万一人工智能股市崩盘了怎么办？
内容总结：
人工智能热潮正推动全球股市狂飙，但过度膨胀的泡沫风险已拉响警报。自2022年ChatGPT面世以来，美国股市市值激增21万亿美元，其中亚马逊、博通、英伟达等十大科技巨头贡献了55%的涨幅。AI概念催生的财富神话不断上演：甲骨文创始人埃里森借股价暴涨短暂登顶全球首富，今年上半年美国GDP全部增长来自IT投资热潮，西方三分之一的风险资金涌向AI领域。
狂热背后是"工业革命级"变革预期的驱动。红杉资本等机构宣称AI将重塑经济格局，投资人甚至将这场竞赛形容为"创造数字神明"的角逐。然而现实数据却显露寒意：瑞银报告指出，目前西方头部AI企业年收入仅500亿美元，相较于2025-2028年全球预计投入的29万亿美元数据中心建设资金（不含能源成本）可谓杯水车薪。更严峻的是，麻省理工学院研究显示95%的企业在生成式AI投资上获得"零回报"。
历史经验敲响警钟。对冲基金Praetorian指出当前态势堪比互联网泡沫时期的全球光纤过度建设，瑞银警告"估值红灯已亮起"，阿波罗投资公司数据表明AI股估值甚至超过1999年互联网泡沫顶峰。就连OpenAI首席执行官奥特曼也承认市场存在过度兴奋，但他与其他科技领袖同时辩称：技术革命伴随投机泡沫是历史常态。
纵观1825年至2000年的51项重大技术创新，37项曾出现泡沫，但多数未阻碍技术最终普及——英国铁路泡沫后仍建成密集路网，美国电灯泡沫破灭却未熄灭人们对光明的追求。真正的危险在于泡沫破裂时的洗牌效应：19世纪铁路、电报巨头多数快速更迭，当下风光无限的"七大科技巨头"能否挺过十年尚属未知。
泡沫破裂的社会冲击取决于四大关键因素：引爆原因、资本规模、资产持久性与损失承担者。分析显示，AI泡沫风险在历史泡沫中仅次于19世纪三次铁路大崩盘。当前AI热潮虽由技术突破点燃，但各国政府通过政策加持投入万亿资金，已形成"技术+政治"双引擎驱动。更令人担忧的是，AI资产贬值速度远超传统基础设施——英伟达尖端芯片数年即可能过时，美国科技企业资产平均寿命仅9年。
若泡沫破裂，损失将由现金流充裕的科技巨头、保险公司及养老基金等承担，银行体系直接风险敞口有限。但美国经济面临独特风险：家庭净资产中股票占比达30%的历史新高，且财富高度集中于富人阶层。咨询公司牛津经济测算显示，金融资产每变动1美元将引发14美分的消费波动，而这一切又与少数AI巨头的命运紧密捆绑。
当"数字神明"降临迟缓或最终缺席，这场由技术乐观主义构筑的繁荣防线，或将难以掩盖美国制度动荡、贸易壁垒高筑与政府债务攀升的现实困境。历史提醒我们：技术创新终将穿越泡沫周期，但狂欢后的退潮必然伴随残酷的洗牌。
中文翻译：
若人工智能股市崩盘会怎样？
我们发现潜在成本已攀升至惊人高位
自2022年ChatGPT发布以来，美国股市市值增长了21万亿美元。仅亚马逊、博通和英伟达等十家企业就贡献了其中55%的涨幅。所有这些企业都乘着人工智能的热潮高歌猛进，且并非个例。在AI热潮推动下，甲骨文公司股价飙升，其创始人拉里·埃里森一度成为世界首富。今年上半年，信息技术投资热潮贡献了美国全部的GDP增长；截至当前，西方三分之一的风险投资流向了AI企业。
市场如此火爆，是因为许多人相信AI将重塑经济格局。红杉资本投资者近期宣称其"规模不亚于甚至将超过工业革命"。资产管理公司Atreides Management的加文·贝克在去年播客中表示，AI领军者不仅追求该技术为企业带来"数万亿乃至数百万亿美元价值"，更是在"竞相创造数字上帝"。这种信念足以 justifying任何规模的投入。
AI真能成为神级存在吗？或许可能，但瑞银最新报告显示，目前其收益"令人失望"。据我们测算，西方头部AI企业年收入总额仅为500亿美元。尽管增长迅速，但相较于摩根士丹利预测的2025至2028年全球数据中心累计2.9万亿美元投资（尚未计入能源成本），仍是九牛一毛。AI收入或持续快速增长，但前提是企业持续认可其效用——这并非必然。麻省理工学院最新研究表明，95%的机构从生成式AI投资中获取"零回报"。
