文 | 定焦One，作者 | 金玙璠，编辑 | 魏佳
截至11月2日，“美股七姐妹”中除英伟达之外的六家，已公布最新业绩。营收几乎全线增长，股价走势却不同。
跌得最狠的，恰恰是增长最快的。美东时间10月28日，谷歌母公司Alphabet公布三季度财报，谷歌云单季收入307亿美元，同比增长103%。然而第二天，股价暴跌12.64%，创下这家公司上市以来最大单日跌幅。
24小时后，亚马逊重演了这一幕。AWS增速39.3%，创近五年新高，盘后股价跳水超过7%。再往前一周，Meta营收增长26%，股价也跌了7%。
另外三家股价相对稳定。微软Azure增速48%，业绩不差，第二天股价微跌0.2%。特斯拉经营利润下降56%，营业利润率只剩2.6%，却因为宣布将全年资本开支从250亿美元压缩到200亿美元，股价反而稳住了。而几乎没在AI基础设施上花钱的苹果，今年股价涨了约26.7%，是“七姐妹”里表现最好的。
“七姐妹”指的是苹果、微软、英伟达、亚马逊、Alphabet、Meta和特斯拉这七家美国科技巨头。英伟达的财季与其余六家不同步，三季度财报要到11月才公布，因此本文主要分析已披露业绩的六家，涉及股价和产业链时仍按七家来讨论。
过去几年，七姐妹的股价走势大体同步，今年彻底变了。截至10月30日，苹果年内涨26.7%，英伟达涨9.7%，Alphabet和特斯拉分别跌10.5%和11%；七家公司占标普500约三分之一的权重，年内加权涨幅只有6.2%，没跑赢大盘。这是自2023年“七姐妹”概念被提出以来，首次集体落于下风。
表面上看，市场在担心AI泡沫，担心巨额投入能否收回。仅上个季度，六家公司的资本开支加起来就超过2000亿美元，其中相当一部分最终流向了英伟达，而英伟达自己上一财年全年的资本开支不过61亿美元。但如果只是害怕泡沫，同样在大举烧钱的微软为什么安然无恙，增速最快的谷歌反而跌得最惨？
要回答这个问题，得先拆解这几家巨头的生意，看清它们各自靠什么赚钱。
01.谁在赚“现在的钱”，谁还在赌“未来”
这六家公司都被贴上“AI概念”的标签，但撑起股价的，是四种不同的逻辑。
微软和亚马逊卖算力，Meta卖广告，苹果卖硬件和服务，特斯拉押注自动驾驶。谷歌特殊一些，两样都占：云和广告。当市场开始用真金白银掂量AI的回报时，模式的优劣就显现出来了。
卖算力，是当下最稳的生意。只要整个行业还在疯狂抢算力，微软、亚马逊、谷歌这三家云厂商未来两三年的收入就大致有了保障。
微软Azure单季收入331.7亿美元，增长48%；AWS收入460亿美元，增长39.3%，年化1840亿美元，是近五年最快；谷歌云收入307亿美元，直接翻了一倍多。
这种增速能否持续，要看手里的订单。微软还没交付的商业合同高达6770亿美元，比去年同期多了73%；谷歌云的积压订单也有5610亿美元，管理层说一半以上会在未来两年内确认为收入；亚马逊积压订单5780亿美元，CEO Jassy在电话会上透露，OpenAI和Anthropic各自追加了上千亿美元的长期采购合同。
云厂商未来几年的收入，很大一部分已经被预定了。眼下的瓶颈在于供给。亚马逊现在平均每14秒就要往数据中心里装一枚AI芯片，产能依然不够。
Meta是另一条路：靠AI把广告卖得更贵。
这个季度营收646亿美元，增长26%，广告展示量多了10%，平均单价涨了14%。管理层在电话会上把增长归功于AI推荐系统的改进。
问题是，广告生意有上限。广告主的总预算是一个有限的池子，AI的价值无非是让广告更精准、更贵，或者增加用户看广告的时长。这个季度单价已经涨了14%，还能不能继续涨？全系产品日活用户超过36亿，增长空间还有多大？当AI投入的边际效益开始递减，广告的增速能否永远跑赢开支的增速？这些都是问题。
苹果几乎没参与这轮军备竞赛。
季度营收1145亿美元，创了纪录，增长12%；全年营收4788亿美元，是公司史上最好成绩。它全年资本开支90亿美元，同比还降了29%。过去两年，市场总批评苹果在AI上落后，同行在大举砸钱，它好像没什么动作。这个季度之后，风向有点变了。当所有人都在为巨额投入和日益紧张的现金流发愁时，“不花钱”反而成了优点。2026年，苹果以27%的涨幅领跑七巨头。
特斯拉是第四种：估值全押在还没兑现的自动驾驶上。
