The AI boom is visible from orbit. Satellite photos of New Carlisle, Indiana, show greenish splotches of farmland transformed into unmistakable industrial parks in less than a year’s time. There are seven rectangular data centers there, with 23 more on the way.

人工智能（AI）的蓬勃发展从太空轨道上都能看得一清二楚。印第安纳州新卡莱尔的卫星照片显示，不到一年的时间里，大片绿色的农田已经变成了清晰可见的工业园区。那里有七个矩形的数据中心，还有23个正在建设中。

Inside each of these buildings, endless rows of fridge-size containers of computer chips wheeze and grunt as they perform mathematical operations at an unfathomable scale. The buildings belong to Amazon and are being used by Anthropic, a leading AI firm, to train and run its models. According to one estimate, this data-center campus, far from complete, already demands more than 500 megawatts of electricity to power these calculations—as much as hundreds of thousands of American homes. When all the data centers in New Carlisle are built, they will demand more power than two Atlantas.

在这些建筑的内部，一排排冰箱大小的电脑芯片容器呼呼作响，嗡嗡轰鸣，以不可思议的规模执行着数学运算。这些建筑属于亚马逊公司（Amazon），并被领先的人工智能公司Anthropic用来训练和运行其模型。据一项估算，这个远未完工的数据中心园区，已经需要超过500兆瓦的电力来支持这些计算——这相当于数十万个美国家庭的用电量。当新卡莱尔（New Carlisle）的所有数据中心建成后，它们所需的电量将超过两个亚特兰大（Atlanta）的用电总量。

The amount of energy and money being poured into AI is breathtaking. Global spending on the technology is projected to hit $375 billion by the end of the year and half a trillion dollars in 2026. Three-quarters of gains in the S&P 500 since the launch of ChatGPT came from AI-related stocks; the value of every publicly traded company has, in a sense, been buoyed by an AI-driven bull market. To cement the point, Nvidia, a maker of the advanced computer chips underlying the AI boom, yesterday became the first company in history to be worth $5 trillion.

投入人工智能领域的能源和资金规模之大令人惊叹。预计今年年底全球在该技术上的支出将达到3750亿美元，2026年将达到5000亿美元（半万亿美元）。自ChatGPT发布以来，标准普尔500指数四分之三的涨幅都来自人工智能相关股票；从某种意义上说，每家上市公司的价值都受益于人工智能驱动的牛市。为更有力地说明这一点，支撑人工智能繁荣的先进计算机芯片制造商英伟达（Nvidia）昨天成为历史上第一家市值达到5万亿美元的公司。

Here’s another way of thinking about the transformation under way: Multiplying Ford’s current market cap 94 times over wouldn’t quite get you to Nvidia’s. Yet 20 years ago, Ford was worth nearly triple what Nvidia was. Much like how Saudi Arabia is a petrostate, the U.S. is a burgeoning AI state—and, in particular, an Nvidia-state. The number keeps going up, which has a buoying effect on markets that is, in the short term, good. But every good earnings report further entrenches Nvidia as a precariously placed, load-bearing piece of the global economy.

这是另一种思考当前变革的方式：即使将福特（Ford）目前的市值乘以94，也仍不足以达到英伟达（Nvidia）的市值。然而，20年前，福特的市值几乎是英伟达的三倍。就像沙特阿拉伯是一个“石油国家”一样，美国正成为一个新兴的“人工智能国家”——尤其是一个“英伟达国家”。这个数字（指英伟达的市值）持续上涨，短期内对市场产生了提振作用，这是好的。但每一份亮眼的财报都进一步巩固了英伟达在全球经济中一个举足轻重但又岌岌可危的地位。

America appears to be, at the moment, in a sort of benevolent hostage situation. AI-related spending now contributes more to the nation’s GDP growth than all consumer spending combined, and by another calculation, those AI expenditures accounted for 92 percent of GDP growth during the first half of 2025. Since the launch of ChatGPT, in late 2022, the tech industry has gone from making up 22 percent of the value in the S&P 500 to roughly one-third. Just yesterday, Meta, Microsoft, and Alphabet all reported substantial quarterly-revenue growth, and Reuters reported that OpenAI is planning to go public perhaps as soon as next year at a value of up to $1 trillion—which would be one of the largest IPOs in history. (An OpenAI spokesperson told Reuters, “An IPO is not our focus, so we could not possibly have set a date”; OpenAI and The Atlantic have a corporate partnership.)

