英伟达本周早些时候宣布，将向OpenAI投资1000亿美元，以支持其大规模数据中心建设，此举加剧了投资者对人工智能领域潜藏危险金融泡沫的担忧。该领域上市公司和私营公司估值所依据的营收与盈利计算存在逻辑矛盾。
尽管英伟达此次投资规模空前，但这家人工智能芯片制造商已进行一系列“循环交易”——向自身客户注资或提供贷款。各行业都在一定程度上存在供应商融资，但此类循环交易可能使投资者对人工智能的真实需求产生虚高预期。
在以往科技泡沫中，营收“迂回投资”和科技公司为自身客户融资的做法，在泡沫最终破裂时都加剧了损害。尽管此类融资目前对英伟达营收的贡献比例看似有限，但该公司作为全球最具价值上市公司的主导地位意味着其股票已被“精准定价”，任何微小失误都可能对其估值产生远超常规的冲击，进而波及金融市场乃至更广泛的经济领域。
整个人工智能热潮在多大程度上是由英伟达的资金支撑的，这一问题的答案难以明晰，这也正是其令人不安之处。该公司已达成多项投资和融资交易，其中许多交易单笔金额过小，不足以被视为“重大交易”，无需在财务文件中披露，但这些交易累积起来可能产生深远影响。
此外，存在诸多相互关联的循环环节——英伟达投资了一家公司，如OpenAI，而OpenAI又从某云服务商采购服务，该云服务商恰是英伟达的投资对象，然后该云服务提供商又从英伟达购买或租赁图形处理器（GPU）——要理清资金流向绝非易事。
错综复杂的投资网络
在英伟达构建的迂回投资网络里，最为引人注目的两个范例当属OpenAI与Coreweave。除最新注资外，英伟达早在2024年10月就参与了这家快速成长的人工智能公司高达66亿美元的融资轮。英伟达还投资了Coreweave，后者为OpenAI提供数据中心算力支持，同时也是英伟达的客户。截至6月底，英伟达持有Coreweave约7%的股份，目前价值约为30亿美元。
企业从英伟达投资中获得的收益远不止资金本身。英伟达在OpenAI和Coreweave等公司持有股权，使这些企业能够以远低于市场利率的条件获得数据中心项目债务融资，而若没有英伟达的支持，它们无法获得如此优惠利率的融资机会。Seaport Global Securities分析师杰伊·戈德堡（Jay Goldberg）将此类交易比作有人请父母共同签署抵押贷款合同，这能让贷款方确信资金能顺利收回。
为数据中心融资的初创公司通常只能接受高达15%的借款利率，而像微软这样的大型成熟公司可能只需支付6%到9%的利率。在英伟达的支持下，OpenAI和Coreweave的融资利率已接近微软或谷歌享有的水平。
英伟达还签署了一项63亿美元的协议，旨在购买Coreweave无法售予他方的云计算资源。此前该芯片制造商已同意在四年内斥资13亿美元采购Coreweave的云服务。与此同时，Coreweave迄今为止已采购至少25万块英伟达图形处理器——据称其中多数为单价约3万美元的H100 Hopper型号——这意味着Coreweave购买英伟达芯片的支出约达75亿美元。从本质上讲，英伟达对Coreweave的所有投资均以营收形式实现回流。
英伟达还与其他所谓“新云”公司达成了类似的云计算交易。据《The Information》报道，今年夏天英伟达同意在四年内斥资13亿美元，从Lambda租用约1万块自家生产的人工智能芯片——Lambda与Coreweave一样，都运营着数据中心。此外，双方还签署了另一份价值2亿美元的协议，将在未指明的时间段内租用约8000块芯片。
对于那些笃定人工智能存在泡沫的人而言，Lambda的交易显然是泡沫存在的有力佐证。Lambda租用英伟达芯片，然后将使用时间转租给英伟达？它以图形处理器本身价值作为抵押获取借款，进而购置这些芯片。
除对OpenAI和Coreweave进行大额投资外，这家人工智能芯片制造商还在多家上市企业中持有数百万美元股份，这些公司要么采购其图形处理器，要么从事相关芯片技术研发。其中包括芯片设计公司Arm、高性能计算企业Applied Digital、云服务公司Nebius Group以及生物技术公司Recursion Pharmaceuticals。（英伟达近期还斥资50亿美元收购了英特尔4%的股份。与Arm类似，英特尔生产的芯片在某些场景下可替代英伟达的图形处理器，但多数情况下与之形成互补。）