难怪越来越多人质疑AI投资已陷入非理性狂热。对冲基金Praetorian Capital指出："环球电讯正重获新生"——该公司在互联网泡沫时期曾过度建设跨大陆光缆。瑞银另一份报告警告："该领域估值确实亮起红灯，几乎没有留给现金流失望的余地"。阿波罗投资公司托尔斯滕·斯洛克注意到，AI股票估值已超过1999年互联网泡沫时期。就连OpenAI掌门人、AI最狂热布道者萨姆·奥尔特曼也在敲警钟："投资者整体是否对AI过度兴奋？我的看法是肯定的。"
这番表态看似惊人，但奥尔特曼之流同时辩称泡沫是新科技诞生的常态。高盛前分析师迈克尔·帕雷克指出："技术热情总是超越技术现实"。旧金山联储2008年研究显示："历史表明，重大技术创新期常伴随投机泡沫，因投资者对生产力真实进步反应过度"。2018年学术研究考察1825至2000年间51项创新，发现其中37项伴随泡沫。
多数泡沫并未阻碍相关技术席卷全球。英国1840年代和1860年代经历两次铁路泡沫，却依然建成密集铁路网。19世纪末美国投资者狂追电灯公司损失惨重，但如今美国人依然需要夜间照明。AI很可能重演此景。泡沫来去无常，数字上帝却可能永恒。
然而崩盘仍将造成重大影响。历史教训是：科技泡沫破裂时，龙头企业常被新锐取代。阿拉斯代尔·奈恩在《推动市场的引擎》中写道："当现金流出现问题，所有最大最成功的照明公司都经历了控制权变更"。铁路、电报和电话早期主导企业也多被快速取代。谁还记得1960年代美国电子泡沫中的Vulcatron，或互联网热潮时期的家喻户晓企业科宁？若十余年后所有"七巨头"科技企业和顶级AI初创仍能存活，堪称奇迹。
对全社会而言，科技崩盘的影响差异巨大。1960年代美国电子泡沫破裂几乎未伤及经济；1870年代铁路泡沫破裂却导致美国史上最长萧条期。我们对历次科技泡沫的分析发现，最关键因素包括：繁荣的引爆点、投资资本性质及损失承担者。
先说引爆点。经济史学家威廉·奎因和约翰·特纳在《繁荣与萧条》中区分了政治与技术引爆点。政治力量（通过调整法规或税收等）催生的泡沫比技术引发的破坏更大。政治诱因促使投资者羊群效应。宽松房产税、低利率和金融自由化导致1980年代末日本巨大资产泡沫，破裂后经济停滞数十年。相反，技术引爆点危害较小：互联网狂热后未出现长期萧条。
资本投资的规模与持久性同样关键。1840年代英国商界对铁路陷入疯狂，1844至1847年间投资占GDP比重从5%飙升至13%。泡沫破裂后投资腰斩，英国失业率翻倍。
资本配置方式也至关重要。1980年代日本电子企业大量资本支出最终徒劳无功。相反，若创造持久资产，泡沫亦能造福社会。铁路狂热构建了英国铁路网主干，尽管盈利姗姗来迟。1990年代末美国铺设的数千万英里光缆远超当时互联网需求，但近年却助力流媒体和视频通话等数据密集型服务。
决定崩盘严重性的最终因素是谁承担损失。当大量散户各损失少量资金时，经济损害有限——美国电子和互联网热潮后便是如此。相反，1860年代英国铁路泡沫破裂时，损失集中在银行体系，导致大量坏账，银行缩减信贷加剧衰退。
AI在这份"恶行名录"中居于何处？我们选取十次历史泡沫，从引爆点、累计资本支出、投资持久性和投资者群体四个维度评估。按我们粗略估算，潜在AI泡沫严重程度仅次于19世纪三次重大铁路泡沫。
AI繁荣由技术引爆，但政客正在火上浇油。2017年奠基论文《注意力就是全部》发表，2022年OpenAI推出ChatGPT，这些发展与政治无关。然而近期各国政府开始扶持AI领军企业。特朗普政府承诺放宽监管，提供实现"全球主导"所需的基础设施和人才。海湾国家政府正投入数万亿美元用于AI投资。
AI资本支出性质亦令人担忧。目前其规模按历史标准尚属适度：我们最宽估算显示，过去四年美国AI企业投资占年GDP的3-4%，而1840年代英国铁路投资占GDP比重达15-20%。但若数据中心建设预测准确，情况将改变。更异常的是，极大比例资本正投向快速贬值的资产。英伟达尖端芯片数年后必将显得笨拙。我们估计美国科技企业资产平均寿命仅九年，远低于1990年代电信资产的十五年。