当季营收280亿美元，和上一年持平，经营利润却下滑了56%。卖车这门生意的利润在缩水，市场愿意给它高估值，赌的是自动驾驶和机器人。FSD付费订阅174万，增长67%，北美交付的新车超过75%选装了FSD；Robotaxi也已在奥斯汀正式载客，无监督里程一个季度增长了八倍。
但这些离赚钱还远。特斯拉与SpaceX正在评估的潜在合并，也让它的资本配置多了不少变数。
谷歌跌得最狠，因为两门生意同时遭了质疑。
搜索广告收入增长12%，低于市场预期（约15%），AI搜索的变现效率还不如传统搜索。云业务收入翻倍，代价是单季资本开支514亿美元，自由现金流转为负的91亿美元。加上SpaceX等股权投资的亏损拖累，净利润下滑了24%。财报全面低于预期，市场直接用脚投票。
一位关注美股的投资人Roy对「定焦One」表示，把这六家都当成一个“AI概念”来炒，从一开始就不合理。同样是投AI，有的当年就赚回来，有的要等好几年；有人是用客户的钱建基础设施，有人是用自己的利润赌未来。
02.现金流为负，市场紧张了
这个季度，六家公司的股价涨跌，和自由现金流的状况对应。所谓自由现金流，就是公司赚的钱，刨去建厂房、买设备、买芯片的开销后，手里真正剩下的部分。
自由现金流为负的三家，谷歌、Meta、亚马逊，股价在10月都跌了。自由现金流为正的另外三家，微软、特斯拉、苹果，股价都稳住了。
Roy表示，前两年，市场还愿意为“敢花钱”下注；现在，花钱多的得先证明自己还得起。
变化背后，是这轮AI投入的体量实在太大。微软、亚马逊、谷歌、Meta四家这个季度的资本开支合计2029亿美元，创单季纪录，今年全年计划加起来超过7000亿美元，比去年多了75%。业界预计明年这一规模将达到1.3万亿美元。
钱花到这种程度，市场当然要关注：花出去的钱到底什么时候能回来？高盛在今年一季度的行业回顾里概括过这个转向：对AI的关注，已经从“谁投得最多”，变成“投出去的钱离变回现金还有多远”。
“现金流为负”的三家，处境也不一样。
谷歌和亚马逊的钱都花在了数据中心、芯片和网络设备上。花出去的钱，大概率收得回来，只是需要时间。谷歌这个季度资本开支514亿美元，同比翻了一倍多。底气是谷歌手里5610亿美元的积压订单，管理层说一半以上会在两年内变成收入。亚马逊的逻辑相同：AWS积压订单5780亿美元，OpenAI和Anthropic各自追加了上千亿美元的长期采购。
值得一提的是，微软这个季度的资本开支同样翻倍，达到394亿美元，不过因为赚得多，经营现金流有580亿美元，扣掉开支还剩186亿美元。
Meta的情况不同。它的钱也花在数据中心上，单季资本开支349亿美元，全年计划花1350到1450亿美元，和云厂商是一个量级。区别在于，云厂商的数据中心是替客户建的，建好了客户付钱；而Meta的数据中心是建给自己用的，投入能不能收回来，还是看广告能多赚多少。
2022年Meta大举押注元宇宙，一年之内股价跌了75%，后来靠裁员和砍预算才缓过来。眼下广告业务固然比四年前更赚钱，但这次赌注的体量是当年的数倍。
更值得关注的是，Meta正在从“靠自身利润为AI竞赛提供资金”转向“靠资本市场募资”：长期债务一年内从587亿美元涨到809亿美元，还宣布要设立没有投票权的C类股。标普最近已将Meta的评级展望从“稳定”下调至“负面”，理由正是资本支出和自由经营现金流存在重大不确定性。
另外两家，特斯拉主动收缩，苹果干脆没怎么花钱。
特斯拉把主要资源投向Robotaxi和机器人，但相比此前更大的规划已有所收敛：全年资本开支从250亿美元以上控制到200亿美元，自由现金流回到正的15.3亿美元，手里还备着300亿美元的信贷额度。
苹果几乎没有参与这轮投入。它不建大规模数据中心，AI投入集中在对现有产品和服务的渐进式集成，而非大规模自研训练集群，全年资本开支90亿美元，同比还降了29%，单季经营现金流就有280亿美元。
但苹果也有代价：自研AI进展不顺，今年1月宣布每年花约10亿美元买谷歌的Gemini模型来重建Siri；供应链也被波及，AI服务器抢走了大量高带宽内存产能，普通存储芯片的成本大涨，Mac mini已经提价了200美元。
市场反应其实很清晰：回款路径明确的，市场还愿意等；路径不明朗的，压力最大；投入保守、或者已经开始收缩的，反而成了避风港。
03.座次重排，洗牌才到中场
