INTENSIVE READING · INTERACTIVE

财经外刊精读 · 互动版 OpenAI's revenue revision, the chip selloff and the debt behind the boom

2026 年 10 月 9 日 · 星期五 本期：AI 交易 · 营收口径 · 算力融资与私募信贷

30 秒速读

先建立框架，再进原文。默认只给英文，点摘要段落（或右上角「译」）出中文对译，点数字卡展开中文。

A Financial Times report that OpenAI's annualized revenue was closer to $50 billion than the $70 billion investors had been told knocked 1.4% off the Nasdaq 100 and 3.4% off chipmakers. The gap is accounting, not demand.

《金融时报》报道称，OpenAI 的年化营收接近 500 亿美元，而非投资者此前听到的 700 亿美元；这一消息令纳斯达克 100 指数下跌 1.4%、芯片股指数下跌 3.4%。这 200 亿美元的差距来自会计口径，而不是需求出了问题。

$50bn
OpenAI's run rate — the annualized revenue the company reported to investors for the end of September.
OpenAI 的年化营收 —— 公司在 9 月末向投资者报告的年化营收口径。
$20bn
The gap — how much lower that was than the roughly $70 billion figure reported by media outlets late last month.
差距规模 —— 比上月末媒体广泛报道的约 700 亿美元低出的部分。
3.4%
Chip selloff — the slump in a gauge of chipmakers, in a session when the Nasdaq 100 lost 1.4%, its worst day in seven weeks.
芯片股跌幅 —— 芯片股指数当日的跌幅；同日纳斯达克 100 跌 1.4%，为七周来最差单日。
$50bn+
Vendor financing — the debt Broadcom is arranging for the custom AI chip it is developing with OpenAI, with Apollo and Blackstone in talks.
供应商融资 —— 博通正为它与 OpenAI 共同开发的定制 AI 芯片筹措的债务，阿波罗与黑石参与洽谈。
2 in 3
Still rising — the share of S&P 500 members that advanced anyway, as money rotated into defensive sectors.
仍在上涨 —— 标普 500 中依然上涨的成分股比例，资金转向了防御性板块。
~70%
December hike odds — what markets implied after Fed officials Waller and Musalem both signalled more tightening ahead.
12 月加息概率 —— 在美联储官员沃勒与穆萨莱姆双双暗示还将进一步收紧后，市场的隐含定价。

