The returns on trillions of dollars of spending are deeply uncertain
数万亿美元支出的回报仍高度不确定YOU AIN’T seen nothing yet. Last year America’s biggest technology companies, including Amazon, Google and Microsoft, spent $450bn on infrastructure, much of it to power artificial intelligence. This was just an amuse-bouche. For the main course they will this year spend $900bn on chips, data centres, power and so forth, with a $1.4trn pudding to follow in 2027. To fund this feast they have borrowed more than $400bn this year. The AI capex boom is fast becoming the largest investment surge in history.你还没见识真正的大场面。去年，美国最大的科技公司，包括 Amazon、Google 和 Microsoft，在基础设施上花费了 4500 亿美元，其中很大一部分用于支持人工智能。这还只是 amuse-bouche，也就是开胃小点。今年的主菜，是它们将在芯片、数据中心、电力等方面花费 9000 亿美元；而 2027 年还会有一道 1.4 万亿美元的甜点。为了资助这场盛宴，它们今年已经借入超过 4000 亿美元。AI capex boom，也就是 AI 资本开支热潮，正迅速成为历史上最大规模的投资浪潮。If superintelligence is in reach, building football fields’ worth of compute could also be history’s most valuable capital-allocation exercise. And yet capital spending can still generate disappointing returns for investors. Since peaking in June, the share prices of the biggest AI firms have fallen by 15%, as worries have mounted that healthy returns will be hard to come by. After Alphabet, Google’s parent, reported second-quarter earnings on July 22nd, its shares dropped by 7%. Results are due from Microsoft and Meta on July 29th and from Amazon the next day. On July 28th South Korea’s benchmark index, dominated by Samsung Electronics and SK Hynix, two big chipmakers, fell by a tenth.如果 superintelligence，也就是超级智能，已经触手可及，那么建设大片如足球场般规模的 compute，也就是算力，也可能成为历史上最有价值的资本配置行动。然而，资本支出仍可能给投资者带来令人失望的回报。自 6 月见顶以来，最大 AI 公司的股价已经下跌 15%，因为市场越来越担心，高回报将很难获得。7 月 22 日，Google 母公司 Alphabet 公布第二季度财报后，其股价下跌 7%。Microsoft 和 Meta 将于 7 月 29 日公布业绩，Amazon 则在次日公布。7 月 28 日，South Korea 的基准指数下跌十分之一；该指数主要由两大芯片制造商 Samsung Electronics 和 SK Hynix 主导。A back-of-the-envelope calculation finds that covering AI capex through identifiable AI income requires revenue on the order of $2.5trn per year, more than tech’s entire combined revenue today. Only a small minority of consumers seem willing to pay for personal AI subscriptions, so the real money will have to be made in selling to enterprises. Traders might use Microsoft’s Copilot to create better financial models, for instance, while schools could teach children with models from Google. For now, though, sales in the trillions are a long way off.一个 back-of-the-envelope calculation，也就是粗略估算，发现如果要用可识别的 AI 收入覆盖 AI 资本开支，每年需要约 2.5 万亿美元收入，这比整个科技行业目前的总收入还要多。似乎只有少数消费者愿意为个人 AI 订阅付费，因此真正的大钱必须来自面向企业的销售。例如，交易员可能使用 Microsoft 的 Copilot 创建更好的金融模型，而学校可以用 Google 的模型教孩子。不过就目前而言，数万亿美元级别的销售额仍然遥不可及。For AI revenues to soar, more firms will have to use AI, what economists call an increase at the “extensive margin”, and use it more deeply, the “intensive margin”. At the extensive margin, roughly 20% of American firms used AI “in any…business functions” in the previous fortnight, according to the Census Bureau’s latest two-weekly survey. In Britain, official figures say about a third of firms claim to use AI, though the question is put differently. That indicates rapid technological diffusion by any standard: after all, four years ago no one used large language models.要让 AI 收入飙升，更多企业必须使用 AI，这就是经济学家所说的 “extensive margin”，也就是广度边际增加；同时企业还必须更深入地使用 AI，这就是 “intensive margin”，也就是强度边际增加。在广度边际上，根据 Census Bureau 最新的双周调查，大约 20% 的美国企业在过去两周内，在“任何……业务职能”中使用过 AI。在 Britain，官方数据显示约三分之一企业声称使用 AI，尽管问题问法不同。无论按什么标准看，这都表明技术扩散非常迅速：毕竟四年前还没有人使用大型语言模型。Yet those numbers may have recently levelled off. In May economists at the Census Bureau found that “AI use remained relatively steady in many sectors over the last six months.” A survey by Jon Hartley of the University of Texas at Austin and others finds that after peaking at 46% of the workforce in mid-2025, now about 33% of people use AI at work. If, say, a third of firms across the OECD club of mostly rich countries adopt AI, then to generate $2.5trn of AI revenues the firms would have to spend about $100,000 a year on average. Is that plausible?