1
2
Business | Schumpeter
商业 | 熊彼特专栏
英文部分选自经济学人20261017期商业版块
Business | Schumpeter
商业 | 熊彼特专栏
The AI boom is threatened by an epidemic of unreliable numbers
AI热潮正面临不可靠数据泛滥的威胁
Investors can no longer put their trust in data
投资者再也无法信任各类数据
Last week the Financial Times wrote that OpenAI’s revenue was nearing $50bn. A massive number, though massively lower than the $70bn reported by Axios the week before. Tech stocks fell. Journalists were not on top of their figures, cried investors. Tech stocks rose. If investors could be swayed so easily, they were not on top of their investments, cried Hock Tan. The boss of Broadcom, a chipmaker, had compared markets to a casino, wrote Bloomberg. Tech stocks fell—and have continued to do so. Markets may have been flying. But when it comes to the numbers that matter, they have been flying blind.
上周，《金融时报》报道，OpenAI营收接近500亿美元。这个数字固然庞大，但还是远低于一周前Axios新闻报道的700亿美元，科技股随即下跌。投资者抱怨记者没有把数据核实清楚，随后科技股又反弹上涨。芯片厂商博通首席执行官陈福阳表示，如果投资者如此容易受消息左右，恐怕他们自己也没搞清楚投的是什么。彭博社报道，他还把市场比作赌场，于是科技股再度下跌，并且持续走低。市场此前或许一路高飞，但面对关键数字，往往是一路盲飞。
Radical change has always been hard to count. The merger wave of the 1980s left corporate balance-sheets full of goodwill, that problematic wedge between the book value of a company’s assets and what it was bought for. Measuring other intangible assets like brands and patents was an even bigger head-scratcher in the 1990s and 2000s. The rise of stock-based compensation in the 2010s further obscured the picture of corporate health, particularly for tech firms. Every investment boom brings with it a rogues’ gallery of acronyms, adjusted numbers and valuation techniques, sometimes designed to elucidate and other times to obscure.
颠覆性变革向来难以量化统计。20世纪80年代的并购浪潮，让企业资产负债表堆满商誉，这一颇有争议的项目，反映的是企业资产账面价值和其收购价格之间的差额。到了20世纪90年代和21世纪头十年，如何衡量品牌和专利等其他无形资产更是一大难题。2010年代股权激励的普及，进一步模糊了企业真实经营状况，科技企业尤甚。每一轮投资热潮都会冒出一大堆五花八门的缩写、调整后的数据和估值方法；这些工具有时是为了厘清业务，有时却是为了掩盖真相。
One problem today is that the reported revenues of OpenAI and Anthropic are not really revenues at all. Rather, they are vague riffs on a famously vague Silicon Valley metric: ARR. The “A” is for “annual”, in the sense that it bears some relationship to the Gregorian calendar (monthly sales are multiplied by 12). The first “R” traditionally stands for “recurring”, which means only that the firm hopes they will continue. And “revenue”, the second “R”? Anthropic counts sales of its models made through cloud providers, whereas OpenAI does so only after subtracting its partners’ cut.
当下的一个问题是：OpenAI和Anthropic对外公布的营收，根本算不上真正意义上的营收。它们本质上只是硅谷一个素来模糊的指标ARR的变体解读。A代表年度（annual），意思是和公历年度挂钩，但也只是将月度销售额直接乘12；第一个R按惯例代表持续性（recurring），实际上仅仅是企业希望这笔收入可以持续；第二个R，也就是营收（revenue）该如何理解呢？Anthropic会把通过云服务商渠道卖出模型产生的全部销售额计入统计；而OpenAI要先扣除合作方分成之后，才将这部分收入计算在内。
注释：
AI估值密码大揭秘：ARR如何让科技巨头身价暴涨？
https://www.xueqiu.com/4656299966/338632208
Until the pair list, investors are trapped in a strange interregnum where the faintest suggestion of news about numbers that are themselves merely suggestive shakes the market. On October 14th Anthropic reportedly told investors that it had revenues of $17bn during the third quarter and that, on its own adjusted measure, it was profitable. Still, it is unclear how much sales had grown since July, when ARR was $65bn. New numbers are good for little when investors do not know what to compare them with.
在这两家企业上市之前，投资者只能被困在一个奇怪的信息真空期：即便只是关于某些数字的一丝风吹草动都足以撼动市场，而这些数字本身也不过是一些暗示性的指标。据报道，10月14日，Anthropic向投资者透露，公司第三季度的营收达170亿美元，按公司自行调整的口径计算已实现盈利。然而，7月公布的ARR为650亿美元，第三季度的销售额究竟比那时增长了多少，仍不清楚。如果投资者不知道拿什么来比较，再新的数字也没有多少价值。
Wacky financials deserve wacky valuations. Coatue, an investment firm, reckons that the labs should be worth some multiple of their pre-training EBITDA (that is, their earnings before interest, tax, depreciation and amortisation—and before the cost of training their models, an enormous expense). This might remind investors of the oil industry (where EBITDAX, earnings before exploration, is common); or private equity (where inflating EBITDA in order to borrow more has been raised to something of an art form); or even WeWork, the office-leasing company whose creative financial reporting (“community-adjusted EBITDA”) earned it a deserved place in the pantheon of bad accounting. Perhaps they could live even more dangerously. How about price-to-token-price ratio? Or existentially-adjusted earnings per share?