最后是损失承担者问题。摩根士丹利认为，即将投入的2.9万亿美元数据中心投资中，近半来自科技巨头的现金流。这些公司负债极少，如需为投资融资可大幅增加借款。它们约占标普500指数市值的五分之一，但在投资级债券市场借款占比仅2%，资产负债表看似坚如磐石。
其他大投资者包括保险公司、养老基金、主权财富基金和富豪家族。今年8月，大型债券投资者PIMCO和私人信贷公司Blue Owl为Meta在路易斯安那州290亿美元数据中心扩建提供资金。若AI投资价值归零，这些投资者将受损，但不太可能摧毁金融体系。由于美国银行未直接参与大量AI融资，其风险暴露主要通过非银行贷方间接形成。
但另一方面，美国经济处于历史罕见境地：个人股票市场风险暴露程度空前。股票资产约占美国家庭净值的30%，高于2000年初互联网泡沫顶峰期的26%。这些资产集中在富人阶层，其消费近来推动经济增长。咨询公司牛津经济研究院指出，金融财富每变动1美元，消费支出约波动14美分。而这些波动比以往更依赖少数科技巨头——其前景将由AI决定。
过去一年，技术革命的承诺让人暂时忽视美国制度动荡、贸易壁垒加剧和政府巨额借贷的阴暗现实。若数字上帝未能降临，甚至姗姗来迟，坠落将异常惨烈。
本文由《经济学人》授权翻译，原文发表于www.economist.com
英文来源：
What if the AI stock market blows up?
We find that the potential cost has risen alarmingly high
Since the release of ChatGPT in 2022, the value of America’s stockmarket has risen by $21trn. Just ten firms—including Amazon, Broadcom and Nvidia—account for 55% of the rise. All are riding high on enthusiasm for artificial intelligence, and they are not the only ones. Larry Ellison briefly became the world’s richest man, after AI enthusiasm prompted the share price of Oracle, his firm, to leap. In the first half of the year an IT investment boom accounted for all America’s GDP growth; in the year to date a third of the West’s venture-capital dollars have gone to AI firms.
The market is so hot because many believe AI will transform the economy. Investors at Sequoia Capital, a VC firm, recently argued it will be “as big if not bigger than the Industrial Revolution". In a podcast last year, Gavin Baker of Atreides Management, an asset manager, argued that AI luminaries are not just after the “tens of trillions or hundreds of trillions of value" the tech could add to their firms—they are “in a race to create a Digital God". That belief would justify any amount of spending.