本月早些时候，英伟达还承诺向英国人工智能初创企业投入20亿英镑（约合27亿美元），其中至少5亿英镑将投向英国数据中心运营商Nscale。该企业预计将动用部分资金采购英伟达的图形处理器，用于其正在建设的数据中心。英伟达还表示将通过直接投资以及借助当地风投机构，对多家英国初创企业进行投资，其中部分资金可能以计算资源采购形式回流至OpenAI——无论是直接采购还是通过云服务商间接采购，这些服务商最终都需购买英伟达的图形处理器。
据Dealroom与《金融时报》数据显示，2024年英伟达通过直接投资或旗下企业风险投资部门NVentures，向全球人工智能初创企业投入约10亿美元。这一数额较2022年显著增长——当年OpenAI推出ChatGPT掀起生成式人工智能热潮。
这笔资金最终会有多少以销售形式回流至英伟达？同样难以精确测算。华尔街研究机构NewStreet Research估算，英伟达每向OpenAI投入100亿美元，便能收获价值350亿美元的图形处理器采购或租赁收入，相当于其上一财年年度收入的27%。
互联网泡沫时代的回声
如此回报率无疑使客户融资极具吸引力，但这确实引发了分析师对人工智能估值泡沫的担忧。此类循环交易是以往科技泡沫的典型特征，往往最终对投资者造成反噬。
此次英伟达与OpenAI达成的租赁协议（作为其最新投资的一部分）可能暗藏隐患。通过向OpenAI租赁而非要求其直接购买芯片，使得OpenAI无需为芯片的高折旧率承担会计费用，这最终将改善其利润表现。但这意味着英伟达必须独自承担折旧成本。此外，若人工智能工作负载需求与英伟达首席执行官黄仁勋的乐观预期不符，该公司还将面临图形处理器库存积压风险——届时这些芯片将无人问津。
在部分市场观察者看来，英伟达的最新交易与过往科技繁荣时期的过度行为如出一辙。21世纪初互联网泡沫时期，北电网络（Nortel）、朗讯（Lucent）和思科（Cisco）等电信设备制造商曾向初创企业和电信公司提供贷款以购买其设备。就在2001年泡沫破裂前夕，思科和北电网络向客户提供的融资额已超过其年收入的10%，而五大电信设备制造商的融资总额超过其总收益的123%。
最终，铺设的光纤电缆和安装的交换设备数量远超市场实际需求。当泡沫破裂导致众多客户破产时，电信设备制造商的资产负债表上便背负了坏账。当泡沫破裂时，这种情况导致资产价值缩水幅度远超正常情况下可能出现的水平，在随后的十年间，网络设备企业的市值蒸发了逾90%。
更糟糕的是像光纤巨头环球电讯（Global Crossing）这样直接进行“营收迂回投资”的公司。为达成营收预期，这些公司常在季度末通过向其他公司支付服务费，再由对方回购等值设备的方式完成交易。当泡沫破裂时，环球电讯破产了，其高管最终因营收迂回投资问题支付巨额和解金。
正是对这类交易的记忆，让分析师们至少对英伟达的部分循环投资心存疑虑。Seaport Global分析师戈德堡表示，这些交易带有循环融资印记，堪称“泡沫式行为”的典型范例。
“此举显然会加剧人们对‘循环’问题的担忧。”伯恩斯坦研究公司（Bernstein Research）分析师斯泰西·拉斯贡（Stacy Rasgon）在英伟达宣布对OpenAI进行巨额投资后，在一份投资者报告中写道。诚然，从心怀担忧到陷入危机尚有漫长的距离，但随着人工智能公司估值持续攀升，这段距离正在逐渐缩短。（*）
译者：中慧言-王芳
英伟达本周早些时候宣布，将向OpenAI投资1000亿美元，以支持其大规模数据中心建设，此举加剧了投资者对人工智能领域潜藏危险金融泡沫的担忧。该领域上市公司和私营公司估值所依据的营收与盈利计算存在逻辑矛盾。
尽管英伟达此次投资规模空前，但这家人工智能芯片制造商已进行一系列“循环交易”——向自身客户注资或提供贷款。各行业都在一定程度上存在供应商融资，但此类循环交易可能使投资者对人工智能的真实需求产生虚高预期。
