但它的贷款能否被证明是稳健的？
经
过了三十年
Nvidia 这家美国公司的芯片为全球大部分人工智能提供算力，它用了三十年才达到 1 万亿美元估值。而达到 2 万亿美元只又用了九个月。此后不到两年，它又突破了 5 万亿美元大关。它如今是全球市值最高的公司，价值约 5.4 万亿美元。明年其销售额预计几乎翻倍。一些分析师预测，到 2029 年其年营收将达到 1 万亿美元。
然而，推动 Nvidia 惊人增长的并不只是芯片制造。其老板黄仁勋（Jensen Huang）还借助金融工程来提振对其产品的需求。例如，8 月中旬，Nvidia 同意为俄亥俄州一座将大量使用其芯片的大型数据中心提供高达 1050 亿美元的兜底支持。一周前，它已披露计划，在六家华尔街大型机构的帮助下，撬动超过 5000 亿美元的投资用于
AI
基础设施，方式是为其出售给此类项目的设备价值提供担保。它可能承销某些投资成本的四分之一之多。7 月，Nvidia 还以另一种方式帮助一些客户为大型数据中心融资，即承诺如果未达到既定目标，就补足它们的收入。
企业常常会向自己的客户提供贷款：想想汽车制造商的金融部门就知道了。黄仁勋认为，英伟达只是在帮助那些完全可行的项目释放投资，这些项目否则可能难以以合适的价格借到足够的资金。但在批评者看来，这些交易带有浓厚的 1990 年代末互联网泡沫的气息——当时思科和朗讯等电信设备制造商向购买其设备的电信运营商借出了数十亿美元。当市场对电信公司服务的需求不及预期时，其中一些公司倒闭了，让思科和朗讯蒙受了巨额损失。
正如分析公司 Seaport Research Partners 的 Jay Goldberg 所说，英伟达正在“激发需求”与“创造需求”之间走钢丝。他目前还不认为英伟达已经越过了那条线，尽管它“已经相当接近了”。还有人持更加怀疑的态度，其中包括投资者 Michael Burry，他曾靠做空引发 2007-09 年金融危机的抵押贷款支持证券而大赚一笔。他和其他人正在追问，如果对
AI
芯片的需求增长慢于预期，供应增加、价格下跌，那会发生什么——这不仅会削减英伟达自身的利润，还可能因其对客户的慷慨支持而产生亏损。
庞大度假村的贷款人
这个问题之所以切中要害，是因为英伟达的财务承诺规模巨大且增长迅速。过去三年里，它承诺向初创企业投资超过 $70bn，并向其客户提供 $300bn 的财务支持。一些人已经开始把英伟达称为
AI
”，因为它在为整个行业提供融资方面扮演着如此关键的角色。正如黄先生所言，这确实将有助于行业增长——但它也伴随着风险。
英伟达的金融工程在一定程度上是对其最大客户转变为竞争对手的回应。“超大规模云厂商”，即亚马逊、谷歌、Meta 和微软等科技巨头，约占英伟达营收的一半。今年它们预计将投资约 8000 亿美元，主要用于
AI
基础设施。但它们中的大多数已开始设计自己的芯片，这使它们未来从英伟达的采购充满变数。
对超大规模云厂商而言，这些定制芯片要便宜得多，成本仅为英伟达芯片的五分之一到三分之一。它们在某些任务上表现也更出色，因为超大规模云厂商可以针对自己的软件对其进行定制。更糟糕的是，从英伟达的角度看，少数科技巨头不仅为自己使用而开发芯片，还在向他人推销。例如，谷歌已向 Anthropic 出售了一些专用处理器，后者是一家大型独立
AI
实验室。亚马逊也预计其定制芯片业务将成为可观的收入来源。数据提供商 Bloomberg Intelligence 预测，定制芯片将逐步蚕食英伟达的销售额，在用于
AI
到本十年末将升至这一水平，而今年约为 40%。
超大规模云厂商拥有投资级信用评级，这使其借贷成本保持在低位。新兴的 neocloud 有着相似的开支需求，但收入微薄（见商业版块）。它们的贷款自然要昂贵得多。Google 的母公司 Alphabet 在 11 月发行了 27.5 亿美元的 50 年期债券，年利率为 5.7%。而最大的 neocloud 公司 CoreWeave 在 7 月借入 26 亿美元时的利率几乎是这一水平的两倍。
Nvidia 这家“银行”想要缩小的，正是超大规模云厂商与其他所有公司之间在借贷成本上的这一差距。它采取的方式之一是入股那些本身将成为客户、或会间接帮助推高 Nvidia 芯片需求的初创公司。去年 Nvidia 进行了约 90 笔此类投资，几乎是两年前的两倍。今年它已经又达成了 60 多笔。