把时间拉长来看，这轮AI浪潮已经把七姐妹的座次彻底洗牌了，且还没有结束。
回到这轮AI的起点。2022年11月30日，ChatGPT发布。同年年底，按市值排座次：苹果（2万亿美元）、微软（1.8万亿美元）、谷歌（1.1万亿美元）、亚马逊（8500多亿美元）、特斯拉（不到4000亿美元）、英伟达（3600多亿美元）、Meta（3000亿美元出头）。英伟达排第六。
2023年5月，英伟达市值盘中首次达到1万亿美元。两年半后的今天，它一度登顶全球市值第一。
截至今年10月30日，新的座次是：苹果（5.03万亿美元）、英伟达（4.93万亿美元）、微软（3.88万亿美元）、谷歌（3.44万亿美元）、亚马逊（2.6万亿美元）、特斯拉（1.58万亿美元）、Meta（1.57万亿美元）。七家合计，从7万亿美元扩张到了23万亿美元，占标普500的权重从22%升到了33%。
卖芯片的英伟达涨了13倍，从第六名冲到榜首，市值一度超过特斯拉的三倍。三年前，这两家公司的市值还差不多。Meta涨了近五倍，但排名依然垫底。过程中，特斯拉也一度跌到最后一名。
Roy表示，从这一季的财报里，能看出三条正在改写游戏规则的线索。
第一，谁在收钱、谁在借钱。
六家科技巨头的巨额开支里，最大的一部分流向了英伟达。英伟达自己不搞大基建，上一财年全年资本开支仅61亿美元左右。只要行业还在抢算力，它的收入就有保障。但反过来也一样，它的地位本质上建立在另外几家的开支上，一旦下游开始收缩，它马上承压。
另一边，有些云厂商以前是拿利润建机房，现在是发债、发股、借钱来建。这个季度，谷歌发债净入287亿美元，还计划发行847.5亿美元的股票，亚马逊单季发债481亿美元。
更宏观的数据来自路透社的调查：微软、Alphabet、亚马逊、Meta和甲骨文五家科技企业，2026年前9个月经营活动现金流高达创纪录的5429亿美元，但资本支出翻倍至5569亿美元，首次入不敷出。五家公司合计自由现金流由上年同期1435亿美元的盈余转为140亿美元的赤字。
Roy强调，市场已经注意到了循环融资的隐忧：芯片制造商和云厂商利用自身信用，给下游客户提供贷款或担保，这些资金又流回芯片和算力的采购中，容易制造需求旺盛的假象。
第二，折旧才刚开始。
厂房和芯片采购花出去的不是一次性费用。设备一买入就开始计提折旧，逐季摊进利润表。微软单季折旧摊销同比多了44%，往后只会加速；亚马逊过去12个月高达793亿美元，比它一年前同期的净利润（765亿美元）还多。这意味着，设备老化和贬值的速度，快过了它赚钱的速度。摩根大通更是预测：亚马逊2027年的自由现金流可能滑向负1280亿美元。
这背后的原因是，AI芯片和服务器更新换代太快。花100亿买的高端GPU，以为能摊到十年，但可能五年甚至三年后，因为性能落后，就面临资产减值的风险。
投资人Michael Burry最近公开指控巨头们有意拉长AI芯片的折旧年限，据其估算，2026到2028年可能累计虚增约1760亿美元利润。
这一点上，投资最少的苹果和已经收紧开支的特斯拉，相对轻松。
第三，停下来比继续烧更难。
如果AI的商业回报兑现得比预期要慢，能主动刹车、调整节奏的公司可以保住利润；反之，则会被成本拖住。野村跨资产策略师将这一困境概括为“进退两难”：继续加速则自由现金流承压，减速则AI需求叙事失守。
云厂商签的是长期合同，特斯拉的自动驾驶投入不能停，Meta的AI推荐系统是广告业务的地基，也不能停。
苹果是唯一拥有充分余地的。它的投入规模小，可以等技术路线和商业模式更清晰了再精准切入。
下一个验证点是英伟达11月发布的财报。如果它的数据中心收入及指引仍在加速，说明下游还在加码；如果指引下调，就说明下游开始收缩了。
所以，这一季财报季宣告了一个时代的终结：AI基建的“军备竞赛”阶段，正式让位于“算账”阶段。上半场比的是谁胆大、谁下注猛；下半场，拼的是谁家底厚，谁能在折旧开始吞噬利润、债务成本居高不下的时候，拿出让市场信服的投入产出成绩单。
本文由 AI图鉴自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
♪ Man, focus on One, author ♪