在以往科技泡沫中，营收“迂回投资”和科技公司为自身客户融资的做法，在泡沫最终破裂时都加剧了损害。尽管此类融资目前对英伟达营收的贡献比例看似有限，但该公司作为全球最具价值上市公司的主导地位意味着其股票已被“精准定价”，任何微小失误都可能对其估值产生远超常规的冲击，进而波及金融市场乃至更广泛的经济领域。
整个人工智能热潮在多大程度上是由英伟达的资金支撑的，这一问题的答案难以明晰，这也正是其令人不安之处。该公司已达成多项投资和融资交易，其中许多交易单笔金额过小，不足以被视为“重大交易”，无需在财务文件中披露，但这些交易累积起来可能产生深远影响。
此外，存在诸多相互关联的循环环节——英伟达投资了一家公司，如OpenAI，而OpenAI又从某云服务商采购服务，该云服务商恰是英伟达的投资对象，然后该云服务提供商又从英伟达购买或租赁图形处理器（GPU）——要理清资金流向绝非易事。
错综复杂的投资网络
在英伟达构建的迂回投资网络里，最为引人注目的两个范例当属OpenAI与Coreweave。除最新注资外，英伟达早在2024年10月就参与了这家快速成长的人工智能公司高达66亿美元的融资轮。英伟达还投资了Coreweave，后者为OpenAI提供数据中心算力支持，同时也是英伟达的客户。截至6月底，英伟达持有Coreweave约7%的股份，目前价值约为30亿美元。
企业从英伟达投资中获得的收益远不止资金本身。英伟达在OpenAI和Coreweave等公司持有股权，使这些企业能够以远低于市场利率的条件获得数据中心项目债务融资，而若没有英伟达的支持，它们无法获得如此优惠利率的融资机会。Seaport Global Securities分析师杰伊·戈德堡（Jay Goldberg）将此类交易比作有人请父母共同签署抵押贷款合同，这能让贷款方确信资金能顺利收回。
为数据中心融资的初创公司通常只能接受高达15%的借款利率，而像微软这样的大型成熟公司可能只需支付6%到9%的利率。在英伟达的支持下，OpenAI和Coreweave的融资利率已接近微软或谷歌享有的水平。
英伟达还签署了一项63亿美元的协议，旨在购买Coreweave无法售予他方的云计算资源。此前该芯片制造商已同意在四年内斥资13亿美元采购Coreweave的云服务。与此同时，Coreweave迄今为止已采购至少25万块英伟达图形处理器——据称其中多数为单价约3万美元的H100 Hopper型号——这意味着Coreweave购买英伟达芯片的支出约达75亿美元。从本质上讲，英伟达对Coreweave的所有投资均以营收形式实现回流。
英伟达还与其他所谓“新云”公司达成了类似的云计算交易。据《The Information》报道，今年夏天英伟达同意在四年内斥资13亿美元，从Lambda租用约1万块自家生产的人工智能芯片——Lambda与Coreweave一样，都运营着数据中心。此外，双方还签署了另一份价值2亿美元的协议，将在未指明的时间段内租用约8000块芯片。
对于那些笃定人工智能存在泡沫的人而言，Lambda的交易显然是泡沫存在的有力佐证。Lambda租用英伟达芯片，然后将使用时间转租给英伟达？它以图形处理器本身价值作为抵押获取借款，进而购置这些芯片。
除对OpenAI和Coreweave进行大额投资外，这家人工智能芯片制造商还在多家上市企业中持有数百万美元股份，这些公司要么采购其图形处理器，要么从事相关芯片技术研发。其中包括芯片设计公司Arm、高性能计算企业Applied Digital、云服务公司Nebius Group以及生物技术公司Recursion Pharmaceuticals。（英伟达近期还斥资50亿美元收购了英特尔4%的股份。与Arm类似，英特尔生产的芯片在某些场景下可替代英伟达的图形处理器，但多数情况下与之形成互补。）
