上周，《金融时报》报道称OpenAI的营收接近500亿美元，这一数字虽然庞大，却远低于前一周Axios报道的700亿美元，导致科技股应声下跌；随后博通（Broadcom）总裁陈福阳将市场比作赌场，引发科技股持续下挫。正如20世纪80年代并购潮带来的商誉、90年代和21世纪初的无形资产以及2010年代的股权激励一样，每一次投资热潮都会伴随一堆令人费解的财务缩写和调整后数据。如今，OpenAI和Anthropic报告的所谓营收实际上是硅谷模糊指标“年度经常性收入”（ARR，即将月销售额乘以12）的变体，且两家公司的计算口径截然不同：Anthropic将通过云提供商实现的模型销售额全部计入，而OpenAI则扣除了合作伙伴的分成。10月14日，Anthropic向投资者透露其第三季度营收为170亿美元，并按自身调整后的标准实现了盈利，但由于外界不清楚这与其7月份650亿美元的ARR相比增长了多少，投资者在两家公司上市前如同盲人摸象。投资公司Coatue甚至提议按“预训练EBITDA”（即扣除高昂的模型训练成本之前的息税折旧摊销前利润）的倍数来为AI实验室估值，这让人联想到石油行业的EBITDAX或WeWork臭名昭著的“社区调整后EBITDA”。
公开市场的数据同样模糊不清。去年市场还在争论大型科技公司是否因过慢减记计算资产价值而夸大了盈利，今年则面临杠杆率和资本回报率两大黑箱：亚马逊、谷歌、Meta和微软做出的2.4万亿美元表外承诺已超过它们从债券市场的借款总额；微软在4月称其AI业务的“年度经常性收入运行率”（ARRR）为370亿美元，亚马逊在7月称其ARRR为250亿美元。然而，高盛集团的研究人员指出，如果云巨头要在今明两年的数据中心投资上获得10%至20%的回报率，且AI开发商要获得30%的利润率，那么每年需要产生约1万亿美元的营收，这意味着企业从AI中获得的生产率提升价值必须远超这一数字，但目前企业老板们根本无法衡量其AI支出的回报。
更严重的是，衡量美国企业活力和宏观经济的基本指标也已失真。科技巨头在数据中心上的支出立即推高了半导体和工业企业的利润，但在自身财务报表中却作为折旧缓慢体现，加上对非上市AI实验室持股的一次性估值上调，短期内人为夸大了整体盈利：第二季度标普500指数未调整的盈利增长率高达惊人的50%，若剔除Alphabet和亚马逊对私人投资的账面价值上调，这一比例则降至32%。在宏观层面，存储芯片价格飙升导致部分软件产品的通胀被错误高估，而智库Epoch AI的报告指出，由于英伟达跨境销售的统计方式问题，过去一年美国的GDP增长率被足足低估了0.3个百分点。与此同时，美国数据收集机构资金匮乏，且自21世纪初安然丑闻以来负责审查企业的会计监管机构预算遭到大幅削减——该机构近期给四大审计公司打出了最高分，这恐怕又是另一组好得令人难以置信的数字。
Last week, tech stocks whipsawed after the Financial Times reported OpenAI’s revenue was nearing $50bn—well below the $70bn reported by Axios a week earlier—and Broadcom boss Hock Tan compared the market to a casino. Like 1980s merger goodwill, 1990s intangible assets, and 2010s stock-based compensation, the AI boom relies on murky metrics, notably “annual recurring revenue” (ARR), which simply multiplies monthly sales by 12. Even definitions diverge: Anthropic counts gross model sales made through cloud providers, whereas OpenAI deducts partners’ cuts. When Anthropic reportedly told investors on October 14th that third-quarter revenue hit $17bn and reached adjusted profitability, investors could not gauge growth against July’s $65bn ARR figure. Meanwhile, investment firm Coatue suggests valuing AI labs on “pre-training EBITDA”—ignoring massive model-training costs—echoing the oil industry’s EBITDAX or WeWork’s infamous “community-adjusted EBITDA”.
Public markets offer little more clarity. Concerns over slow depreciation of computing assets have given way to questions over leverage and returns: Amazon, Google, Meta, and Microsoft hold $2.4trn in off-balance-sheet commitments, exceeding their bond-market debt, while Microsoft’s $37bn “annual revenue run rate” in April and Amazon’s $25bn ARRR in July remain black boxes. Crucially, Goldman Sachs researchers estimate that for cloud giants to earn a 10-20% return on 2026-27 data-centre investments and AI developers to make a 30% margin, the industry needs $1trn in annual revenue, requiring even larger corporate productivity gains that bosses currently cannot measure.
Even basic gauges of corporate and economic health are distorted. Tech giants’ data-centre spending immediately boosts semiconductor profits while depreciating slowly on their own books, and write-ups of stakes in private AI labs further inflate results: S&P 500 unadjusted second-quarter earnings growth was 50%, falling to 32% when excluding Alphabet’s and Amazon’s private-investment write-ups. Macro statistics are similarly skewed: soaring memory-chip prices overstate software inflation, and think-tank Epoch AI estimates America’s GDP growth was undercounted by 0.3 percentage points over the past year due to how Nvidia’s cross-border sales are measured. Meanwhile, American statistical agencies are starved of funds, and the post-Enron accounting watchdog has had its budget slashed even as it awards top marks to the big four audit firms.
Source: The AI boom is threatened by an epidemic of unreliable numbers
Subtitle: Investors can no longer put their trust in data
Dateline: Oct 15th 2026\n