然而，这些数字最近可能已经趋于平稳。5 月，Census Bureau 的经济学家发现，“过去六个月，许多行业的 AI 使用率保持相对稳定。”University of Texas at Austin 的 Jon Hartley 及其合作者进行的一项调查发现，在 2025 年中期达到劳动力 46% 的峰值之后，如今约 33% 的人在工作中使用 AI。假设在 OECD 这个主要由富裕国家组成的俱乐部中，有三分之一企业采用 AI，那么为了产生 2.5 万亿美元 AI 收入，这些企业平均每年必须花费约 10 万美元。这可信吗？Perhaps, though for now few firms treat AI as a core technology, which limits how much they are willing to spend on it. According to a survey by the European Central Bank, in late 2025 only a tenth of euro-area companies using AI reported doing so “intensively”. The Bundesbank finds that about half of German firms using AI do so for 5% of working hours or less. Ivan Yotzov of the Bank of England, and colleagues, has found that the average American executive uses AI for 1.7 hours a week—enough time to create a decent PowerPoint presentation, but not much more.也许可信，但目前很少有企业把 AI 视为 core technology，也就是核心技术，这限制了它们愿意在 AI 上花多少钱。根据 European Central Bank 的一项调查，2025 年末，使用 AI 的 euro-area 公司中，只有十分之一表示自己是在 “intensively”，也就是密集地使用 AI。Bundesbank 发现，约一半使用 AI 的德国企业，其 AI 使用时间占工作时间的 5% 或更少。Bank of England 的 Ivan Yotzov 及其同事发现，美国高管平均每周使用 AI 1.7 小时——这足够做出一份不错的 PowerPoint 演示文稿，但也就仅此而已。For these dilettantes, free or ultra-cheap AI models are often good enough. Official data from Britain suggest that close to half of British businesses using AI do not pay for it, presumably making do with the free tier of an American model or an open-source Chinese one. Ramp, a fintech firm, finds that a fraction of firms spend thousands of dollars a month per employee on AI. The median firm’s monthly spending per worker in June, however, was $10.66. Intuit, a software firm which tracks small and medium-sized businesses in America, Britain and Canada, reports that about one in ten has paid for a dedicated AI tool.对这些 dilettantes，也就是浅尝辄止的业余使用者来说，免费或超低价 AI 模型通常已经足够好。Britain 官方数据显示，接近一半使用 AI 的英国企业并不为此付费，大概只是凑合使用美国模型的免费层，或中国开源模型。金融科技公司 Ramp 发现，有一小部分企业每月为每名员工在 AI 上花费数千美元。然而，6 月份企业每名员工的月度支出中位数只有 10.66 美元。软件公司 Intuit 追踪 America、Britain 和 Canada 的中小企业，它报告称，大约十分之一中小企业为专用 AI 工具付过费。Total AI spending can only be guessed at, because data sources are murky and the picture is changing fast. Exponential View, a consultancy, counts $175bn of generative-AI revenue, on an annualised basis, in June. In a recent paper Anton Korinek of Anthropic and Patrick McKelvey of the Bank of Canada estimate total “AI services” revenue. Adapting their methodology, we reckon this was $220bn, again annualised, in the first quarter of this year. Ramp’s data imply that 2-3% of business spending now goes on AI, pointing to $170bn a year.AI 总支出只能估算，因为数据来源并不清晰，而且局面变化很快。咨询公司 Exponential View 统计，按年化口径计算，6 月 generative-AI，也就是生成式 AI 收入为 1750 亿美元。在最近一篇论文中，Anthropic 的 Anton Korinek 和 Bank of Canada 的 Patrick McKelvey 估算了总 “AI services” revenue，也就是 AI 服务收入。我们调整他们的方法后估计，今年第一季度这一数字为 2200 亿美元，同样是年化口径。Ramp 的数据则暗示，如今 2-3% 的企业支出流向 AI，对应每年 1700 亿美元。All these complex calculations roughly tally with a much simpler one: adding up the AI revenue of the firms selling most of the AI. Anthropic pulls in perhaps $75bn, annualised; OpenAI makes tens of billions; Google, via its AI model Gemini, and Microsoft probably get a bit less. SpaceX may have a few billion dollars’ worth of revenue from enterprise AI this year. Meta also makes a few bucks from AI. Add this up and you land at roughly $150bn a year.所有这些复杂计算，大致都与一种简单得多的方法相吻合：把那些销售大部分 AI 的公司收入加起来。Anthropic 年化收入可能达到 750 亿美元；OpenAI 收入达数百亿美元；Google 通过其 AI 模型 Gemini，以及 Microsoft，可能收入略少一些。SpaceX 今年可能从企业 AI 获得数十亿美元收入。Meta 也从 AI 中赚了一些钱。把这些加起来，结果大约是每年 1500 亿美元。