As at 2 November, six of the “seven sisters of the United States of America”, with the exception of Weeda, had published their performance updates. Revenues grew almost entirely, while stock prices varied.
The worst falls are the fastest growing. On 28 October, American Eastern Time Alphabet, the Google parent company, published a three-quarter financial newspaper, which generated $30.7 billion in single-season revenues, an increase of 103 per cent over the same period. However, the next day, stock prices fell by 12.64 per cent, the largest single-day decline since the company was listed.
Twenty-four hours later, the Amazon repeated that. AWS grew at a rate of 39.3 per cent, reaching a high of nearly five years, with post-stock price jumps exceeding 7 per cent. The previous week had seen a 26 per cent increase in revenues and a 7 per cent decline in stock prices.
The other three were relatively stable. Microsoft Azure grew by 48 per cent, performing well, and the next day the stock price fell slightly by 0.2 per cent. Tesla lost 56 per cent of its operating profits and had only 2.6 per cent of its operating margins, but stock prices stabilized as a result of the announcement that capital spending would be reduced from $25 billion to $20 billion for the year. The apple, which is barely spent on AI infrastructure, has increased its share price by about 26.7 per cent this year, the best performance of the Seven Sisters.
“Seven Sisters” refers to the seven American giants of technology, Apple, Microsoft, Weidar, Amazon, Alphabet, Meta and Tesla. The financial season of Ingweida was not synchronized with the remaining six, and the quarterly financial statements were not released until November, so the present paper mainly analyses the six cases where performance has been disclosed, with stock prices and industrial chains still being discussed as seven.