原文精读 · Full Text

34 段（含标题）。点每段右上角的「译」展开中文翻译与结构配色；配图已就近排在相关段落之后。
颜色 = 句子成分，中英同色表示对应
主语 Subject 谓语 / 动词 Verb 宾语 · 表语 Object 从句 · 非谓语 Clause 状语 Adverbial 插入语 · 引述 Attribution
TITLE
The AI trade meets a $20 billion reality check
这轮 AI 交易遇上一场 200 亿美元的「现实核查」
PARA 01
A Financial Times report that OpenAI's annualized revenue was about $20 billion below the levels it had signalled to investors1 knocked the artificial-intelligence trade on Thursday, sending chipmakers sharply lower.
《金融时报》的一篇报道称 OpenAI 的年化营收比它此前向投资者释放的水平低约 200 亿美元打击了周四的人工智能板块，令芯片股大幅走低。
PARA 02
The ChatGPT maker told investors its annualized revenue was approaching $50 billion at the end of September, according to the Financial Times, which cited financial documents shared with backers. That is well short of the $70 billion figure reported by several media outlets late last month.
这家 ChatGPT 的开发商告诉投资者其年化营收在 9 月底已接近 500 亿美元，据《金融时报》援引其与投资方共享的财务文件报道。这远低于上月末多家媒体报道的约 700 亿美元。
Data-center racks: the other side of the AI trade is physical spending — chips, racks and power must be paid for upfront, while the revenue arrives years later.
Data-center racks: the other side of the AI trade is physical spending — chips, racks and power must be paid for upfront, while the revenue arrives years later. 数据中心机柜：AI 交易的另一面是实打实的硬件支出 —— 芯片、机柜与电力，都要先付钱，收入却在数年之后才到。 配图：Unsplash 无版权图库素材（非新闻原图），仅作示意。
PARA 03
The Nasdaq 100 fell 1.4%, its worst day in seven weeks, and a gauge of chipmakers slumped 3.4%. Nvidia, Oracle and the memory makers fell — Nvidia nearly 3%, Oracle 5.8%, and Micron and SK hynix more than 4% each.
纳斯达克 100 指数下跌 1.4%，为七周来最差单日，芯片股指数则重挫 3.4%。英伟达、甲骨文与存储厂商下跌—— 英伟达近 3%、甲骨文 5.8%，美光与 SK 海力士各跌逾 4%。
Red numbers on a trading screen: on 8 October 2026 the Nasdaq 100 fell 1.4%, its worst day in seven weeks, and a gauge of chipmakers slumped 3.4%.
Red numbers on a trading screen: on 8 October 2026 the Nasdaq 100 fell 1.4%, its worst day in seven weeks, and a gauge of chipmakers slumped 3.4%. 交易屏幕上的红色数字：2026 年 10 月 8 日，纳斯达克 100 指数跌 1.4%，为七周来最差单日，芯片股指数重挫 3.4%。 配图：Unsplash 无版权图库素材（非新闻原图），仅作示意。
PARA 04
The $20 billion gap is not a demand shock. It comes from how the two leading AI developers count sales made through cloud partners such as Amazon Web Services and Google Cloud.
这 200 亿美元的差距并不是需求出了问题。它来自两家领先 AI 公司对「经由 AWS、谷歌云这类云伙伴完成的销售」采用的不同计法。
PARA 05
Anthropic includes revenue from cloud-partner sales in its headline figure;3 OpenAI does not. Investors trying to compare the two on the same basis produced the larger $70 billion number by 'grossing up' OpenAI's revenue to match Anthropic's method.
Anthropic计入把经由云伙伴的销售收入记进它的头条数字；OpenAI 不计入。试图在同一口径上比较两者的投资者得出了更大的 700 亿美元 —— 办法是把 OpenAI 的营收「放大」到与 Anthropic 的方法对齐。
One business, three very different numbers OpenAI run-rate reported to investors vs the circulating figure vs rival Anthropic's 2026 target, in $bn OpenAI (company) $50bn OpenAI (grossed up) $70bn Anthropic (target) $100bn
Chart 1 | One business, three very different numbers: OpenAI's reported run rate of about $50 billion, the $70 billion figure that circulated after grossing up cloud-partner sales, and Anthropic's year-end target of roughly $100 billion How to read the bars: all three are annualized run-rate figures, not audited annual revenue, and they are not computed on the same basis. OpenAI's $50 billion is the figure the company gave investors for the end of September; $70 billion is that number grossed up to include cloud-partner sales, the method Anthropic uses. Unit: US$ billion. Sources: Financial Times report of 8 October 2026, as summarised by Reuters/CNA and Investopedia; Anthropic target from investor reporting compiled by the Financial Times and Axios.
PARA 06
Annualized revenue, or 'run rate', multiplies one recent month's revenue by twelve. It has become a popular metric among fast-growing Silicon Valley startups, but it is not the same as reported annual revenue.
年化营收，也就是所谓「run rate」，把最近一个月的营收乘以 12。它已成为硅谷高增长初创公司爱用的指标，但它并不等同于财报口径的年度营收。
PARA 07
If a customer pays $100 for an AI service through a cloud provider, Anthropic's accounting allows it to record the full $100 as top-line revenue,2 with the cloud provider's cut booked as an expense. OpenAI records only its own share of such sales.
如果一位客户通过云厂商为一项 AI 服务支付 100 美元，Anthropic 的会计处理允许它把 100 美元全额记作营业收入，再把云厂商抽走的那部分记为费用。而 OpenAI 只记录这类销售中它自己分成的那部分。
PARA 08
The difference comes down to who each company thinks is selling to the customer — who controls the relationship and is responsible for delivering the product. Both are compliant; they simply describe the same transaction differently.
这一差别归根结底在于两家公司各自认为是谁在向客户销售 ——谁掌握客户关系、谁负责交付产品。两家都合规；只是对同一笔交易给出了不同的描述。
Calculator and statements: the dispute is not about demand but about accounting — the same business can report two revenue figures $20 billion apart, depending on whether sales are counted gross or net.