美国目前似乎处于一种良性的人质局面。人工智能相关支出对国家GDP增长的贡献，如今已超过所有消费者支出之和；另据一项计算，2025年上半年，这些人工智能支出占GDP增长的92%。自ChatGPT于2022年末推出以来，科技行业在标普500指数中的市值占比已从22%上升到大约三分之一。就在昨天，Meta、微软和Alphabet都报告了可观的季度营收增长；路透社也报道称，OpenAI计划最快于明年上市，估值可能高达1万亿美元——这将是史上规模最大的首次公开募股（IPO）之一。（OpenAI的一位发言人告诉路透社，“首次公开募股不是我们的重点，所以我们不可能设定日期”；OpenAI和《大西洋月刊》存在企业合作关系。）

Many people believe that growth will only continue. “We’re gonna need stadiums full of electricians, heavy equipment operators, ironworkers, HVAC technicians,” Dwarkesh Patel and Romeo Dean, AI-industry analysts, wrote recently. Large-scale data-center build-outs may already be reshaping America’s energy systems. OpenAI has announced that it intends to build at least 30 gigawatts’ worth of data centers—more power than all of New England requires on even the hottest day—and CEO Sam Altman has said he’d eventually like to build a gigawatt of AI infrastructure every week. Other major tech firms have similar ambitions.

许多人认为这种增长只会持续下去。人工智能行业分析师Dwarkesh Patel和Romeo Dean最近写道：“我们将需要数以体育场计的电工、重型设备操作员、钢铁工人以及供暖、通风和空调（HVAC）技术人员。”大规模数据中心的建设可能已经正在重塑美国的能源系统。OpenAI已宣布，它计划建设至少30吉瓦的数据中心——这比整个新英格兰地区在最热的日子所需的电力还要多——而且首席执行官萨姆·奥特曼（Sam Altman）也表示，他最终希望每周建造1吉瓦的人工智能基础设施。其他主要的科技公司也有类似的雄心。

Listen to the AI crowd talk enough, and you’ll get a sense that we may be on the cusp of an infrastructure boom. And yet, something strange is happening to the economy. Even as tech stocks have skyrocketed since 2022, the companies’ share of net profits from S&P 500 companies has hardly budged. Job openings have fallen despite a roaring stock market, 22 states are in or near a recession, and despite data centers propping up the construction industry, U.S. manufacturing is in decline.

如果你听够了人工智能圈内人士的讨论，你会感觉到我们可能正处于一场基础设施建设热潮的边缘。然而，经济却正在发生一些奇怪的事情。尽管自2022年以来科技股一路飙升，但这些公司在标普500指数公司净利润中所占的份额却几乎没有变化。尽管股市一片繁荣，但职位空缺却有所减少，22个州处于或接近衰退，而且尽管数据中心支撑着建筑业，美国制造业却仍在下滑。

It’s clear that AI is both drowning out and obscuring other stories about the wobbling American economy. That’s a concern. But even worse: What if AI’s promise for American business proves to be a mirage? What happens then?

很明显，人工智能既掩盖了也模糊了美国经济摇摇欲坠的其他方面。这是一个令人担忧的问题。但更糟糕的是：如果人工智能对美国商业的承诺最终只是一场海市蜃楼呢？那时会发生什么？

The yawning gap between data-center expenditures and the rest of the economy has caused whispers of bubble to rise to a chorus. A growing number of financial and industry analysts have pointed out the enormous divergence between the historic investments in AI and the tech’s relatively modest revenues. For instance, according to The Information, OpenAI likely made $4 billion last year but lost $5 billion (making the idea of a $1 trillion IPO valuation that much more staggering). From July through September, Microsoft’s investments in OpenAI resulted in losses totaling more than $3 billion. For that same time period, Meta reported rapidly growing costs due to its AI investments, spooking investors and sending its stock down 9 percent.