定性宽松
其中一些支票旨在推广开放权重
AI
模型，用户可以免费下载并加以改造，这与 Anthropic、Open
AI
和 Google 等公司的专有产品不同——用户通常通过订阅方式使用这些产品，且其内部运作机制是隐藏的。8 月，Nvidia 同意向 Poolside 支付款项，这是一家正在构建
AI
编码模型，斥资 60 亿美元授权其软件，另加 10 亿美元换取股权。它还同意以 129 亿美元收购 Hugging Face——一个托管开放权重模型的平台。此类投资背后的意图，是通过催生出大量
AI
产品和公司，来拉动对 Nvidia 芯片的需求，而这些产品和公司独立于超大规模云厂商。
然而，Nvidia 越来越不只是投资有前景的
AI
公司，还更直接地帮助它们为大型项目融资。7 月初，它宣布了一项新策略：承诺将 neocloud 从新建数据中心获得的收入补足至约定的下限。这些承诺的具体条款因交易而异，往往持续六年。在此期间，Nvidia 承诺按固定价格购买“算力”——用行业术语来说就是如此。
如果该 neocloud 能以更高价格卖出相关容量，Nvidia 将分得差价的一部分。这种安全网让 neocloud 的未来收入变得可预测得多，从而降低了它们为新建数据中心而承担的债务成本。这反过来又刺激了对 Nvidia 处理器的需求。
Sharon
AI
，一家澳大利亚 neocloud，预计将利用这类价值约 49 亿美元的兜底承诺，部署约 40,000 块 Nvidia 芯片。另一家 neocloud 公司 Firmus 计划购买多达 170,000 块。咨询公司 SemiAnalysis 的分析师 Dan Nishball 认为，此类安排不必成为该行业的永久特征，而只是为贷款机构“争取时间”，让它们对向此类项目放贷更加放心。
以 CoreWeave 为例，这是一家计划出租 AI 算力的 neocloud 公司。
Nvidia 已向 CoreWeave 投资超过 20 亿美元，并持有该公司约 11% 的股份。
CoreWeave 从 Nvidia 购买 AI 芯片来建设数据中心。
Nvidia 同意在 2032 年之前从 CoreWeave 购买最多 63 亿美元的未使用数据中心容量。
这是 Nvidia 达成的众多采用复杂财务安排的交易之一。
它已在财务上与自己的许多客户纠缠在一起：neocloud 提供商、AI 实验室，甚至投资公司。
- Neocloud
- AI 实验室
- 金融公司
Nvidia 在其中许多公司持有股权。
- 股权
它还为其许多客户提供了财务兜底。理论上，这些承诺可能使其面临近 3000 亿美元的成本。
- 股权
- 担保与采购承诺
Nvidia 还在对大型
AI
实验室采取一种创造性的方式，这些实验室需要大量算力，却在大量烧钱。它在俄亥俄州支持的那个巨型数据中心归
SB
Energy 所有，后者是日本企业集团 SoftBank 旗下的一个部门。Open
A
I
将成为租户。就 Nvidia 而言，它将为该土地和建筑的租约以及其所依赖的购电合同提供担保。作为回报，该项目将使用 150 万个 Nvidia 处理器，并预计在项目生命周期内经历多轮“升级周期”。
即便是从 Amazon 和 Google 采购芯片的 Anthropic，也未能逃脱这张网。据报道，通过一系列错综复杂的协议，它签署了一份价值 350 亿美元的协议，从 Lambda 租用云计算容量，而 Nvidia 持有这家 neocloud 的股份。Lambda 反过来又使用由另一家 neocloud Hut 8 正在开发的设施，而该设施的全部容量此前已被 Nvidia 租下（如今可能已大部分转租出去）。
Nvidia 正在探索各种创造性的方式，为这类项目引入更多外部资本。它与华尔街大型公司达成的 5000 亿美元合作，将吸引主权财富基金、保险公司和养老基金等机构投资者。这些投资者将为独立载体提供融资，由这些载体购买 Nvidia 硬件、建设基础设施并出售算力。
