芯片失事：英伟达能否避免坠毁？

⽬前的AI数据中⼼建设格局，依赖于两样东西：英伟达芯⽚和借来的资⾦。或许⼈们开始拿英伟达芯⽚去借钱是不可避免的。随着这股狂热持续发酵，我开始担忧AI数据中⼼繁荣背后的脆弱性；当我更深⼊地审视这⼀领域的⾦融层⾯时，并没有感到安⼼。 Nvidia has plowed plenty of money into the AI space, with more than 70 investments in AI companies just this year,according to PitchBook data. Among the billions it’s splashed out, there’s one important category: neoclouds,asexemplified by CoreWeave, the publicly traded, debt-laden company premised on the bet that we will continue buildingdata centers forever. CoreWeave and its ilk have turned around and taken out debt to buy Nvidia chips to put in their datacenters, putting up the chips themselves as loan collateral — and in the process effectively turning $1 in Nvidia investmentinto $5 in Nvidia purchases. This is great for Nvidia. I’m not convinced it’s great for anyone else. 根据PitchBook的数据，Nvidia已经向AI领域投⼊了⼤量资⾦，仅今年就对AI公司进⾏了70多项投资。在其花出去的数⼗亿美元中，有⼀个重要类别：新型云服务（neoclouds），其代表是CoreWeave——这家已上市、背负⼤量债务的公司，押注于我们将永远不断地建设数据中⼼。CoreWeave及其同类公司随后又举债购买Nvidia芯⽚，⽤于其数据中⼼，并将这些芯⽚本⾝作为贷款抵押——在这个过程中，实际上把Nvidia投⼊的1美元转化成了5美元的Nvidia采购额。这对Nvidia来说⾮常不错。我并不确信这对其他任何⼈来说也是好事。 Do you have information about loans in the AI industry? You can reach Liz anonymously atlopatto.46 on Signal using a non-work device. 你是否掌握有关AI⾏业贷款的信息？你可以使⽤⾮⼯作设备，通过Signal匿名联系Liz，账号是lopatto.46。 There has been a lot of talk about the raw technical details of how these chips depreciate, and specifically whether thesechips lose value so fast they make these loans absurd. While I am impressed by the sheer amount of nerd energy put intothis question, I do feel this somewhat misses the point: the loans mean that Nvidia has an incentive to bail out this industryfor as long as it can because the majority of GPU-backed loans are made using Nvidia’s own chips as collateral. ⼈们已经⼤量讨论了这些芯⽚在技术层⾯上的贬值⽅式，尤其是它们是否贬值得如此之快，以⾄于让这些贷款显得荒谬。尽管我对这个问题背后投⼊的巨⼤“技术宅”精⼒印象深刻，但我确实觉得这在某种程度上偏离了重点：这些贷款意味着，英伟达有动⼒在⼒所能及的范围内为这个⾏业兜底，因为绝⼤多数以GPU为抵押的贷款，使⽤的正是英伟达⾃家的芯⽚作为抵押物。 Of course, that also means that if something goes wrong with Nvidia’s business, this whole sector is in trouble. And judgingby the increasing competition its chips face, something could go wrong soon. 当然，这也意味着，⼀旦英伟达的业务出了问题，整个⾏业都会陷⼊⿇烦。⽽从其芯⽚所⾯临的竞争⽇益加剧来看，问题可能很快就会出现。 Can startups outrun chip depreciation — and is it happening faster than they say?初创公司能否跑赢芯⽚贬值的速度——⽽且这种贬值是否⽐他们所说的来得更快？ Loans based on depreciating assets are nothing new. For the terminally finance-brained, products like GPUs register asinterchangeablewidgets (in the sense of “anunnamed article considered for purposes of hypothetical example,” not“gadget” or “software application”) not substantively different from trucks, airplanes, or houses. Soa company likeCoreWeave can package some chips up with AI customer contracts and a few other assets and assemble a valuable enoughbundle to secure debt, typically for buying more chips. If it defaults on the loan, the lender can repossess the collateral, thesame way a bank can repossess a house. 以会贬值的资产为基础的贷款并不是什么新鲜事。对那些在“⾦融脑”已到末期的⼈来说，像GPU这样的产品只是可互换的⼩部件（这⾥指的是“⽤于假设性⽰例的未命名物品”，⽽⾮“⼩玩意⼉”或“软件应⽤”），与卡车、飞机或房屋在本质上并⽆不同。因此，像CoreWeave这样的公司可以把⼀些芯⽚、AI客户合同以及其他资产打包在⼀起，组装成⼀个⾜够有价值的资产组合，从⽽获得债务融资，通常是⽤来购买更多芯⽚的。如果它在贷款上违约，放贷⽅就可以收回抵押品，就像银⾏可以收回房屋⼀样。 One way lenders can hedge their bets against risky assets is by pricing the risk into the interest rate. (There is another wayof understanding debt, and we will get there in a minute.) A 10-year mortgage on a house is currently 5.3 percent.CoreWeave’s first GPU-backed loan, made in 2023, had 14 percent interest in the third quarter of this year. (The ratefloats.) 放贷⽅对冲⾼风险资产的⼀种⽅式，是把风险计⼊利率。（关于如何理解债务，还有另⼀种⽅式，我们稍后会谈到。）⽬前，⼀笔10年期的住房抵押贷款利率为5.3%。CoreWeave在2023年发放的⾸笔以GPU作抵押的贷款，在今年第三季度的利率为14%。（该利率是浮动的。） “You have so many forces acting in making them a natural monopoly, andthis amplifies that.” “有如此多的⼒量在推动它们成为⼀种天然垄断，这也放⼤了这种趋势。” Another way lenders can try to reduce their risk is by asking for a high percentage of collateral relative to the loan. This isexpressed as a loan-to-value ratio (LTV). If I buy a house for $500,000, I usually have to contribute a downpayment — callit 20 percent — and use my loan for the rest. That loan, for $400,000, means I have a (LTV) ratio of 80 percent. 放贷⽅降低风险的另⼀种⽅式，是要求相对于贷款⾦额较⾼⽐例的抵押品。这通常⽤贷款价值⽐（loan-to-value ratio，LTV）来表⽰。⽐如我以50万美元买⼀套房⼦，通常需要先⽀付⼀笔⾸付——假设是20%——其余部分通过贷款⽀付。那么这笔40万美元的贷款，就意味着我的LTV⽐例是80%。 GPU loans’ LTV vary widely, based on how long the loan is, faith in companies’ management teams, and other contractfactors, says Ryan Little,the senior managing director of equipment financing at Trinity Capital, who has made GPU loans.Some of these loans have LTVs as low as 50 percent; others are as high as 110 percent. GPU-backed loans are competitive,and Trinity Capital has occasionally lost deals to other lenders as well as vendor financing programs. GPU贷款的LTV差异很⼤，取决于贷款期限、对公司管理团队的信⼼以及其他合同因素。Trinity Capital设备融资⾼级董事总经理、曾经发放过GPU贷款的Ryan Little表⽰，其中⼀些贷款的LTV低⾄50%，⽽另⼀些则⾼达110%。以GPU作抵押的贷款竞争激烈，Trinity Capital有时也会在交易中输给其他放贷⽅以及⼚商融资项⽬。 The majority of these loans are made on Nvidia chips