数据中心支出与经济其他部分之间巨大的差距，使得关于泡沫的低语逐渐演变成一片声浪。越来越多的金融和行业分析师指出，人工智能领域巨大的历史性投资与其产生的相对微薄的收入之间存在巨大差异。例如，根据《The Information》的报道，OpenAI去年营收可能达到40亿美元，但亏损了50亿美元（这使得其1万亿美元的首次公开募股（IPO）估值想法更加令人震惊）。从七月到九月，微软对OpenAI的投资导致了总计超过30亿美元的亏损。在同一时期，Meta报告称，由于其在人工智能领域的投资，成本快速增长，这吓坏了投资者，并使其股价下跌了9%。

Much is in flux. Chatbots and AI chips are getting more efficient almost by the day, while the business case for deploying generative-AI tools remains shaky. A recent report from McKinsey found that nearly 80 percent of companies using AI discovered that the technology had no significant impact on their bottom line. Meanwhile, nobody can say, beyond a few years, just how many more data centers Silicon Valley will need. There are researchers who believe there may already be enough electricity and computing power to meet generative AI’s requirements for years to come.

许多事情都处于不断变化之中。聊天机器人和人工智能芯片几乎每天都在提高效率，但部署生成式人工智能工具的商业前景依然不确定。麦肯锡（McKinsey）最近的一份报告发现，近80%使用人工智能的公司都认为，这项技术对他们的利润没有产生显著影响。与此同时，没有人能预测几年后，硅谷到底还需要多少数据中心。甚至有研究人员认为，现有的电力和计算能力可能已经足以满足未来几年生成式人工智能的需求。

The economic nightmare scenario is that the unprecedented spending on AI doesn’t yield a profit anytime soon, if ever, and data centers sit at the center of those fears. Such a collapse has come for infrastructure booms past: Rapid construction of canals, railroads, and the fiber-optic cables laid during the dot-com bubble all created frenzies of hype, investment, and financial speculation that crashed markets. Of course, all of these build-outs did transform the world; generative AI, bubble or not, may do the same.

经济上最糟糕的设想是，在人工智能领域史无前例的投入短期内（甚至可能永远）无法带来利润，而数据中心正是这些担忧的核心。过去的基础设施热潮也曾出现过类似的崩溃：运河、铁路的快速建设，以及互联网泡沫时期铺设的光纤电缆，都曾引发了炒作、投资和金融投机的狂潮，最终导致市场崩溃。当然，所有这些大规模建设确实改变了世界；生成式人工智能，无论最终是否会形成泡沫，也可能产生同样的影响。

This is why OpenAI, Google, Microsoft, Amazon, and Meta are willing to spend as much as possible, as rapidly as possible, to eke out the tiniest advantage. Even if a bubble pops, there will be winners—each company would like to be the first to build a superintelligent machine. For now, many of these tech companies have cash to burn from their other ventures: Alphabet and Microsoft both made more than $100 billion in profit over the previous fiscal year, while Meta and Amazon both made more than $50 billion. But at some point in the near future, data-center spending will likely outpace even these enormous cash flows, reducing Big Tech’s liquidity and worrying investors. And so, as the AI arms race continues to escalate, the companies are beginning to raise outside money—in other words, take on debt.

这就是为什么OpenAI、谷歌、微软、亚马逊和Meta这些公司愿意尽可能多地、尽可能快地投入资金，只为争取到哪怕一丁点优势。即使泡沫破裂，也总会有赢家——每家公司都渴望成为第一个建造出超级智能机器的人。目前，许多科技公司可以通过其他业务积累的资金来大手笔投入：Alphabet和微软在上一个财年的利润都超过1000亿美元，而Meta和亚马逊的利润也都超过了500亿美元。但在不久的将来，数据中心的开销很可能会超过这些巨大的现金流，从而降低科技巨头的流动性，并引发投资者的担忧。因此，随着人工智能军备竞赛不断升级，这些公司也开始筹集外部资金——换句话说，就是开始举债。