Nvidia 既不会为这些交易预先投入任何现金，也不会承担任何债务。不过在部分情况下，它会提供最高达交易总价四分之一的“残值支持”，以弥补相关硬件在一定期限后出现的价值缺口。它还将提供技术支持，让贷款方更有信心，相信他们所融资的设备会得到妥善使用。
黄仁勋将此描述为让算力成为“一种可投资资产类别”的方式。
支撑这张义务之网的，是两个根本性假设：Nvidia 的芯片将保值，以及对算力的需求将持续快速增长。这两点都没有保证。
先从芯片说起。黄仁勋表示，它们是“可替代的”且耐用，软件升级会在售出多年后仍提升其可用性。他主张，它们应被视为可作银行抵押的资产，可以作为贷款和担保的抵押品。但一种
AI
芯片是否保值仍有争议。Burry 先生认为，大型云服务商通过将折旧年限设为五到六年而非两到三年，从而虚增了利润。
到目前为止，黄仁勋在这场争论中占据上风。旧款芯片仍有需求。8 月，CoreWeave 宣布了一份涉及 Nvidia
A
100 芯片的合同，期限延续至 2029 年。
A
100 于 2020 年发布，此后 Nvidia 又推出了多款更新的型号。也有证据表明，旧款芯片确实保值。SemiAnalysis 估计，租用 Nvidia 的
H
100——这款
AI
行业的主力芯片——按一年期合同租用，每小时约 2.80 美元。这仅比该芯片 2023 年初发布时便宜了约十分之一。
然而，这类芯片的寿命和保值性可能仅仅是稀缺性的产物。当算力受限时，企业别无选择，只能让旧款芯片继续运转。许多企业将这类芯片用于推理，也就是说，当
AI
模型响应查询，这是一项远比训练模型要求更低的任务，而训练模型才需要使用最新的硬件。
Nvidia 自己每年都会发布亮眼的新芯片，明确意图就是取代前代产品。这使它陷入一种尴尬的处境：既要主张旧芯片应保值，又要主张新芯片应取而代之。找到平衡点将十分棘手。
任性的指引
随着供给增加，算力价格理应下降。超大规模云厂商乃至像 Open
AI
这样的实验室都在推出专为推理打造的芯片，因为推理支出已超过训练支出。Nvidia 享有令人艳羡的约 75% 毛利率，而规模较小的竞争对手
AMD
约为 55%。Tim Davis 是一家
AI
硬件公司的创始人，该公司已被芯片设计商 Qualcomm 收购。他认为，随着选择日益增多，芯片制造的经济性将“开始被压缩”。
芯片价格面临的最大风险是需求本身。只要在
AI
持续繁荣，Nvidia 卖出它的芯片，它的客户填满自己的数据中心，而它的各种担保很少被触发。但如果需求并不像市场预期的那样广阔，neocloud 及其他
AI
供应商可能难以出售其产能或实现盈利。这反过来可能触发 Nvidia 的兜底承诺，迫使它购买未使用的算力、弥补价格缺口或为不需要的电力付费。同样的放缓也会削弱 Nvidia 的销售额，从而减少可用于履行这些义务的现金流。这种情景并不取决于需求崩塌；仅仅是增长令人失望，就可能意味着麻烦。
Nvidia 可能不得不掏出多少钱？它已承诺约 $25bn 的未来股权投资。它背负约 $33bn 的债务。它对客户的潜在负债约为 $300bn，但只在经济下行时才会生效，因此不出现在其资产负债表上。这些包括为 Open
AI
的数据中心提供的 $105bn 担保；通过华尔街合作安排的至高 $125bn；以及约 $67bn 的其他兜底承诺。
Nvidia 几乎肯定不会最终支付接近其担保全额的资金。它的承诺分散在多年之间，因此不可能一次性全部到期。它为 Open
AI
的数据中心，举例来说，从 2028 年开始运行 20 年。如果 Open
AI
停止支付租金，Nvidia 可以找到另一个租户，从而大幅降低其风险敞口。
更重要的是，Nvidia 自身的财务状况足够稳健，足以承受这些负债。投资银行 Morgan Stanley 估计，随着担保生效，Nvidia 的“全部计入”债务将从明年初的 $53bn 升至 2029 年初的 $200bn。
但这被目前价值 $99bn 的现金和流动性证券储备，以及今年将产生约 $200bn 现金的业务所抵消。只有在灾难性衰退导致 Nvidia 所有担保同时到期、利润几乎完全蒸发的情况下，公司才会陷入危险——就目前情况而言。