本月早些时候，英伟达还承诺向英国人工智能初创企业投入20亿英镑（约合27亿美元），其中至少5亿英镑将投向英国数据中心运营商Nscale。该企业预计将动用部分资金采购英伟达的图形处理器，用于其正在建设的数据中心。英伟达还表示将通过直接投资以及借助当地风投机构，对多家英国初创企业进行投资，其中部分资金可能以计算资源采购形式回流至OpenAI——无论是直接采购还是通过云服务商间接采购，这些服务商最终都需购买英伟达的图形处理器。
据Dealroom与《金融时报》数据显示，2024年英伟达通过直接投资或旗下企业风险投资部门NVentures，向全球人工智能初创企业投入约10亿美元。这一数额较2022年显著增长——当年OpenAI推出ChatGPT掀起生成式人工智能热潮。
这笔资金最终会有多少以销售形式回流至英伟达？同样难以精确测算。华尔街研究机构NewStreet Research估算，英伟达每向OpenAI投入100亿美元，便能收获价值350亿美元的图形处理器采购或租赁收入，相当于其上一财年年度收入的27%。
互联网泡沫时代的回声
如此回报率无疑使客户融资极具吸引力，但这确实引发了分析师对人工智能估值泡沫的担忧。此类循环交易是以往科技泡沫的典型特征，往往最终对投资者造成反噬。
此次英伟达与OpenAI达成的租赁协议（作为其最新投资的一部分）可能暗藏隐患。通过向OpenAI租赁而非要求其直接购买芯片，使得OpenAI无需为芯片的高折旧率承担会计费用，这最终将改善其利润表现。但这意味着英伟达必须独自承担折旧成本。此外，若人工智能工作负载需求与英伟达首席执行官黄仁勋的乐观预期不符，该公司还将面临图形处理器库存积压风险——届时这些芯片将无人问津。
在部分市场观察者看来，英伟达的最新交易与过往科技繁荣时期的过度行为如出一辙。21世纪初互联网泡沫时期，北电网络（Nortel）、朗讯（Lucent）和思科（Cisco）等电信设备制造商曾向初创企业和电信公司提供贷款以购买其设备。就在2001年泡沫破裂前夕，思科和北电网络向客户提供的融资额已超过其年收入的10%，而五大电信设备制造商的融资总额超过其总收益的123%。
最终，铺设的光纤电缆和安装的交换设备数量远超市场实际需求。当泡沫破裂导致众多客户破产时，电信设备制造商的资产负债表上便背负了坏账。当泡沫破裂时，这种情况导致资产价值缩水幅度远超正常情况下可能出现的水平，在随后的十年间，网络设备企业的市值蒸发了逾90%。
更糟糕的是像光纤巨头环球电讯（Global Crossing）这样直接进行“营收迂回投资”的公司。为达成营收预期，这些公司常在季度末通过向其他公司支付服务费，再由对方回购等值设备的方式完成交易。当泡沫破裂时，环球电讯破产了，其高管最终因营收迂回投资问题支付巨额和解金。
正是对这类交易的记忆，让分析师们至少对英伟达的部分循环投资心存疑虑。Seaport Global分析师戈德堡表示，这些交易带有循环融资印记，堪称“泡沫式行为”的典型范例。
“此举显然会加剧人们对‘循环’问题的担忧。”伯恩斯坦研究公司（Bernstein Research）分析师斯泰西·拉斯贡（Stacy Rasgon）在英伟达宣布对OpenAI进行巨额投资后，在一份投资者报告中写道。诚然，从心怀担忧到陷入危机尚有漫长的距离，但随着人工智能公司估值持续攀升，这段距离正在逐渐缩短。（*）
译者：中慧言-王芳
Nvidia’s announcement earlier this week that it is investing $100 billion into OpenAI to help fund its massive data center build out has added to a growing sense of unease among investors that there is a dangerous financial bubble around AI, and that the revenues and earnings math underpinning the valuations of both public and private companies in the sector just doesn’t add up.