Will AI really become god-like? Perhaps, but a recent report by UBS, a bank, finds that revenues to date have “been disappointing". By our reckoning, total revenues from the tech accruing to the West’s leading AI firms are now $50bn a year. Although such revenues are growing fast, they are still a tiny fraction of the $2.9trn cumulative investment in new data centres globally that Morgan Stanley, another bank, forecasts between 2025 and 2028—a figure which excludes energy costs. AI revenues could continue to grow quickly, but only if firms continue to believe the tech is useful to them, and this is not guaranteed. A recent study by researchers at the Massachusetts Institute of Technology finds that 95% of organisations are getting “zero return" from investments in generative AI.
No wonder more people are asking if AI investment has become irrationally exuberant. “Global Crossing is reborn," argues Praetorian Capital, a hedge fund, referring to the firm that hugely overbuilt cross-continental fibres in the dotcom era. “Valuations in the space are indeed flashing red and leave little room for cashflow disappointments," according to another report by UBS. Torsten Slok of Apollo, a private-investment firm, has noted that AI stocks are more richly valued than dotcom stocks in 1999. Even Sam Altman, boss of OpenAI and one of AI’s most fervent evangelists, is sounding the alarm. “Are we in a phase where investors as a whole are overexcited about AI? My opinion is yes."
This may seem like a striking admission, but Mr Altman and his ilk also argue that bubbles are normal when new technologies emerge. “Tech enthusiasm always runs ahead of tech realities," according to Michael Parekh, a former analyst at Goldman Sachs, yet another bank. “History tells us that periods of major technological innovation are often accompanied by speculative bubbles as investors overreact to genuine advances in productivity," reads a study published in 2008 by the Federal Reserve Bank of San Francisco. An academic study in 2018, which examined 51 innovations from between 1825 and 2000, found that 37 were accompanied by bubbles.
Most did not prevent the technologies that inflated them from sweeping the world. In Britain there were two big railway bubbles, in the 1840s and the 1860s; the country nevertheless has lots of railways. American investors went loopy over electric-light companies in the late 1800s and lost a lot of money, but today Americans still want to see at night. AI may well follow suit. Bubbles come and go, but Digital God would be eternal.
And yet a crash would still have big consequences. One lesson from history is that, when tech bubbles burst, leading firms give way to upstarts. “The biggest and most successful lighting companies all experienced a change of control when cashflow became an issue," wrote Alasdair Nairn in “Engines That Move Markets", which covers the late 19th century. Many firms that dominated the early days of railways, the telegraph and the telephone were also fast supplanted. Who remembers Vulcatron, from America’s electronics bubble of the 1960s, or Corning, a household name during the dotcom boom? It will be a miracle if, in a decade or so, all the “magnificent seven" listed tech firms, and the biggest AI startups, still exist.
For society at large, the consequences of tech crashes vary enormously. The bursting of America’s electronics bubble of the 1960s barely grazed the economy; the bursting of its railway bubble in the 1870s resulted in the longest slump in American history. Our analysis of past technological bubbles finds that a number of factors matter most: what kick-starts the boom, the nature of the capital invested and who bears the losses.
Take the spark first. In their book “Boom and Bust", William Quinn and John Turner, two economic historians, distinguish between political and technological sparks. Bubbles inflated by politicians—through changing regulations or taxes, say—cause more damage than those inflated by new technologies. Political sparks encourage investors to move as a herd. Lenient property taxes, low interest rates and financial liberalisation led to a gargantuan asset bubble in Japan in the late 1980s. For decades after it burst, Japan’s economy remained sluggish. By contrast, technological sparks do less damage: no long slump followed the dotcom mania.