Revenue is rising extraordinarily fast. But perhaps not fast enough for investors, as calculations by Phurichai Rungcharoenkitkul of the Bank for International Settlements suggest. The crux is that AI usage, as proxied by the consumption of tokens, is probably growing even more quickly than revenue. This may be because more people are flitting from one free model to the next, rather than paying. In Mr Rungcharoenkitkul’s analysis, this has an important implication: at least part of the AI capex splurge represents zero-sum competition. Firms are using some of the extra compute they are building not to expand the paying market, but to poach customers from rivals. The paper implies that perhaps one-third of the investment, or even more, is unlikely to make a satisfactory return.收入增长速度极快。但 Bank for International Settlements 的 Phurichai Rungcharoenkitkul 的计算表明，对投资者来说，这可能还不够快。关键在于，以 tokens 消耗量作为 proxy，也就是代理指标来看，AI 使用量很可能比收入增长得更快。这可能是因为越来越多人不是付费，而是在一个免费模型和另一个免费模型之间来回切换。在 Rungcharoenkitkul 先生的分析中，这有一个重要含义：AI 资本开支狂潮中至少有一部分代表 zero-sum competition，也就是零和竞争。企业正在使用它们建设的部分额外算力，不是为了扩大付费市场，而是为了从竞争对手那里抢客户。该论文暗示，也许三分之一甚至更多的投资，不太可能获得令人满意的回报。That said, revenues may soon grow even more quickly, if two conditions are met. The first is that AI boosts productivity markedly. For now, there is little evidence that AI is transforming businesses. Few firms are saving money by replacing workers with bots. According to Mr Yotzov’s study, nine in ten executives report no impact of AI on their firm’s productivity over the past three years. If that were to change, though, more firms would see reason to devote greater resources to the technology. They would also be happier to absorb price rises, juicing revenue further.话虽如此，如果两个条件得到满足，收入很快可能会增长得更快。第一个条件是 AI 显著提高 productivity，也就是生产率。目前，几乎没有证据表明 AI 正在改造企业。很少有公司通过用机器人替代员工来省钱。根据 Yotzov 先生的研究，十名高管中有九名表示，过去三年 AI 对其公司的生产率没有影响。不过，如果这种情况发生变化，更多企业就会看到投入更多资源到这项技术的理由。它们也会更愿意接受价格上涨，从而进一步推高收入。The second is that AI adds to “intangible capital”. For firms to make the most of AI they cannot simply pay for a chatbot, but must instead rework their processes, including their staff and their use of data, from top to bottom. The historical evidence suggests that for every $1 of investment in computer hardware, companies have made $5-10 of these intangible investments—which implies they would need to spend trillions of dollars a year if AI is to reach its potential.第二个条件是 AI 增加 “intangible capital”，也就是无形资本。企业若要充分利用 AI，不能只是为一个聊天机器人付费，而必须从上到下重塑自己的流程，包括员工安排和数据使用方式。历史证据表明，企业每投资 1 美元计算机硬件，就会进行 5 至 10 美元这类无形投资。这意味着，如果 AI 要发挥潜力，企业每年需要花费数万亿美元。So far, however, there is little evidence of an intangible-investment boom. Data-organisation firms like Palantir have revenues in the billions, not hundreds of billions. The share of American workers leaving their jobs is close to an all-time low, suggesting little reorganisation of labour. New data indicate that American companies’ investment in “organisational capital”—essentially efforts to improve their routines, processes, culture, supplier relationships and data flows—has been declining as a share of GDP. When AI takes over the economy, you will be able to feel it.然而，到目前为止，几乎没有证据显示出现了 intangible-investment boom，也就是无形投资热潮。像 Palantir 这样的数据组织公司，收入规模是数十亿美元，而不是数千亿美元。美国劳动者离职比例接近历史最低水平，说明劳动力重组很少。新数据显示，美国公司在 “organisational capital”，也就是组织资本上的投资，占 GDP 比重一直在下降。所谓组织资本，本质上是改善企业惯例、流程、文化、供应商关系和数据流的努力。当 AI 真正接管经济时，你一定会感觉到。全篇重点词汇：capex /ˈkæpeks/ 资本开支。capital expenditure; money spent on long-term assets such as equipment, infrastructure or data centres.compute /kəmˈpjuːt/ 算力。computing power used to train or run AI models, especially through specialised chips.extensive margin /ɪkˈstensɪv ˈmɑːdʒɪn/ 广度边际。growth that comes from more firms or people adopting a technology or activity.intensive margin /ɪnˈtensɪv ˈmɑːdʒɪn/ 强度边际。growth that comes from existing users using a technology or activity more deeply or more often.intangible capital /ɪnˈtændʒəbl ˈkæpɪtl/ 无形资本。non-physical assets such as processes, skills, data systems, organisational routines and know-how.经济学商业人评论复制ID☞：JJXREN【经济学商业人评论】翻译爱好者非官方组织,每天几分钟,学习财经英语,为考研英语雅思阅读提供复习材料,免费的双语杂志和经典剪报内容，在知识上和精神上给您学习生活双重享受，内容仅限个人学术研究或者个人翻译作品欣赏用途,请勿转载用于商业。
夜雨聆风