Here is where the bubble dynamics get complicated. Tech firms don’t want to formally take on debt—that is, directly ask investors for loans—because debt looks bad on their balance sheets and could reduce shareholder returns. To get around this, some are partnering with private-equity titans to do some sophisticated financial engineering, Paul Kedrosky, an investor and a financial consultant, told us. These private-equity firms put up or raise the money to build a data center, which a tech company will repay through rent. Data-center leases from, say, Meta can then be repackaged into a financial instrument that people can buy and sell—a bond, in essence. Meta recently did just this: Blue Owl Capital raised money for a massive Meta data center in Louisiana by, in essence, issuing bonds backed by Meta’s rent. And multiple data-center leases can be combined into a security and sorted into what are called “tranches” based on their risk of default. Data centers represent an $800 billion market for private-equity firms through 2028 alone. (Meta has said of its arrangement with Blue Owl that the “innovative partnership was designed to support the speed and flexibility required for Meta’s data center projects.”)

泡沫的运作方式在这里变得复杂起来。科技公司不愿正式承担债务——即直接向投资者借款——因为债务会影响它们的资产负债表，并可能降低股东回报。投资者兼金融顾问保罗·凯德罗斯基（Paul Kedrosky）告诉我们，为了规避这种情况，一些公司正与私募股权巨头合作，进行一些复杂的金融工程。这些私募股权公司出资或筹集资金建设数据中心，然后科技公司会通过支付租金来偿还。比如，Meta公司的数据中心租赁合同可以被重新打包成一种可供人们买卖的金融工具——实质上就是一种债券。Meta最近就这么做了：蓝猫资本（Blue Owl Capital）通过发行以Meta租金为担保的债券，为Meta在路易斯安那州的一个大型数据中心筹集了资金。多个数据中心租赁合同还可以合并成一种证券，并根据其违约风险被分作不同的“分层”（tranches，指将贷款组合成不同风险和回报的投资产品）。仅到2028年，数据中心就将为私募股权公司带来8000亿美元的市场。（Meta曾评价与蓝猫资本的这项安排称，这种“创新型合作旨在支持Meta数据中心项目所需的速度和灵活性。”）

In this way, the data-center financing ends up being a real-estate deal as much as an AI deal. If this sounds complicated, it’s supposed to: The complexity, investment structure, and repackaging make exactly what is going on hard to parse. And if the dynamics also sound familiar, it’s because not two decades ago, the Great Recession was precipitated by banks packaging risky mortgages into tranches of securities that were falsely marketed as high-quality. By 2008, the house of cards had collapsed.

如此一来，数据中心的融资与其说是一笔人工智能交易，不如说更像是一笔房地产交易。如果这听起来很复杂，那正是它本来的面貌：这种复杂性、投资结构以及资产重新打包（证券化）的方式，使得究竟发生了什么难以弄清楚。而如果这些运作模式听起来也很耳熟，那是因为不到二十年前，大衰退（指2008年金融危机）正是由银行将高风险抵押贷款打包成多层证券（“分级证券”）引发的，而这些证券却被虚假地宣传为高质量产品。到2008年，这座“纸牌屋”就彻底倒塌了。

Data-center build-outs aren’t the same as subprime mortgages. Still, there is plenty of precarity baked into these investments. Data centers deteriorate rapidly, unlike the more durable infrastructure of canals, railroads, or even fiber-optic cables. Many of the chips inside these buildings become obsolete within a few years, when Nvidia and its competitors release the next wave of bleeding-edge AI hardware. Meanwhile, the returns on scaling up chatbots are, at present, diminishing. The improvements made by each new AI model are becoming smaller and smaller, making the idea that Silicon Valley can spend its way to superintelligence more tenuous by the day.