Calculator and statements: the dispute is not about demand but about accounting — the same business can report two revenue figures $20 billion apart, depending on whether sales are counted gross or net. 算盘与报表：整场争议的分歧点不是需求，而是一个会计问题 —— 同一门生意，用总额法和净额法可以算出两个相差 200 亿美元的营收。 配图：Unsplash 无版权图库素材（非新闻原图），仅作示意。
PARA 09
The distinction matters because run-rate revenue is shorthand for valuing buzzy private companies, where investors talk in terms of enterprise value to sales — a simple price tag divided by revenue.4 If one figure excludes sales a peer includes, that multiple changes without any change in demand.
这个区别很重要因为年化营收是给当红私营公司估值的快捷口径，投资者常说的「企业价值／销售额」倍数，无非是标价除以营收。如果一家的数字不含对方计入的销售，需求没变，这个倍数却会变。
PARA 10
Analysts put the gap in blunter terms. 'The $20 billion is simply the value of sales flowing through AWS, Google and Microsoft Azure,' said Shay Boloor of Futurum Equities. 'It is another reminder that the winner of the AI model race keeps looking like whoever owns the distribution toll booth.'
分析师们把这个差距说得更直白。「这 200 亿美元，无非是经由 AWS、谷歌与微软 Azure 完成的销售价值，」Futurum Equities 的 Shay Boloor 说。「这再次提醒我们：AI 模型竞赛的赢家，看上去始终是握着分销收费站的那一方。」
PARA 11
None of this means demand has stalled: OpenAI's revenue in the third quarter was still growing at more than 70% year on year. The revision reflects a change of definition rather than a change of business.
这一切都不意味着需求停滞：OpenAI 的营收在三季度仍同比增长逾 70%。这次修正改的是口径，不是生意。
PARA 12
Private companies are not bound by the standardised accounting and disclosure rules that public ones must follow, so their headline numbers can be presented in ways that are hard to compare. A listing would force both OpenAI and Anthropic to show what is actually being counted.
私营公司不受上市公司必须遵守的标准化会计与披露规则约束，因此它们的头条数字可以用难以横向比较的方式呈现。上市将迫使OpenAI 与 Anthropic 都说明：到底把什么计进了营收。
PARA 13
Its rival is moving in exactly that direction: Anthropic confidentially filed for a listing in June targeting a valuation of about $2 trillion, with backers expecting annualized revenue to pass $100 billion this year.
它的对手正朝那个方向走：Anthropic已于 6 月秘密递交上市申请，目标估值约 2 万亿美元，投资方预计其年化营收今年将突破 1000 亿美元。
PARA 14
The revenue question lands in the middle of a borrowing boom. Broadcom is arranging more than $50 billion in financing for the custom AI chip it is developing with OpenAI,5 with Apollo Global Management and Blackstone among the lenders in talks, according to the Wall Street Journal.
营收这个问题正落在一场借贷热潮的正中央。博通正筹措逾 500 亿美元资金，用于它与 OpenAI 共同开发的定制 AI 芯片，据《华尔街日报》，参与洽谈的贷款方包括阿波罗全球管理与黑石。
PARA 15
Oracle is negotiating with Apollo and Goldman Sachs to fund hardware for a one-gigawatt facility through a leasing entity kept off its own balance sheet.7 The structure is designed to bridge the gap between when chips must be paid for and when cloud revenue arrives, while limiting Oracle's own borrowing.
甲骨文正在洽谈对象是阿波罗与高盛为一座 1 吉瓦设施的硬件融资，方式是设立一个不并入自身资产负债表的租赁实体。这一结构意在填平「必须先付芯片钱」与「云收入随后才到」之间的时间差，同时控制甲骨文自身的负债。
PARA 16
The Wall Street Journal noted that 'even the Nvidia chips needed for a one-gigawatt data center could cost tens of billions of dollars.' SpaceX, meanwhile, has approached lenders about a $40 billion debt structure to buy Nvidia processors.
《华尔街日报》指出「即便只算一座 1 吉瓦数据中心所需的英伟达芯片，也可能要数百亿美元。」与此同时，SpaceX已就一笔用于购买英伟达处理器的 400 亿美元债务方案接洽贷款方。
PARA 17
These deals show where the money is coming from: as public bond markets and internal cash flows come under pressure from data-center expansion, demand for AI capital is spilling into private credit and leasing through separate entities.6
这些交易说明钱从哪里来：随着公共债市与企业自有现金流因数据中心扩张而承压，AI 领域的资金需求正涌向私募信贷与「借独立实体租赁」这两条通道。
The AI money loopHow capital, chips, debt and revenue now feed each other1Capital flows infunds, private credit2Compute is orderedgigawatts of chips3Suppliers book salesNvidia, Broadcom4Suppliers arrange debtover $50bn for chips5Sites must be poweredgas by truck, queues6Revenue must repay itrun rate under scrutinySix steps: capital flows in, compute is ordered, suppliers book sales, suppliers arrange the debt, sites must be powered, and revenue must repay it all.The dashed line is the feedback loop: whether revenue covers the cost decides whether the next round of capital arrives.
Chart 2 | The AI money loop: how capital, chips, debt and revenue now feed each other Editor's schematic, not a chart from the original article. Six steps trace the loop: capital flows in, gigawatts of compute are ordered, suppliers book the sales, suppliers help arrange the debt, sites must be built and powered, and revenue has to repay it all. The dashed line back to step 1 is the feedback loop: whether revenue covers the cost decides whether the next round of capital arrives.
PARA 18
The physical side is just as awkward. Oracle is trucking natural gas to several data centers to keep building on schedule while pipelines are delayed.
物理层面同样棘手。甲骨文正把天然气用卡车运往多个数据中心以便在管道延期时保持工期。
PARA 19
The workaround is expensive. Energy analysts estimate that trucked gas costs roughly four times as much as gas delivered by pipeline, and powering just 100 megawatts can require a large trailer every 40 minutes.