然而，如果 Nvidia 的承诺持续增长，情况可能会发生变化。金融研究公司 CreditSights 的 Andy Li 担心，它会一直“踩油门”，直到“有什么东西崩掉”。SemiAnalysis 估计，Nvidia 在其 neocloud 后备计划覆盖的每 100 兆瓦数据中心容量上，承担约 $5.9bn 的担保。
SemiAnalysis 认为，如果 Nvidia 增加更多此类担保，仅这一项举措带来的风险敞口到 2028 年底就可能达到 $175bn。
Nvidia 并非唯一一家使用越来越复杂的金融工具来润滑客户需求的公司。
AMD
已提出向 Open 出售自身大量股份
AI
以及 Meta，以换取购买其芯片的巨额合同。今年 4 月，Google 组建了一个企业财团，帮助为 Anthropic 购买其价值 350 亿美元的芯片提供融资。代表 Google 开发这些芯片的 Broadcom 同意，如果 Anthropic 未能付款，将承担任何缺口。
在锁定销售和贷款的竞赛中，存在一种危险：那些原本可能难以找到融资的投机性项目将会被建成。如果该行业的回报不及预期，或者比这些错综复杂的交易所假设的兑现时间更长，结果将是大量昂贵而闲置的芯片。没有哪家公司比 Nvidia 更深地卷入这张网中。黄仁勋认为，
AI
基础设施的建设正在“全速推进”。但 Nvidia 的投资者应当意识到，他们为这轮繁荣提供的融资越多，他们将来不得不承受的任何崩盘也就越多。
■
But will its loans prove sound?
I
t took thirty
years for Nvidia, an American firm whose chips power much of the world’s artificial intelligence, to reach a valuation of $1trn. Getting to $2trn took only nine more months. It passed the $5trn mark less than two years after that. It is now the world’s most valuable company, worth around $5.4trn. Next year its sales are expected almost to double. Some analysts predict it will take in $1trn in annual revenue by 2029.
It is not just chipmaking that has spurred Nvidia’s stunning growth, however. Its boss, Jensen Huang, has also resorted to financial engineering to boost demand for its wares. In mid-August, for example, Nvidia agreed to provide a backstop worth up to $105bn for a vast data centre in Ohio that will use lots of its chips. A week earlier it had revealed plans to mobilise more than $500bn of investment in
AI
infrastructure with the help of six big Wall Street firms, by guaranteeing the value of the equipment it sells to such projects. It may underwrite as much as a quarter of the cost of some investments. In July Nvidia helped some customers finance big data centres in another way, by promising to top up their income if it misses set targets.