数据中心的建设与次级抵押贷款有所不同。然而，这些投资本身就蕴含着巨大的不确定性。数据中心老化速度很快，这与运河、铁路甚至光纤电缆等更耐用的基础设施不同。当英伟达（Nvidia）及其竞争对手发布下一代尖端人工智能硬件时，这些数据中心内部的许多芯片在几年内就会过时。与此同时，目前扩大聊天机器人规模所带来的回报正在减少。每一个新人工智能模型所带来的改进都变得越来越小，这使得硅谷能够通过投入资金达到超级智能的这种想法，日益变得站不住脚。

The people who are paying attention to this cycle are getting anxious. On a scale from one to 10, the AI-bubble concern is: people posting memes of Christian Bale’s character from The Big Short, squinting in disbelief at his computer monitor. If tech stocks fall because of AI companies failing to deliver on their promises, the highly leveraged hedge funds that are invested in these companies could be forced into fire sales. This could create a vicious cycle, causing the financial damage to spread to pension funds, mutual funds, insurance companies, and everyday investors. As capital flees the market, non-tech stocks will also plummet: bad news for anyone who thought to play it safe and invest in, for instance, real estate. If the damage were to knock down private-equity firms (which are invested in these data centers) themselves—which manage trillions and trillions of dollars in assets and constitute what is basically a global shadow-banking system—that could produce another major crash.

关注这一周期的人们正变得焦虑不安。从1到10的尺度来看，人们对人工智能泡沫的担忧程度，就像看到有人发布《大空头》（The Big Short）电影中克里斯蒂安·贝尔（Christian Bale）饰演的角色，不敢置信地眯着眼睛盯着电脑屏幕的表情包一样。如果科技股因为人工智能公司未能兑现承诺而下跌，那么那些高度杠杆化（即大量借贷）投资了这些公司的对冲基金，可能会被迫进行贱卖。这可能会引发一个恶性循环，导致金融损害蔓延到养老基金、共同基金、保险公司和普通投资者身上。随着资本逃离市场，非科技股也将暴跌：这对那些原以为可以稳妥行事而投资（例如）房地产的人来说，无疑是个坏消息。如果这种损害甚至击垮了私募股权公司本身（这些公司投资了数据中心，管理着数万亿美元的资产，并构成了实质上的全球“影子银行系统”），那可能会引发另一次重大崩溃。

For now, money is still pouring into the AI industry. But there’s also something circular about these investments. To wit: OpenAI has agreed to pay $300 billion to Oracle for new computing capacity, Oracle is paying Nvidia tens of billions of dollars for chips to install in one of OpenAI’s data centers, and Nvidia has agreed to invest up to $100 billion in OpenAI as it deploys Nvidia chips. Attempts to illustrate these circular investments have produced a series of byzantine charts that one software engineer referred to on X as “the technocapital hyperobject at the end of time.” The consensus seems to be that although this is legal, it likely cannot go on forever.

目前，资金仍在源源不断地涌入人工智能行业。然而，这些投资中也存在着某种循环性。具体来说，OpenAI 已同意向甲骨文（Oracle）支付 3000 亿美元，以获取新的计算能力；甲骨文则向英伟达（Nvidia）支付数百亿美元购买芯片，用于安装在 OpenAI 的某个数据中心里；而英伟达又同意向 OpenAI 投资高达 1000 亿美元，作为其部署英伟达芯片的回报。试图用图表说明这些循环投资时，往往会产生一系列错综复杂的图示，一位软件工程师在 X（原推特）上将此称为“时间尽头的技术资本超实体”（the technocapital hyperobject at the end of time）。普遍的共识是，尽管这种做法是合法的，但它不太可能永远持续下去。

Maybe it will all work out. Three years ago, the generative-AI industry made functionally no revenue; today, it produces tens of billions of dollars annually, a rate of growth that, eventually, could catch up with all of this spending. Generative-AI tools are currently used by hundreds of millions of people, and it’s hard to imagine that simply ceasing overnight. Perhaps OpenAI or Anthropic will pull off superintelligence, allowing them to, in the words of the Bloomberg columnist Matt Levine, “create God and then ask it for money.”

也许一切都会顺利。三年前，生成式人工智能行业几乎没有收入；如今，它每年产生数百亿美元的收入，这种增长速度最终可能会赶上所有的投入。生成式人工智能工具目前被数亿人使用，很难想象它们会一夜之间停止运行。也许OpenAI或Anthropic将成功实现超级智能，正如彭博社专栏作家马特·莱文（Matt Levine）所说，这将使他们能够“创造上帝，然后向上帝要钱”。

Data centers take time to approve and build; power plants and transmission lines take perhaps even more. Labor is limited, supply chains hit snags, investment waxes and wanes—meaning that even if these data centers are built at the tremendous scale desired by Altman and his competitors, construction and energy constraints may keep the boom from growing too irresponsibly.