这个变通办法代价高昂。能源分析师估算卡车运输的天然气成本约为管道输气的四倍，而仅给 100 兆瓦供电，就可能需要每 40 分钟送到一个大型拖车罐。
PARA 20
Oracle also issued a force majeure notice to the New Mexico developer, shielding it from some payments if delays continue. Its free cash flow is negative,8 and is expected to stay that way until more AI data centers are finished and earning revenue.
甲骨文还发出了对墨西哥州开发商的不可抗力通知，若延期持续可免除部分付款。其自由现金流为负，并预计在更多 AI 数据中心建成创收之前一直如此。
PARA 21
The pressure is already reaching the listing pipeline. Firmus Grid, an Australian data-center company backed by Nvidia, is set to postpone its IPO and weigh a private round instead. Analysts called it an ominous sign for other proposed data-center listings.
这种压力已经传导到上市管道。英伟达投资的澳大利亚数据中心公司 Firmus Grid打算推迟 IPO，转而考虑一轮私募融资。分析师称，这对其他待上市的数据中心项目是不祥之兆。
PARA 22
The question investors are asking is simple: can revenue keep pace with the spending? 'Investors are becoming a bit skeptical about how long this huge AI spending can last with borrowing costs rising in a significant way,' said Matt Maley of Miller Tabak.
投资者在问的问题很简单：营收跟得上支出吗？「随着借贷成本显著上升，投资者开始有点怀疑这轮巨额 AI 支出还能持续多久，」Miller Tabak 的 Matt Maley 说。
PARA 23
Yet the selling stayed narrow: more than two-thirds of S&P 500 members advanced while small caps were little changed, as investors rotated toward defensive sectors rather than leaving equities.9
但抛售范围很窄：标普 500 中超过三分之二的成分股上涨小盘股几乎持平，资金是转向防御性板块，而不是撤离股市。
PARA 24
'My biggest surprise is the tech selloff remains concentrated, with more than two-thirds of S&P 500 constituents higher and small caps close to flat,' said Chris Murphy, co-head of derivatives strategy at Susquehanna International Group. 'Strength in staples and low-volatility stocks points to rotation toward defensives rather than broad-based equity selling.'
「最让我意外的是，这轮科技股抛售依然很集中 —— 标普 500 超过三分之二的成分股上涨，小盘股接近持平，」Susquehanna 国际集团衍生品策略联席主管 Chris Murphy 说。「必需消费与低波动股走强，说明这是转向防御，而不是全面抛售股票。」
PARA 25
Macro adds to the pressure. Brent crude jumped more than 4% to settle above $104 a barrel after fresh attacks on tankers, threatening a recent recovery in Middle East flows and keeping inflation expectations alive.
宏观面加重了这种压力。布伦特原油跳涨逾 4%，收在每桶 104 美元上方导火索是油轮再度遇袭 —— 这威胁到中东运输近期的回暖，也让通胀预期继续存在。
PARA 26
Bonds offered a reminder of the cost of money: long-dated Treasury yields retreated from more than two-decade highs after solid demand at a 30-year auction. The rally extended after President Trump said he would hold off attacking Iran until after the November 3 midterm elections.
债市提醒了市场资金的价格：长期美国国债收益率从二十多年高位回落此前 30 年期国债拍卖需求稳健。特朗普总统称 11 月 3 日中期选举前不会攻击伊朗后，涨势进一步延续。
PARA 27
The rate outlook is the other half of the story. Fed Governor Christopher Waller said further increases will probably be needed, and St. Louis Fed President Alberto Musalem said rates should rise over the next six to nine months. Markets put the odds of a December hike at nearly 70%.
利率前景是故事的另一半。美联储理事沃勒表示，很可能还需要进一步加息，圣路易斯联储主席穆萨莱姆称，未来六到九个月利率应当上行。市场认为 12 月加息的概率接近 70%。
PARA 28
Higher rates bite precisely where AI is most exposed. Most of the spending is upfront, in land, chips and power, while the revenue arrives years later which is exactly why so much of the financing is being pushed into private credit and off-balance-sheet entities.
更高的利率正好咬在AI 最脆弱的地方。大部分支出是前置的 —— 土地、芯片与电力，而收入要数年之后才到，这正是如此多融资被推向私募信贷与表外实体的原因。
PARA 29
The chip cycle is already showing its own strains: Samsung Electronics reported record quarterly operating profit of about 107 trillion won, up more than sevenfold, yet its shares fell and the company is cutting fourth-quarter smartphone supply by 20% to 30%, blaming memory prices.
芯片周期已经显露出自身的张力：三星电子公布约 107 万亿韩元的季度营业利润纪录，同比增长逾七倍，但股价仍然下跌，且公司正因存储涨价把四季度手机供应削减 20%–30%。
PARA 30
Chipmakers have risen more than 80% this year, which sets a high bar for the earnings season that begins next week. 'The market's going sideways in anticipation,' said Terry Sandven of US Bank Wealth Management. 'What's equally telling is that year-to-date, the path of least resistance for equities has been up.'
芯片股今年涨幅已超 80%，这为下周开始的财报季设下了很高的门槛。「市场在观望中横盘，」美国合众银行财富管理的 Terry Sandven 说。「同样说明问题的是，今年以来股票最省力的方向一直是向上。」
PARA 31
Three things to watch now: the third-quarter earnings season, whether the Broadcom and Oracle financings actually close, and Anthropic's filings, which will offer the first audited look at the economics of a frontier AI lab.
眼下要盯三件事：三季度财报季；博通与甲骨文的融资能否真正落地；以及 Anthropic 的申报文件 —— 它将首次以经审计的视角，展示一家前沿 AI 实验室的经济账。
PARA 32
For readers, the practical lesson is to ask what is inside a revenue number. Run rate and reported revenue are different things, gross and net are different things, and a $20 billion difference can come down to a definition rather than a slowdown.
对读者来说，实用的经验是：去问一个营收数字里到底装了什么。年化与财报营收不是一回事，毛额与净额也不是一回事，200 亿美元的差距，可能只是定义不同，而不是业务变慢。
PARA 33
The AI trade is no longer a bet on whether the technology works. It is a bet on whether the money adds up — and this week, the market asked for the receipts.
这轮 AI 交易已不再是对「技术行不行」的押注。它而是在押注钱能不能对得上账 ——而本周，市场要凭证。

