支撑人工智能的循环融资上周迈入了重量级行列，芯片制造商英伟达（Nvidia）、超威（AMD）和博通（Broadcom）决定作为基石投资者，出资50亿美元参与Meta平台公司（Meta Platforms）250亿美元的股票出售。
The circular financing underpinning AI took a step into the big league last week with the decision by chip makers Nvidia, AMD and Broadcom to act as $5 billion cornerstone investors in Meta Platforms’ $25 billion stock sale.
这项投资出台之际，外界对人工智能基础设施所需资金规模的担忧日益加剧，并对AI股票造成了沉重打击。随着科技股遭遇挫折，这三家芯片制造商在周五吐光了当周的所有涨幅。
The investment came just before rising concerns about the scale of the finance needed for artificial-intelligence infrastructure hit AI stocks hard. The three chip makers on Friday gave up all their gains for the week as tech stocks stumbled.
循环融资会使任何低迷行情恶化。而且，购买一家部分资金用于购买你产品的融资公司的股份，可以说是最典型的循环融资。这些公司本质上是在对AI押注进行加码，这使得AI股票变得更加波动，因为好消息对它们更有利，而坏消息则更糟糕。
Circular financing would worsen any downturn. And buying shares in a company raising money in part to buy your product is as circular as it gets. The companies are essentially doubling down on AI bets, making AI stocks even more volatile because good news is better for them and bad news worse.
Meta在筹集资金方面并没有遇到困难。在强劲需求以及其他类型股票的大规模发售推动下，该公司扩大了最初的股票发售规模。因此，虽然该公司确实需要资金，但它并不需要从供应商那里获得资金。那么，为什么它们会预先承诺购买呢？
Meta wasn’t having a problem raising money. It upsized its initial stock offer after strong demand alongside a large sale of other types of shares. So while the company certainly needs money, it didn’t need it from its suppliers. Why, then, were they precommitting to buy?
原因可能仅仅是因为它们有能力这样做。既然这些芯片制造商资金充足，为什么不转移一部分资金来换取Meta未来的股份呢？
It could just be because they can. Since the chip makers are flush with cash, why not transfer some over for a stake in Meta’s future?
然而，这一动机违背了基本的投资理论。如果一家公司拥有过多资金，它应该将其返还给股东，而不是试图代表股东挑选股票。股东们可以在AI巨头之间进行挑选，而不是将选择权交给英伟达首席执行官黄仁勋（Jensen Huang）——尽管他很聪明。应该让他把天才般的才华集中在制造更好的芯片上。
That motive runs afoul, though, of basic investment theory. If a company has too much money, it should hand it back to its shareholders, not try to pick stocks on their behalf. The shareholders can pick and choose between the AI giants, rather than leave their choices to Nvidia Chief Executive Jensen Huang, smart as he is. Let him focus his genius on making better chips.
这就留下了一个令人不安的答案：Meta是一个如此庞大的客户，以至于它对芯片制造商拥有影响力。按照这种解读，它们购买股票是确保这家超大规模企业购买其芯片所需的“付费参与”的一部分。
That leaves a troubling answer: Meta is such a big customer that it has power over the chip makers. On this read, their stock purchases are part of a pay-to-play needed to ensure the hyperscaler buys their chips at all.
这将是对AI领域循环融资背后通常的供应商融资的一种颠倒，传统的融资是由芯片制造商资助薄弱的数据中心运营商，条件是这些运营商使用其芯片。
This would be an inversion of the usual vendor financing behind the circularity in AI, which is led by chip makers financing weak data-center operators on condition they use their chips.