While Nvidia’s latest announcement is by far the largest example, the AI chipmaker has engaged in a series of “circular” deals in which it invests in, or lends money to, its own customers. Vendor financing exists to some degree in many industries, but in this case, circular transactions may give investors an inflated perception of the true demand for AI.
In past technology bubbles, revenue “roundtripping” and tech companies financing their own customers have exacerbated the damage when those bubbles eventually popped. While the share of Nvidia’s revenues that are currently being driven by such financing appears to be relatively small, the company’s dominance as the world’s most valuable publicly-traded company means that its stock is “priced for perfection” and that even minor missteps could have outsized impact on its valuation—and on financial markets and perhaps even the wider economy.
The extent to which the entire AI boom is backstopped by Nvidia’s cash isn’t easy to answer precisely, which is also one of the unsettling things about it. The company has struck a number of investment and financing deals, many of which are too small individually for the company to consider “material” and report in its financial filings, even though collectively they may be significant.
In addition, there are so many interlocking rings of circularity—where Nvidia has invested in a company, such as OpenAI, that in turn purchases services from a cloud service provider that Nvidia has also invested in, which then also buys or leases GPUs from Nvidia—that disentangling what money is flowing where is far from easy.
Tangled webs of investment
Two of the most prominent examples of Nvidia’s web of circuitous investments are OpenAI and Coreweave. In addition to the latest investment in OpenAI, Nvidia had previously participated in a $6.6 billion investment round in the fast-growing AI company in October 2024. Nvidia also has invested in CoreWeave, which supplies data center capacity to OpenAI and is also an Nvidia customer. As of the end of June, Nvidia owned about 7% of Coreweave, a stake worth about $3 billion currently.
The benefits that companies get from a Nvidia investment extend beyond the cash itself. Nvidia’s equity stakes in companies such as OpenAI and Coreweave enable these companies to access debt financing for data center projects at potentially significantly lower interest rates than they would be able to access without such backing. Jay Goldberg, an analyst with Seaport Global Securities, compares such deals to someone asking their parents to be a co-signer on their mortgage. It gives lenders some assurance that they may actually get their money back.
Startups financing data centers have often had to borrow money at rates as high as 15%, compared to 6% to 9% that a large, established corporation such as Microsoft might have to pay. With Nvidia’s backing, OpenAI and Coreweave have been able to borrow at rates closer to what Microsoft or Google might pay.
Nvidia has also signed a $6.3 billion deal to purchase any cloud capacity that CoreWeave can’t sell to others. The chipmaker had previously agreed to spend $1.3 billion over four years on cloud computing with CoreWeave. Coreweave, meanwhile, has purchased at least 250,000 Nvidia GPUs so far—the majority of which it says are H100 Hopper models, which cost about $30,000 each—which means Coreweave has spent about $7.5 billion buying these chips from Nvidia. So in essence, all of the money Nvidia has invested in Coreweave has come back to it in the form of revenue.
Nvidia has struck similar cloud computing deals with other so-called “neo-cloud” companies. According to a story in The Information, Nvidia agreed this summer to spend $1.3 billion over four years renting some 10,000 of its own AI chips from Lambda, which like Coreweave runs data centers, as well as a separate $200 million deal to rent some 8,000 more over an unspecified time period.
For those who believe there’s an AI bubble, the Lambda deal is clear evidence of froth. Those Nvidia chips Lambda is renting time on back to Nvidia? It bought them with borrowed money collateralized by the value of the GPUs themselves.