核心术语 · Glossary

编号与正文上标一一对应，点正文里的数字可跳到对应卡片。默认只显示英文解释（练阅读）；想看中文，点卡片标题。
1annualized revenue/ˈænjuəlaɪzd ˈrevənjuː/ 点击看中文解释 ↓

Annualized revenue takes one month's (or one quarter's) sales and scales it up to a full year — one month times twelve. It is a speedometer, not an odometer: it shows how fast a company is growing right now, not what it actually earned.

Why it matters here ▸ OpenAI said its figure was about $50 billion at the end of September. The $70 billion in circulation was the same business measured with someone else's ruler.

年化营收 —— 把最近一个月（或一个季度）的销售额放大成一整年，最简单的算法是「一个月乘以 12」。它像是车速表而不是里程表：反映的是当下的增长速度，而不是公司真正赚到的钱。

2026≈ 9 GW→2027≈ 15 GW→2028≈ 32 GWNet shortfall = power still missing in 2028 after on-site generation and fuel cells; 34% of demand.
拓展知识：这类指标在硅谷很流行，因为它能让一家成立三年的公司说出一个「像大公司」的数字。但读的时候要问三件事：① 用的是哪个月（旺季还是淡季）；② 是毛额还是净额；③ 有没有把一次性的、还没收到钱的合同算进去。真正上市之后，财报上会同时给出实际年度营收 —— 两个数字往往相差不小，这本身不是造假，只是口径不同。
Extended note: Annualised metrics are popular in Silicon Valley because they let a three-year-old company quote a number that sounds like a big company's. When you meet one, ask three questions: which month was used, is it gross or net, and does it include one-off or not-yet-collected contracts? After an IPO, the audited annual revenue appears alongside it, and the two often differ materially — that is a difference of definition, not necessarily of honesty.
2gross versus net revenue/ɡrəʊs ˈvɜːsəs net ˈrevənjuː/ 点击看中文解释 ↓

When a middleman handles the sale, a company can book either the gross amount (everything the customer paid) or the net amount (only what it keeps). Same customer, same product, two very different revenue lines.

