050亿美元的托底担保；此前一周又联合六家华尔街巨头启动超过5,000亿美元的基础设施投资计划，并为部分项目提供高达25%的设备残值担保；过去三年间，英伟达向初创企业承诺投资超700亿美元，并向客户提供了3,000亿美元的各类财务支持，被业界戏称为“AI中央银行”。

这种深度介入金融担保的举措，源于其面临的核心客户向竞争对手转型的巨大压力。亚马逊、谷歌、Meta和微软等超大规模云服务商贡献了英伟达约50%的营业收入，今年这四家巨头的资本开支预计达8,000亿美元，但其自研定制芯片的成本仅为英伟达芯片的五分之一至三分之一。彭博行业研究预测，到本十年末，定制芯片在AI算力市场的份额将从目前的40%攀升至50%。面对单座1吉瓦设施需部署50万颗芯片、耗资470亿美元（其中芯片占40%）的庞大开销，以及自身将芯片涨价逾15%的举措，英伟达不得不全力扶持新兴云服务商（neoclouds）以开辟新增长极。

为缩小新兴算力公司与大型科技巨头之间的借贷成本差距，英伟达构建了复杂的担保与股权网络：其持有CoreWeave公司11%的股份（价值逾20亿美元），并承诺在2032年前认购至多63亿美元的闲置算力；同时通过类似机制协助Anthropic达成价值350亿美元的算力租赁协议。摩根士丹利预计，随着担保生效，英伟达的总负债将从明年初的530亿美元飙升至2029年初的2,000亿美元。尽管英伟达坐拥990亿美元现金储备且今年预计创造2,000亿美元现金流，短期内具备充足的风险缓冲垫，但若下游算力需求增速放缓，激增的表外担保可能引发严重的资产减值与连锁违约风险。

Nvidia's market valuation required three decades to reach trn, added a second trillion within nine months, and crossed trn less than two years later to achieve a current capitalization of .4trn. Anticipating doubled sales next year and potential trn annual revenues by 2029, chief executive Jensen Huang has deployed aggressive financial engineering to underwrite industry demand. In August, Nvidia pledged a bn backstop for a hyperscale Ohio data centre and unveiled a bn infrastructure alliance with Wall Street institutions offering up to 25% equipment residual-value coverage. Cumulatively, the firm has committed over bn to startups and extended bn in client credit facilities, functioning as the de facto central bank of AI.

This balance-sheet mobilization counters mounting defection risks among core hyperscalers. Amazon, Google, Meta, and Microsoft provide roughly 50% of Nvidia’s revenue and plan bn in aggregate capex this year; however, their proprietary custom chips cost between one-fifth and one-third of Nvidia processors. Bloomberg Intelligence projects custom silicon will capture 50% of the AI accelerator market by 2030, up from 40% currently. Given that a standard 1GW data facility requires 500,000 chips costing bn—with silicon comprising 40% of outlays—Nvidia is actively subsidizing independent neocloud operators to absorb inventory while simultaneously pushing through processor price increases exceeding 15% amid rising component expenses.

To bridge borrowing spreads between investment-grade hyperscalers and cash-burning neoclouds, Nvidia has constructed an elaborate web of reciprocal commitments. The company holds an 11% stake in CoreWeave worth over bn, guaranteeing up to .3bn in unused compute purchases through 2032, alongside facilitating Anthropic’s bn capacity arrangement via Lambda. Morgan Stanley forecasts Nvidia’s all-in obligations will surge from bn to bn by early 2029. Although fortified by bn in liquid reserves and bn in projected annual operating cash generation, an unexpected deceleration in artificial intelligence commercialization could convert massive off-balance-sheet guarantees into acute financial liabilities, leaving Nvidia exposed to severe asset deflation across global compute markets.

Source: Nvidia is the central bank of AI

Subtitle: But will its loans prove sound?

Dateline: Sep 3rd 2026