如果这种倒置确实在起作用——看起来很有可能——这表明芯片供应领域的竞争比许多人想象的更为激烈。如果 Meta 能让死对头 AMD 和英伟达相互制衡，说明它对两家的产品都很满意。加上 Alphabet 的张量处理单元和亚马逊的 Trainium 芯片也在争夺 AI 市场，大型芯片制造商要想维持溢价将变得更加困难。
If such an inversion is what’s at work—and it looks like it could be—it shows competition in chip supply is stronger than many thought. If Meta is able to play fierce rivals AMD and Nvidia off against each other, it must be happy with both products. With Alphabet’s tensor processing units and Amazon’s Trainium chips also competing for AI, it will become harder for the big chip-makers to demand premium prices.
眼下，芯片制造商总共投入 Meta 股票的 50 亿美元，在热火朝天的 AI 融资圈子里只是九牛一毛，哪怕这是“付费入场”。但这属于 AI 公司之间互联互通日益加深的一个令人担忧的趋势，这越来越意味着一家公司出问题会波及其他公司。
Now, the $5 billion in total put into Meta stock by the chip makers is small beer in the heady world of AI finance, even if it is pay-to-play. But it’s part of a worrisome growth in interconnections between AI companies, which increasingly means problems at one firm will create trouble elsewhere.
这主要体现在债务端，芯片制造商和巨型“超大规模厂商”——Alphabet、亚马逊、Meta 和微软——都在利用强劲的信用评级和庞大的现金储备，帮助实力较弱的算力供应商。
Most of this comes on the debt side, where both chip makers and the giant “hyperscalers”—Alphabet, Amazon, Meta and Microsoft—have been using strong credit ratings and cash piles to help weaker suppliers of computing power, or compute.
供应商融资的基本原理就像汽车经销商一样耳熟能详。如果你买不起崭新的宝马，宝马会借钱给你。你提走车，它完成销售，你连本带息还贷。皆大欢喜。
The basic principle of vendor financing is familiar from the car dealership. If you can’t afford the shiny new Beamer right now, BMW will lend you the money. You get the car, it gets the sale, and you pay off the loan with interest. Everyone’s a winner.
但用于支持芯片销售的结构更为复杂，在许多方面风险大得多。在某些情况下，不是直接借钱，而是类似于宝马联手一家银行，承诺如果次级借款人违约，以最低价格回购车辆。与此同时，它还在扩大产能，以满足突然能获得价格合理贷款的次级借款人带来的所有新需求。
The structures being used to back chip sales are more complicated, and in many ways look a lot more risky. In some cases, rather than lend money, it is as though BMW teamed up with a bank, promising to buy back the car for a minimum price if a subprime buyer defaulted on a loan. At the same time, it is ramping up production to meet all the new demand from subprime borrowers suddenly able to access reasonably priced car loans.
会出什么岔子？问题不仅在于次级借款人可能不还钱。而在于如果他们遇到麻烦，很可能是因为某种影响所有参与方的因素，比如经济衰退或算力需求不及预期。
What could possibly go wrong? It isn’t just that the subprime borrowers might not repay. It’s that if they run into trouble, it’s probably because of something that affects everyone involved, such as an economic downturn or weaker-than-expected demand for compute.
那么，出借人或担保人不仅要为一个数据中心的问题承担责任，而是要承担许多个。这很可能伴随着对新芯片需求的崩溃，以及大量二手芯片或降价算力涌入市场。
Then the lender, or guarantor, isn’t just on the hook for problems at one data center, but lots. This would probably come alongside a collapse in demand for new chips and a flood of secondhand chips or cut-price compute hitting the market.
这并非理论推演。这种情况在2001年就曾发生在电信设备公司身上，当时那些不可靠的互联网泡沫时期客户无力偿还债务，导致这些公司最终核销了数十亿美元的供应商融资。
This isn’t theory. It happened in 2001 to the telecom-equipment companies, which ended up writing off billions of dollars in vendor financing when flaky dot-com customers failed to pay.
记作营收的资金并未反映真实需求，因为这些资金经历了循环流转：作为贷款流出，却在同一天返回以推高销售数据。这些设备制造商售出的设备最终流向了二手市场，并对标价构成了冲击，而当时它们的销售已经陷入困境。
Money booked as revenue didn’t reflect real demand because it had round-tripped, going out as a loan only to come back the same day to boost sales figures. The equipment these gear makers had sold ended up on the used market and undermined sticker prices, just when sales were already struggling.