数据中心的审批和建设需要时间；发电厂和输电线路可能需要更长时间。劳动力有限，供应链会遇到瓶颈，投资时高时低——这意味着，即使这些数据中心能达到像（OpenAI CEO）奥特曼（Sam Altman）及其竞争对手所期望的那样巨大的规模，建设和能源方面的限制也可能阻止这场繁荣过度无节制地增长。

In any case, as we approach the end of 2025, data centers have become a peculiar cultural object. Their immense scale is a physical reminder of the economic dominance of Silicon Valley companies and their seemingly unchecked ambition. The uneasiness they inspire economically is rooted in memories of 2008 but also of the tech industry’s own financial chicanery, specifically the 2022 crypto crash, which was facilitated by a circular-payment scheme of its own. (FTX, a crypto exchange, and Alameda Research, a hedge fund, both co-founded by Sam Bankman-Fried, were found to be propping each other up: Alameda bought FTX’s bespoke cryptocurrency, and FTX lent Alameda money from its customers’ accounts.) And so, in some way, the externalities of the data-center boom, be they environmental or economic, are tied up in fears of what happens not when these tech companies fail, but when they succeed.

无论如何，随着我们接近2025年底，数据中心已成为一个独特的文化符号。其巨大的规模直观地体现了硅谷公司在经济上的主导地位以及它们似乎不受限制的野心。数据中心在经济上引发的不安，不仅根植于对2008年金融危机的记忆，也来源于科技行业自身的金融诡计，特别是2022年的加密货币崩盘——那次崩盘也是由一种循环支付方案促成的。（例如，由萨姆·班克曼-弗里德（Sam Bankman-Fried）共同创立的加密货币交易所FTX和对冲基金Alameda Research，被发现相互支撑：Alameda购买FTX的定制加密货币，而FTX则从其客户账户中借钱给Alameda。）因此，在某种程度上，数据中心热潮的外部效应，无论是环境方面的还是经济方面的，都与一种深层担忧息息相关：人们所担心的，不是这些科技公司的失败，而是它们的成功将带来的后果。

Boom and bust can feel like two sides of the same coin: Consider also that if AI companies deliver on their massive investments, it would likely mean producing a technology so capable and revolutionary that it wipes out countless jobs and sends an unprecedented shock wave through the global economy before humans have time to adapt. (Perhaps we will be unable to adapt at all.) If they fail, there will likely be unprecedented financial turmoil as well.

繁荣与衰退可能就像一枚硬币的两面：还要考虑到，如果人工智能公司的大笔投资最终取得成功，那很可能意味着它们将创造出一种极其强大和具有颠覆性的技术，这种技术在人类来得及适应之前，就会淘汰无数工作岗位，并给全球经济带来前所未有的冲击波。（或许我们根本无法适应。）如果它们失败了，也可能会出现前所未有的金融动荡。

The biggest lesson of the past two decades of Silicon Valley is that Meta, Amazon, and Google—and even the newer AI labs such as OpenAI—have remade our world and have become unfathomably rich for it, all while being mostly oblivious or uninterested in the fallout. They have chased growth and scale at all costs, and largely, they’ve won. The data-center build-out is the ultimate culmination of that chase: the pursuit of scale for scale itself. In all scenarios, the outcome seems only to be real, painful disruption for the rest of us.

过去二十年硅谷最大的教训是，Meta、亚马逊和谷歌——甚至包括像OpenAI这样的新兴人工智能实验室——已经重塑了我们的世界，并因此变得极其富有，而他们大多对随之而来的负面影响漠不关心或不感兴趣。他们不惜一切代价追求增长和规模，而且在很大程度上，他们成功了。数据中心的建设是这种追求的最终体现：为了规模而追求规模本身。无论在哪种情况下，结果似乎都只是给我们其他人带来真实的、痛苦的颠覆。