Why it matters here ▸ A $100 sale through a cloud provider: Anthropic can book the full $100 with the platform's cut as an expense; OpenAI books only its own share.

营收的「总额法」与「净额法」—— 当一笔销售经由中间商完成时，公司可以把客户支付的全部金额记作营收（总额法／毛额），也可以只把真正落到自己手上的那部分记作营收（净额法）。同一客户、同一产品，两条营收数字可以差得很远。

拓展知识：会计准则给的判据是「谁是主要责任人」：谁掌握定价权、谁承担存货风险和售后责任，谁就可以按总额法确认收入，否则只能按净额。这解释了为什么电商平台、云厂商、代理分销商的营收常常不可直接比较 —— 一个用总额法，营收会「大」一大截，但利润率会显得很薄；用净额法则相反。看一家公司，永远要连着看营收和利润率。
Extended note: The accounting test is who acts as principal: whoever sets the price and carries inventory and service risk can book gross revenue; otherwise only net. That is why marketplaces, cloud platforms and resellers are hard to compare — a gross reporter shows a much bigger top line and a much thinner margin, and a net reporter the reverse. Always read revenue and margin together.
3cloud-partner sales/klaʊd ˈpɑːtnə seɪlz/ 点击看中文解释 ↓

Sales made through a cloud platform. The customer buys on Amazon Web Services, Google Cloud or Microsoft Azure, the platform handles billing, and the AI company is paid a share.

Why it matters here ▸ Counting them or not is the whole $20 billion question — the two AI labs answer it differently.

云伙伴渠道销售 —— 客户在 AWS、谷歌云或微软 Azure 上购买 AI 服务，由云平台负责开票收款，AI 公司按约定分成。

拓展知识：这类渠道对 AI 公司来说是「量大利薄」：能极快接触到成千上万的企业客户，但平台要抽走一大块，而且客户关系握在平台手里。于是出现了本文那句俏皮话 —— 真正的赢家是「握着分销收费站的人」。对读者的启发是：看到一家 AI 公司增长飞快，先问它的客户到底是谁的客户。
Extended note: For an AI company this channel is high volume, thin margin: instant access to thousands of enterprises, but the platform takes a large cut and owns the customer relationship. Hence the quip about whoever owns the distribution toll booth. The lesson for a reader: when a company grows vertiginously, ask whose customer it really is.
4enterprise value to sales/ˈentəpraɪz ˈvæljuː tə seɪlz/ 点击看中文解释 ↓

A valuation multiple: a company's enterprise value (roughly its price tag, including debt) divided by its revenue. It answers the question 'how many dollars does an investor pay for each dollar of sales?'

Why it matters here ▸ Change the revenue definition and the multiple moves without a single customer changing behaviour — which is why the $50bn/$70bn debate rattled valuations.

企业价值／销售额倍数 —— 用公司的企业价值（大致相当于含债务的「标价」）除以营收，回答「投资者为每一美元销售额付多少钱」。

拓展知识：大多数行业这个倍数在一到三倍之间；高增长软件公司可以到十倍以上，前提是市场相信它未来能赚更多。这类倍数的敏感度极高：分母（营收）口径一变，倍数就变。所以严肃的估值讨论一定先对齐口径，再谈倍数 —— 顺序反了，结论就不可比。
Extended note: Most industries trade at one to three times sales; high-growth software can reach double digits, provided investors believe future profits are coming. Such multiples are extremely sensitive to the denominator: change the revenue definition and the multiple moves. A serious valuation discussion therefore fixes the definition before quoting the multiple — the other order produces numbers that cannot be compared.
5vendor financing/ˈvendə ˈfaɪnænsɪŋ/ 点击看中文解释 ↓

When the supplier helps the customer pay for what it is buying — by arranging loans, leasing the equipment, or guaranteeing the debt.

Why it matters here ▸ Broadcom is arranging more than $50 billion for the chips it is building with OpenAI. You make the chip, then help your customer borrow the money to buy it.

供应商融资 —— 由卖方出钱或出面帮买方解决付款问题：安排贷款、做设备租赁，或者为买方的债务提供担保。

拓展知识：它是双刃的。好处是把订单「钉死」：客户没钱也能买，供应商锁定长期需求。风险是供应商的资产负债表隐性地与客户的偿付能力绑在一起 —— 一旦终端需求或价格不及预期，坏账与减值就出现在卖方账上。历史上有过著名的教训（电信设备、飞机租赁），所以看到「供应商帮客户融资」时，要把它当作一种加杠杆的信号来读。
Extended note: It cuts both ways. It locks in orders: customers who lack cash can still buy, and the supplier secures long-term demand. The risk is that the supplier's balance sheet is now tied to the customer's ability to pay — if end demand or prices disappoint, the write-offs land on the seller. History offers well-known episodes in telecom equipment and aircraft leasing, so read vendor financing as a sign of leverage being added to the chain.
6private credit/ˈpraɪvət ˈkredɪt/ 点击看中文解释 ↓

Lending by private funds rather than banks or public bond markets. The money comes from institutional investors, terms are negotiated in private, and there is no daily market price for the loan.