Besides its large investments in OpenAI and Coreweave, AI chipmaker also holds multi-million dollar stakes in several other publicly-traded companies that either purchase its GPUs or work on related chip technology. These include chip design firm Arm, high-performance computing company Applied Digital, cloud services company Nebius Group, and biotech company Recursion Pharmaceuticals. (Nvidia also recently purchased a 4% stake in Intel for $5 billion. Like Arm, Intel makes chips that in some cases are alternatives to Nvidia’s GPUs, but which for the most part are complementary to them.)
Earlier this month, Nvidia also pledged to invest £2 billion ($2.7 billion) in U.K. AI startups, including at least £500 million in Nscale, a U.K.-based data center operator that will, presumably, be using some of that money to purchase Nvidia GPUs to provision the data centers it is building. Nvidia also said it would invest in a number of British startups, both directly and through local venture capital firms, and some of that money too, will likely come back to OpenAI in the form of computing purchases, either directly, or through cloud service providers, who in turn will need to buy Nvidia GPUs.
In 2024, Nvidia invested about $1 billion in AI startups globally either directly or through its corporate venture capital arm NVentures, according to data from Dealroom and The Financial Times. This amount was up significantly from what Nvidia invested in 2022, the year the generative AI boom kicked off with OpenAI’s debut of ChatGPT.
How much of this money winds up coming right back to Nvidia in the form of sales is again, difficult to determine. Wall Street research firm NewStreet Research has estimated that for every $10 billion Nvidia invests in OpenAI, it will see $35 billion worth of GPU purchases or GPU lease payments, an amount equal to about 27% of its annual revenues last fiscal year.
Echoes of the dotcom era
That kind of return would certainly make this sort of customer financing worthwhile. But it does raise concerns among analysts about a bubble in AI valuations. These kinds of circular deals have been a hallmark of previous technology bubbles and have often come back to haunt investors.
In this case, the lease arrangements that Nvidia is entering into with OpenAI as part of its latest investment could prove problematic. By leasing GPUs to OpenAI, rather than requiring them to buy the chips outright, Nvidia is sparing OpenAI from having to take an accounting charge for the high depreciation rates on the chips, which will ultimately help OpenAI’s bottom line. But it means that instead Nvidia will have to bear this depreciation costs. What’s more, Nvidia will also take on the risk of being stuck with an inventory of GPUs no one wants if demand for AI workloads don’t match Nvidia CEO Jensen Huang’s rosy predictions.
To some market watchers, Nvidia’s latest deals feel all-too-similar to the excesses of past technology booms. During the dot com bubble at the turn of the 21st Century, telecom equipment makers such as Nortel, Lucent, and Cisco lent money to startups and telecom companies to purchase their equipment. Just before the bubble burst in 2001, the amount of financing Cisco and Nortel had extended to their customers exceeded 10% of annual revenues, and the amount of financing the top five telecom equipment makers had provided to customers exceeded 123% of their combined earnings.
Ultimately, the amount of fiber-optic cabling and switching equipment installed far exceeded demand, and when the bubble burst and many of those customers went bust, the telecom equipment makers were left holding the bad debt on their balance sheets. This contributed to a greater loss of value when the bubble burst than would have otherwise been the case, with networking equipment businesses losing more than 90% of their value over the ensuing decade.
Worse yet were companies such as fiber-optic giant Global Crossing that engaged in direct “revenue roundtripping.” These companies cut deals—often at the end of a quarter in order to hit topline forecasts—in which they paid money to another company for services, and then that company agreed to purchase equipment of exactly equal value. When the bubble burst, Global Crossing went bankrupt, and its executives ultimately paid large legal settlements related to revenue roundtripping.
It is memories of these kinds of transactions that have caused analysts to at least raise an eyebrow at some of Nvidia’s circular investments. Goldberg, the Seaport Global analyst, said the deals had a whiff of circular financing and were emblematic of “bubble-like behavior.”
“The action will clearly fuel ‘circular’ concerns,” Stacy Rasgon, an analyst with Bernstein Research, wrote in an investor note following Nvidia’s announcement of its blockbuster investment in OpenAI. It’s a long way from a concern to a crisis, of course, but as AI company valuations get higher, that distance is starting to close.