Companies often lend to their customers: think of the financing arms of carmakers, for example. Mr Huang argues that Nvidia is simply helping to unlock investment for perfectly viable projects that might otherwise struggle to borrow enough at the right price. But to critics, these deals carry more than a whiff of the dotcom boom of the late 1990s, when telecom-equipment-makers such as Cisco and Lucent lent billions of dollars to telecoms providers that bought their gear. When demand for the telecom firms’ services fell short, some of them collapsed, landing Cisco and Lucent with big losses.
As Jay Goldberg of Seaport Research Partners, a firm of analysts, puts it, Nvidia is walking a fine line between “enabling demand” and “creating it”. He does not yet think Nvidia has crossed that line, although it is “getting pretty close”. Others are more sceptical still, including Michael Burry, an investor who made a fortune betting against the mortgage-backed securities that precipitated the financial crisis of 2007-09. He and others are asking what might happen if demand for
AI
chips grows more slowly than expected, supply increases and prices fall, which would not only reduce Nvidia’s own profits but potentially also generate losses on its lavish support for its customers.
Lender of vast resort
The question is pertinent because Nvidia’s financial commitments are huge and growing fast. Over the past three years it has pledged over $70bn in investment in startups and offered $300bn in financial support to its customers. Some have taken to calling Nvidia the “central bank of
AI
”, because it plays so pivotal a role in financing the industry. That will indeed, as Mr Huang argues, help the industry grow—but it also carries risks.
Nvidia’s financial engineering is partly a response to its biggest customers’ transformation into rivals. “Hyperscalers”, tech giants such as Amazon, Google, Meta and Microsoft, account for roughly half of Nvidia’s revenue. This year they are projected to invest around $800bn, largely on
AI
infrastructure. But most of them have begun designing their own chips, which puts their future purchases from Nvidia in doubt.
For the hyperscalers, these custom chips are much cheaper, costing between a fifth and a third as much as Nvidia’s. They can also perform some tasks better, because the hyperscalers can tailor them to their own software. Worse, from Nvidia’s point of view, a few of the tech giants are not simply developing chips for their own use, but also marketing them to others. Google, for example, has sold some specialised processors to Anthropic, a big, free-standing
AI
lab. Amazon also expects its custom-chip business to become a sizeable source of revenue. Bloomberg Intelligence, a data provider, predicts that custom chips will gradually eat into Nvidia’s sales, accounting for about 50% of the market for processors used in
AI
by the end of the decade, up from roughly 40% this year.
Hyperscalers have investment-grade credit ratings, which keep their borrowing costs low. Upstart neoclouds have similar spending needs, but little revenue (see Business section). Their loans are naturally much more expensive. Alphabet, Google’s parent company, sold $2.75bn of 50-year bonds in November, at an annual interest rate of 5.7%. The rate at which CoreWeave, the biggest neocloud, borrowed $2.6bn in July was almost double.
It is this gap, between hyperscalers’ borrowing costs and everyone else’s, that the bank of Nvidia would like to narrow. One way it does that is by taking equity stakes in startups that will be customers themselves or that will help fuel demand for Nvidia’s chips indirectly. Last year Nvidia made about 90 such investments, nearly twice as many as two years earlier. This year it has already agreed another 60-odd.
Qualitative easing
Some of these cheques aim to propagate open-weight
AI
models, which users can download free of charge and adapt, unlike proprietary offerings from firms like Anthropic, Open
AI
and Google, which users tend to access via subscriptions and whose inner workings are hidden. In August Nvidia agreed to pay Poolside, a startup building
AI
coding models, $6bn to license its software and a further $1bn for a stake. It has also agreed to buy Hugging Face, a platform hosting open-weight models, for $12.9bn. The intention behind such investments is to fuel demand for Nvidia’s chips by creating a proliferation of
AI
products and companies that are independent of the hyperscalers.
Increasingly, however, Nvidia is not simply investing in promising
AI
firms, but also helping them finance big projects more directly. In early July it announced a new stratagem in which it promises to top up neoclouds’ income from new data centres to an agreed floor. These undertakings, the exact terms of which vary from deal to deal, often last for six years. Throughout that period, Nvidia promises to pay a set price for “compute”, as the jargon has it. If the neocloud manages to sell the capacity in question at a higher price, Nvidia receives a share of the difference. This safety-net makes neoclouds’ future revenues much more predictable and so lowers the cost of the debt they take on to build new data centres. That, in turn, spurs demand for Nvidia’s processors.