Why it matters here ▸ With data-center spending far beyond internal cash flow, AI capital demand "is spilling into private credit", according to the Wall Street Journal's account of these deals.

私募信贷 —— 借钱给你的不是银行，也不是公开债市，而是私募基金；资金来自机构投资者，条款私下谈定，贷款本身没有每日报价。

拓展知识：它近年规模暴增，原因是快、灵活、可以承接传统银行不愿做的复杂结构。代价是两件事：不透明（没有公开定价，估值靠模型）与期限错配（资金多来自可赎回的中短期份额，而资产往往是十年期的项目）。当这类资金大量涌入同一个主题（比如 AI 基建）时，行业风险会被集中放大约束在少数机构手里。
Extended note: Private credit has boomed for being fast, flexible and willing to take structures banks avoid. Two costs follow: opacity, because loans have no public price and are valued by models, and duration mismatch, since the money often comes from redeemable shorter-term vehicles while the assets are decade-long projects. When such capital piles into one theme, sector risk becomes concentrated in relatively few hands.
7off-balance-sheet/ˌɒf ˈbæləns ʃiːt/ 点击看中文解释 ↓

Financing structured so the debt does not appear as a liability on the company's own balance sheet — typically through a separate entity that borrows and owns the assets, which the company then leases.

Why it matters here ▸ Oracle would have investors set up a company that buys the chips and leases them to Oracle, bridging the gap "while limiting Oracle's own borrowing".

表外融资 —— 通过架构设计让债务不出现在公司自己的资产负债表上：通常是由一个独立实体去借款、买下资产，公司再向这个实体付租金。

拓展知识：它解决的是「时间差」：芯片要现在付钱，云收入要几年后才到，如果全记在自己账上，资产负债率和信用评级立刻受压。但会计上并非随便就能出表：只要公司承担了实质风险或享有实质收益，规则要求并表。所以读这类结构，要找的是承诺（最低租约、担保、回购义务）—— 承诺比负债难藏。
Extended note: It solves a timing gap: chips must be paid for now while cloud revenue arrives years later, and booking all of it directly would hit leverage and credit ratings at once. Accounting rules, however, only allow it when the company has genuinely transferred the risks and rewards; where it keeps them, consolidation is required. So when you meet such a structure, look for the commitments — minimum leases, guarantees, repurchase obligations. Commitments are harder to hide than debt.
8free cash flow/friː kæʃ fləʊ/ 点击看中文解释 ↓

The cash left after a company pays for everything it needs to keep running and to grow. If it is negative, the business is consuming more cash than it generates and must keep raising money.

Why it matters here ▸ Oracle's free cash flow is negative and is expected to stay that way until more AI data centers are finished and earning revenue.

自由现金流 —— 一家公司扣掉维持经营和继续扩张所需的全部投入之后，手里剩下的现金。如果为负，说明它花掉的比赚到的多，必须持续融资。

拓展知识：它是最难粉饰的科目之一，因为现金进出有银行流水。判断一家重资产公司，先看自由现金流的方向和缺口由谁补：如果是靠股权融资补，会稀释老股东；靠债务补，会抬高利息负担；靠供应商垫资（如本文的供应商融资），则把风险转移给了上游。三种补法的含义完全不同。
Extended note: It is one of the hardest lines to dress up, because cash movements leave a bank trail. For a capital-heavy company, first check the direction of free cash flow and who is funding the gap: equity issuance dilutes existing owners, debt raises the interest burden, and supplier credit shifts risk upstream. The three are far from equivalent.
9rotation toward defensives/rəʊˈteɪʃn təˈwɔːd dɪˈfensɪvz/ 点击看中文解释 ↓

Investors selling one group of shares and buying another — here, moving out of high-growth technology and into sectors whose demand holds up in a downturn, such as staples and utilities.

Why it matters here ▸ Chip stocks fell hard, yet more than two-thirds of S&P 500 members rose and staples and low-volatility shares were strong: rotation, not a broad exit.