古怪的财务数据，自然得配得上古怪的估值方法。投资公司Coatue认为，这些人工智能实验室的估值应当按照其预训练EBITDA（即息税折旧摊销前利润）的特定倍数来计算。这里的EBITDA甚至还要剔除训练模型的成本，而这一笔开支往往庞大。这种算法或许会让投资者联想到常用EBITDAX的石油行业，即在息税折旧摊销前利润的基础上，剔除勘探费用；或者私募股权行业，业内人士为了借到更多钱，几乎把抬高EBITDA变成了一门艺术；再或是办公空间租赁公司WeWork，该公司曾创造性地提出“经社群调整后的EBITDA”这一财务指标，让它理所当然地跻身会计乱象的反面典型之列。或许这些AI公司还可以玩得更大胆一些，来个“词元价格市盈率”，或者“存在性风险调整后的每股收益”？
Public markets offer more clarity. But not much more. Last year the accounting question du jour was whether big tech firms were writing down the value of their computing assets too slowly, thus overstating their earnings. This issue was not so much resolved as replaced by two bigger ones. How leveraged are the giants? The $2.4trn of off-balance-sheet commitments made by Amazon, Google, Meta and Microsoft exceeds what they have borrowed from bond markets. And how to measure their return on this capital? In April Microsoft said its AI business had an “annual revenue run rate” of $37bn. In July Amazon put its ARRR at $25bn. Despite the additional “r” these figures, too, are black boxes.
公开市场相对更清楚一些，但也清楚不到哪儿去。去年，会计界最关心的是，大型科技公司是否过于缓慢地调低计算设备的账面价值，从而夸大了自身账面利润？这个问题与其说得到了解决，不如说是被两个更大的问题盖过了。首先，这些科技巨头的融资杠杆到底多大？亚马逊、谷歌、Meta和微软作出的表外承诺总额高达2.4万亿美元，甚至超过了它们通过债券市场借入的资金。其次，又该如何衡量这些资本的投资回报？今年4月，微软称其人工智能业务的“年化营收运行率”已达到370亿美元。7月，亚马逊则公布，其所谓的“年化营收运行率”已达到250亿美元。尽管比上文提到的ARR多了一个“R”，但这些数字仍然像黑箱一样。
Yet perhaps nothing matters quite as much to the AI trade than where this revenue is actually coming from. According to researchers at Goldman Sachs, a bank, if the cloud giants are to earn a return between 10% and 20% on what they invest in data centres this year and next, and AI developers are to earn a 30% margin, some $1trn of annual revenue is required. That means the value of productivity gains companies extract from the technology should far exceed that. Therein lies yet another measurement problem. Ask a boss how he measures his returns from spending on the technology and prepare to be met with a blank stare.
然而，对人工智能投资热潮而言，或许没有什么比厘清这些收入的来源更为重要。高盛研究人员测算，如果云计算巨头要从今明两年对数据中心的投资中获得10%至20%的回报率，同时AI开发商要实现30%的利润率，就需要约1万亿美元的年营收。这意味着，企业利用这项技术提高生产率所创造的价值，应当远超这一数额。可这又该如何衡量？问一位企业负责人怎样衡量AI技术的投资回报，很可能只会换来茫然的目光。
ARRrrrrr-gh!
Even the most basic measures of corporate America’s vitality have been rendered unintelligible. The spending of tech giants on data centres immediately lifts the profits of semiconductor and industrial firms, but flows through their own financial statements far more slowly as depreciation, inflating aggregate earnings in the short term. One-off adjustments for public firms’ stakes in private labs fatten them further. During the second quarter the unadjusted earnings growth of the S&P 500 index was an astonishing 50%. Exclude the write-ups of private investments by Alphabet and Amazon and it falls to a still-remarkable 32%.
ARR之困
如今，就连衡量美国企业整体经营状况最基本的指标，也让人看不懂了。科技巨头在数据中心上的支出，会立即推高半导体企业和工业企业的利润；但这些支出在科技巨头自身的财务报表中，却要以折旧的形式缓慢计入费用，短期内抬高企业整体的账面盈利。同时，上市公司对所持非上市人工智能实验室股权作出的一次性估值调整，又进一步推高了利润。今年第二季度，标普500指数成分股公司未经调整利润增速达到惊人的50%。剔除Alphabet和亚马逊所持非上市公司股权投资的估值上调收益后，利润增速降至32%，但仍相当可观。
Measuring the impact of AI on the economy has become similarly fraught. Soaring prices for memory chips mean that inflation has been incorrectly overstated for some software products. One widely discussed report by Epoch AI, a think-tank, suggests that GDP growth in America has been undercounted by a full 0.3 percentage points during the past year because of the way Nvidia’s cross-border sales are measured. At the same time, America’s data-collection bodies have been starved of funding and are declining in competence. Another casualty of deregulation is the accounting watchdog that has inspected firms since the Enron scandal in the 2000s, whose budget has been slashed. The big four audit firms have been given top marks in its recent review—one last set of numbers that might prove too good to be true.
评估人工智能对经济的影响，同样变得困难重重。存储芯片价格飙升，导致部分软件产品的价格涨幅被错误高估。智库 Epoch AI一份引发广泛讨论的报告指出，由于英伟达跨境销售的统计方式存在问题，过去一年美国GDP增速被低估了整整0.3个百分点。与此同时，美国的数据统计机构长期经费不足，专业能力也在下降。放松监管还伤及一直负责会计审计的监管机构。自21世纪初安然（Enron）财务丑闻以来，其预算已遭大幅削减。在最近一轮检查中，四大会计师事务所均获得了优异评价，这或许也是一组好得令人难以置信的数字。
翻译组：
Yuqing，已识乾坤大，犹怜草木青
Mai，做有意义的事，做有意思的人
Humi，在躺平、侧卧和睡梦中寻找诗和远方
审校组：
Lucian，灵魂在杰作中冒险
Rex, the one who widens your horizon
---END---