但愿芯片制造商能比当年的电信公司更多地关注信贷质量，而不是只盯着销售额。芯片制造商和超大规模云服务商也拥有稳健的资产负债表，这也是他们当初能够为较弱的客户和供应商提供融资的原因。
Hopefully the chip makers are less focused on sales and more focused on credit quality than the telcos were. The chip makers and hyperscalers also have solid balance sheets, which is why they’re able to finance weaker customers and suppliers in the first place.
但如果情况真的恶化，人们就会清楚地看到，他们已经在人工智能押注上加倍下注——而且规模极其庞大。
But if things do go badly, it will become clear that they’ve doubled down on their AI bets—and at enormous scale.
在涉及英伟达的众多案例中，有两个案例尤为突出。它是数据中心运营商CoreWeave的第二大股东，该公司股价今年已下跌23%。CoreWeave是英伟达的大客户，英伟达已承诺租用其直到2032年拥有的任何闲置算力。在投资者眼中，CoreWeave已被视为陷入困境的信贷标的；其2031年到期债券的交易收益率比美国国债高出10个百分点以上，高于6月份时的略超4个百分点。
Two examples stand out among many involving Nvidia. It is the second-biggest shareholder in data-center operator CoreWeave, whose shares are down 23% this year. CoreWeave is a huge customer of Nvidia, and Nvidia has promised to rent any spare compute it has until 2032. CoreWeave is treated as distressed credit by investors; its 2031 bonds are trading at more than 10 percentage points above Treasurys, up from just over 4 points in June.
第二个案例在俄亥俄州，英伟达为软银旗下SB Energy正在为OpenAI建造的巨型数据中心提供了1050亿美元的后盾支持。和许多州一样，在期中选举到来之前，俄亥俄州的数据中心已成为政治争议的焦点。
The second is in Ohio, where Nvidia has provided a $105 billion backstop for the monster data center being built for OpenAI by SoftBank’s SB Energy. As in many states, data centers in Ohio have become politically contentious ahead of the midterms.
就像汽车制造商为客户提供融资一样，这种做法在进展顺利时是没有问题的。芯片制造商既能获得额外的销售额，又能获得利息——或者在本案中，获得其所购股票的收益。
Like a carmaker financing its customers, this is fine when it works. The chip maker gets both an extra sale and the interest—or, in this case, gains on the stocks they bought.
如果一切不如预期发展，他们将陷入一场输家通吃的双重押注的劣势一方。
If it doesn’t all work out, though, they end up on the bad side of a double-or-nothing bet.
由于算力市场饱和，其处理器需求崩塌，而他们借出用于过往销售的资金并未全部收回。更糟的是，他们将不得不为旧芯片的担保残值负责。
Demand for their processors collapse because the compute market is saturated, and the money they lent for past sales doesn’t all get paid back. Worse, they will be on the hook for the guaranteed residual value of old chips.
这令投资者感到担忧。“可怕的是，倘若出现回调，固定收益领域和股权领域都存在循环关联性，”丹麦诺和诺德控股首席执行官卡西姆·库泰表示。
This is worrying investors. “What’s scary is the circularity both in fixed income and the equity world should there be a correction,” said Kasim Kutay, CEO of Denmark’s Novo Holdings.
这种相互依存也使市场缺乏多元化。在财务上相互依赖的公司股价应当更紧密地共同波动。
This interdependence also makes the market less diversified. Shares of companies that depend financially on each other should move more closely together.
股东应当为承担这些日益增长的风险支付稍低的价格——并思考如果摇摇欲坠的AI交易倒塌，他们将面临多大的风险敞口。
Shareholders should pay a bit less to cover these growing risks—and think about how exposed they would be if the stumbling AI trade topples over.