Sharon
AI
, an Australian neocloud, expects to make use of a backstop of this sort worth about $4.9bn to deploy around 40,000 Nvidia chips. Firmus, another neocloud, plans to buy as many as 170,000. Dan Nishball, an analyst at SemiAnalysis, a consultancy, argues such arrangements need not be a permanent feature of the industry, but instead “buy time” for lenders to become more comfortable about lending to such projects.
Take CoreWeave, a neocloud firm which plans to rent out AI computing capacity.
Nvidia has invested over $2bn in CoreWeave and owns around 11% of the company.
CoreWeave buys AI chips from Nvidia to build data centres.
Nvidia agrees to buy up to $6.3bn of unused data-centre capacity from CoreWeave until 2032.
This is one of many deals Nvidia has done that use complex financial arrangements.
It has become financially entangled with many of its own customers: neocloud providers, AI labs and even investment firms.
- Neocloud
- AI lab
- Financial firm
Nvidia holds equity stakes in many of these companies.
- Equity stake
And it has also provided financial backstops to many of its customers. In theory, these commitments could expose it to costs of nearly $300bn.
- Equity stake
- Guarantees & purchase commitments
Nvidia is also taking a creative approach with big
AI
labs, which need huge amounts of compute but are bleeding cash. The giant data centre it is backing in Ohio is owned by
SB
Energy, a unit of SoftBank, a Japanese conglomerate. Open
A
I
will be the tenant. Nvidia, for its part, will provide guarantees for the lease on the land and buildings and for the power-purchase contract on which it will rely. In return, the project will use 1.5m Nvidia processors, with multiple “upgrade cycles” expected over the life of the project.
Even Anthropic, which buys chips from Amazon and Google, has not escaped this web. Through a convoluted series of agreements, it has reportedly signed a $35bn deal to rent cloud-computing capacity from Lambda, a neocloud in which Nvidia has a stake. Lambda, in turn, uses a facility being developed by Hut 8, another neocloud, whose entire capacity Nvidia had leased (and may now have largely sublet).
Nvidia is exploring creative ways to draw more outside capital into such projects. Its $500bn partnership with big Wall Street firms will court institutional investors such as sovereign-wealth funds, insurers and pension funds. These investors will finance independent vehicles that will buy Nvidia hardware, build infrastructure and sell compute. Nvidia will neither contribute any cash upfront for these deals nor take on any debt. In some cases, though, it will provide “residual-value support” of up to a quarter of the deal’s total price to cover shortfalls in the value of the hardware involved after a certain period. It will also provide technical support, giving lenders more confidence that the equipment they finance will be used well. Mr Huang frames this as a way to make compute “an investable asset class”.
Underpinning this web of obligations are two fundamental assumptions: that Nvidia’s chips will retain their value and that demand for compute will continue to grow at a rapid pace. Neither is assured.
Start with the chips. Mr Huang says they are “fungible” and durable, with software upgrades enhancing their usefulness for years after they are sold. They should be seen as bankable assets, he argues, that can serve as collateral for loans and guarantees. But how long an
AI
chip retains its value is a matter of debate. Mr Burry has argued that big cloud providers have inflated profits by depreciating them over five or six years rather than two or three.
So far Mr Huang has the better of the argument. Older chips are still in demand. In August CoreWeave announced a contract involving Nvidia’s
A
100 chips that runs into 2029. The
A
100 was launched in 2020 and Nvidia has since rolled out multiple newer models. There is evidence, too, that old chips retain their value. SemiAnalysis estimates that renting Nvidia’s
H
100, a workhorse of the
AI
industry, costs around $2.80 an hour on a one-year contract. That is only about a tenth less than when the chip launched in early 2023.