The size and durability of capital investment is also important. In 1840s Britain, businessmen went truly bananas for railways. From 1844 to 1847 investment rose from 5% to 13% of British GDP. Investment fell by half when the bubble burst—and British unemployment doubled.
Then there is the manner in which capital is deployed. Much of the capex by Japanese electronics firms in the 1980s ultimately served no useful function. By contrast, bubbles can benefit society if they create enduring assets. The railway mania built the backbone of England’s rail network, even if profitability took a long time to arrive. The tens of millions of miles of fibre-optic cable laid across America during the late 1990s were far more than the internet needed at the time. But in recent years this has facilitated data-intensive services such as streaming and video calls.
The final factor determining a crash’s severity is who bears the losses. When lots of individual investors each lose a little, the economic damage is limited. This is what happened after America’s electronics and dotcom booms. Amid the British railway bust of the 1860s, by contrast, losses were concentrated among banks, which ended up with lots of bad loans. They then cut new lending, deepening the downturn.
Where might AI sit in the rogue’s gallery? To judge this, we picked ten historical bubbles and assessed them on each factor—spark, cumulative capex, capex durability and investor group. By our admittedly rough-and-ready reckoning, the potential AI bubble lags behind only the three gigantic railway busts of the 19th century.
The spark of the AI boom was technological, but politicians are adding fuel to the fire. A foundational paper titled “Attention is all you need" was published in 2017. OpenAI released ChatGPT in 2022. These developments had nothing to do with politics. Lately, however, governments have begun to support their AI champions. America’s, under Donald Trump, has promised to trim regulation and help provide the infrastructure and workers needed to achieve “global dominance". Gulf countries’ governments are pouring trillions of dollars into AI investment.
The nature of AI capex is also worrying. For now, the splurge looks fairly modest by historical standards. According to our most generous estimate, American AI firms have invested 3-4% of annual American GDP over the past four years. British railway investment in the 1840s was 15-20% of GDP. Yet if forecasts for data-centre construction are correct, that will change. What is more, an unusually large share of capital investment is being devoted to assets that depreciate quickly. Nvidia’s cutting-edge chips will inevitably look clunky in a few years’ time. We estimate that the average American tech firm’s assets have a shelf-life of just nine years, compared with 15 for telecoms assets in the 1990s.
Last is the question of who would bear the losses from a crash. Almost half the forthcoming $2.9trn in data-centre capex, Morgan Stanley reckons, will come from giant tech firms’ cashflows. These companies can borrow a lot more to fund their investments if they wish, since they have little existing debt. They make up about a fifth of the S&P 500 index’s market value, but as borrowers account for only 2% of the investment-grade bond market. Their balance-sheets look rock-solid.
The other big investors are likely to be insurance companies, pension schemes, sovereign-wealth funds and rich families. In August PIMCO, a big bond investor, and Blue Owl, a private-credit firm, funded Meta’s $29bn data-centre expansion in Louisiana. If the value of all AI investments went to zero such investors would suffer, but would be unlikely to bring down the financial system. Since American banks are not financing much of the AI boom themselves, their exposure to it is mostly indirect, through such non-bank lenders.
In another respect, though, America’s economy is in a historically unique position: individuals’ exposure to the stockmarket has never been so high. Ownership of stocks accounts for about 30% of the net worth of American households, compared with 26% in early 2000, at the peak of the dotcom bubble. Such ownership is concentrated among the rich, whose spending has powered economic growth of late. According to Oxford Economics, a consultancy, consumer spending rises and falls by about 14 cents for every dollar change in financial wealth. These changes, in turn, depend more than ever on a few giant firms whose prospects will be shaped by AI.
Over the past year, the promise of technological revolution has been a welcome distraction from the darker reality of America’s shaky institutions, rising trade barriers and vast government borrowing. Should Digital God fail to arrive, or is even slow to arrive, the fall will be brutal.
© 2025, The Economist Newspaper Limited. All rights reserved. From The Economist, published under licence. The original content can be found on www.economist.com