向防御性板块轮动 —— 卖掉一类股票、买进另一类：这里是卖出高增长的科技股，买进需求在经济下行时依然稳定的板块，比如必需消费与公用事业。

拓展知识：区分「轮动」和「全面抛售」对判断风险很关键。轮动意味着市场还在，只是重新分配仓位 —— 资金池没变，情绪没有崩；全面抛售则表现为指数普跌、相关性升到接近 1、连防御股也跌。判断方法很简单：看当天下跌家数与上涨家数的比例。本文里「三分之二上涨」正是轮动的证据。
Extended note: Telling rotation apart from a broad selloff matters for risk assessment. Rotation means the market is still there and merely reallocating: the pool of money is intact and sentiment has not broken. A broad selloff shows up as everything falling together, correlations near one, and even defensives down. The easiest test is the ratio of decliners to advancers on the day; in this story, two-thirds of the index rising is exactly that evidence.

读后测验 · Check Your Understanding

3 道题。点选项立即判分；解析需要单独点开。
Q1Why was OpenAI's annualized revenue reported as about $20 billion lower than the earlier $70 billion figure?
正确答案：B
答案在第 4、5 段：两家公司对「经由云伙伴完成的销售」计法不同 —— Anthropic 计入，OpenAI 不计入；$70bn 是投资者为了做同口径比较而把 OpenAI 的数字「放大」后的结果。A 项原文明确否认需求出问题（三季度仍同比增长逾 70%）；C 项原文没有任何监管介入；D 项原文也没有提到合同失效。
Anthropic includes revenue from cloud-partner sales in its headline figure; OpenAI does not. Option B restates the mechanism: the gap is a definition, not lost demand. A contradicts the reported 70%-plus growth, C invents a regulator, and D invents lapsed contracts.
Q2According to the piece, why does the distinction between gross and net revenue matter to investors?
正确答案：C
答案在第 9 段：估值倍数（如企业价值／销售额）以营收为分母，口径一变倍数就变，而需求可以纹丝不动。A 与 C 项原文均无依据 —— 私营公司同样要合规、净额法也绝非上市公司专属；B 项把会计口径与税负混为一谈，原文未提税收。
If one figure excludes sales a peer includes, that multiple changes without any change in demand. Option C is the text's own logic: the multiple is built on revenue, so the definition drives the number. A and D invent rules, and B swaps accounting presentation for tax treatment.
Q3Which statement best captures the argument of the article as a whole?
正确答案：D
答案贯穿全文，尤其在第 22、28、33 段：投资者问的不再是「技术行不行」，而是「营收能否跟上支出」，而加息让这个问句更难回答。A 项与原文相反（三季度仍增长逾 70%）；B 项原文只是提醒私营指标口径不统一，并未主张禁止；C 项被第 23、24 段否定 —— 抛售局限在科技板块，标普 500 超过三分之二成分股仍在上涨。
The AI trade is no longer a bet on whether the technology works. It is a bet on whether the money adds up. Option D is the thesis restated in the closing lines. A contradicts the growth data, B over-reads a caution into a ban, and C is refuted by the narrowness of the selloff.
财经外刊精读 · 互动版 · 2026-10-09 · 由 WorkBuddy 自动生成 · 仅供个人英语精读学习使用，版权归原作者与出版方所有
英文原文：Financial Times 关于 OpenAI 年化营收低于此前口径的报道（细节经 Reuters / CNA、Investopedia、Axios、Finimize 英文稿交叉核对），叠加 Bloomberg Markets Wrap（2026-10-08 盘后）、《华尔街日报》关于 Broadcom 为 OpenAI 定制芯片筹措逾 500 亿美元融资的报道，以及 Bloomberg 关于 Oracle 用卡车运天然气为数据中心供电的报道。原文按精读节奏切分为 34 段，引用的原句文字未作改动。
数字交叉核对：Reuters 报道的摩根士丹利研报（2026-10-05）、The Economic Times（Reuters 通稿）、Financial News 英文版、BusinessKorea（英文全文，含 34%／32 吉瓦／9-15-32 吉瓦的逐年口径，以及英伟达 10 月 5 日收盘 238.90 美元、市值约 5.76 万亿美元）。
背景数据：Gallup 民调（2026-03，七成反对／48% 强烈反对）、Data Center Watch（379 个辖区、约 120 个项目受阻）、PJM Interconnection 容量拍卖结算价（每兆瓦日 28.92 → 329.17 美元）、Bloom Energy 数据中心负载预测（80 → 150 吉瓦）。
配图：Unsplash 无版权图库素材（非新闻原图），图表按公开数据自绘；本材料不构成任何投资建议。
← 全部精读