Yet both the longevity and value of such chips may simply be artefacts of scarcity. When compute is constrained, firms have no choice but to keep older chips humming. Many are using such chips for inference, meaning when an
AI
model responds to queries, a far less demanding task than training models, for which the newest hardware is used.
Nvidia itself releases spiffy new chips every year, with the explicit intention of superseding their predecessors. That leaves it in the odd position of arguing both that old chips should retain their value and that its new chips should supplant them. Finding a balance will be tricky.
Wayward guidance
The price of compute should fall as supply increases. Hyperscalers and even labs like Open
AI
are rolling out purpose-built inference chips, since spending on inference has overtaken spending on training. Nvidia enjoys enviable gross margins, of around 75%, against roughly 55% at
AMD
, a smaller rival. Tim Davis, the founder of an
AI
hardware firm acquired by Qualcomm, a chip designer, believes that, as options proliferate, the economics of chipmaking will “start to compress”.
The biggest risk to chip prices is demand itself. As long as spending on
AI
keeps booming, Nvidia sells its chips, its customers fill their data centres and its various guarantees are seldom invoked. But if demand is not quite as expansive as markets expect, neoclouds and other
AI
providers could struggle to sell their capacity or turn a profit. That, in turn, could trigger Nvidia’s backstops, forcing it to buy unused compute, cover price shortfalls or pay for unwanted power. The same slowdown would also dent Nvidia’s sales and thus the cashflow available to meet those obligations. This scenario does not hinge on a collapse in demand; merely disappointing growth could conceivably spell trouble.
How much could Nvidia have to cough up? It has promised around $25bn in future equity investments. It owes around $33bn in debt. Its potential liabilities to customers amount to about $300bn, but only come into play in a downturn and so do not appear on its balance-sheet. These include the $105bn guarantee behind Open
AI
’s data centre; as much as $125bn through the Wall Street partnership; and around $67bn in other backstops.
Nvidia will almost certainly not end up paying anything close to the full amount of its guarantees. Its commitments are spread over many years so could not come due all at once. Its proposed backstop for Open
AI
’s data centre, for example, runs for 20 years, starting in 2028. If Open
AI
stopped paying the lease, Nvidia could find another tenant, vastly diminishing its exposure.
What is more, Nvidia’s own finances are strong enough to weather these liabilities. Morgan Stanley, an investment bank, reckons Nvidia’s “all-in” debt will rise from $53bn early next year to $200bn by the beginning of 2029 as guarantees come into effect. But that is offset by a stash of cash and liquid securities currently worth $99bn, and a business that will generate about $200bn in cash this year. Only a cataclysmic downturn that caused all Nvidia’s guarantees to come due and its profits to evaporate almost entirely would imperil the company—as things stand.
The picture may change, however, if Nvidia’s commitments keep growing. Andy Li of CreditSights, a financial-research firm, worries that it will keep “pushing the pedal” until “something breaks”. SemiAnalysis estimates that Nvidia takes on roughly $5.9bn of guarantees for every 100 megawatts of data-centre capacity covered by its neocloud backstop programme. If it adds more such guarantees, SemiAnalysis reckons its exposure could reach $175bn by the end of 2028 from this initiative alone.
Nvidia is not alone in using ever more complex financial instruments to lubricate customer demand.
AMD
has offered to sell big stakes in itself to Open
AI
and Meta in exchange for mammoth contracts to buy its chips. In April Google assembled a consortium of firms to help finance Anthropic’s purchase of $35bn of its chips. Broadcom, which develops the chips on Google’s behalf, agreed to cover any shortfall if Anthropic fails to pay.
Amid the race to lock in sales and loans, there is a danger that speculative projects that might otherwise struggle to find financing will get built. Should returns in the industry fall short or take longer to materialise than these convoluted deals assume, the result will be a profusion of expensive, idle chips. No company is more entangled in this web than Nvidia. Mr Huang believes that the
AI
infrastructure buildout is “at full steam”. But Nvidia’s investors should be aware that, the more of this boom they finance, the more of any bust they will have to bear.
■