Over the past few years, the stock prices of the seven sisters have been largely synchronized and have completely changed this year. By 30 October, apples had risen by 26.7 per cent during the year, British Wida by 9.7 per cent and Alphabet and Tesla by 10.5 per cent and 11 per cent, respectively; seven companies, which accounted for about one third of the standard 500, had a weighted increase of only 6.2 per cent during the year and did not win the grand drive. For the first time since the concept of “seven sisters” was introduced in 2023, it has fallen collectively.
On the surface, the market is worried about the AI bubble and about the recovery of large amounts of inputs. In the last quarter alone, the combined capital expenditure of six companies exceeded $200 billion, a significant portion of which eventually went to Inverda, whose own capital expenditure for the whole of the previous fiscal year was just $6.1 billion. But if it's just fear of bubbles, why is Microsoft, which is also burning money, safe? The fastest-growing Google falls the worst?
To answer that question, it is necessary to dismantle the businesses of the giants and see how they each make money.
Who's making “current money” and who's betting on “the future”
All six companies are labeled “AI concept”, but the price of the stock is based on four different logics.
Microsoft and Amazon, commercials for Meta, hardware and services for apples, automatic driving for Tesla bets. Google is special, both of which are clouds and advertisements. When markets begin to measure AI ' s returns with real money and silver, the pros and cons of the model emerge.
It's the most stable business in the day. As long as the entire industry continues to struggle, the revenues of the three cloud manufacturers, Microsoft, Amazon and Google, are largely secure for the next two to three years.
Microsoft has a single-season revenue of $33.17 billion, an increase of 48 per cent; AWS revenues of $46 billion, an increase of 39.3 per cent, an annualization of $184 billion, the fastest in the last five years; and Google clouds revenues of $30.7 billion, which have more than doubled directly.
The sustainability of this rate depends on the order in hand. Microsoft ' s outstanding commercial contracts amounted to $67.7 billion, 73 per cent more than in the same period last year; Google Cloud ' s backlog of orders was $56.1 billion, with management saying that more than half would be recognized as revenue over the next two years; Amazon ' s backlog of orders was $57.8 billion, and CEO Jassy disclosed at a teleconference that Open AI and Anthropic had each added tens of billions of dollars in long-term procurement contracts.
A significant portion of the revenues of cloud manufacturers in the coming years has been earmarked. The current bottleneck is supply. On average, the Amazon now has to install an AI chip into the data centre for an average of 14 seconds, and the capacity remains inadequate.
Meta is the other way around: selling advertising more expensive by AI.
This quarter received $64.6 billion, an increase of 26 per cent, a 10 per cent increase in advertising and a 14 per cent increase in average unit prices. Management at the conference attributed the increase to improvements in the AI recommended system.
The problem is that the advertising business has a ceiling. The overall budget of the advertiser is a limited pool, and the value of AI is not only to make advertising more precise and expensive, but also to increase the number of hours that users read it. The unit price has increased by 14% this quarter. Can we keep going? More than 3.6 billion daily users of all products, and how much room for growth? When the marginal benefits of AI inputs begin to decline, can the growth in advertising always win the increase in spending? These are problems.
Apple was hardly involved in the arms race.
Quarterly receipts of $114.5 billion, a record increase of 12 per cent, and annual receipts of $4788.8 billion were the best in company history. It spent $9 billion in capital throughout the year, a 29 per cent decline over the same period. Over the past two years, the market has criticized apples for falling behind on AI, and their peers are throwing money at them, and it seems to be doing nothing. After this quarter, the wind has changed. “no money” becomes an advantage when everyone is worried about huge investments and increasingly tight cash flows. In 2026, apples ran seven giants at 27 per cent of the increase.
Tesla is the fourth: the valuations are all on the outstanding auto-driving.
At a quarterly rate of $28 billion, which was the same as the previous year, operating profits declined by 56 per cent. The profits of the car sales business are shrinking, and the market is willing to give it a high valuation and bet on automatic driving and robots. FSDpaid subscribes to 1.74 million, an increase of 67 per cent, and over 75 per cent of new vehicles delivered in North America have selected FSD; Robotaxi has also officially flown in Austin, an eightfold increase in unsupervised mileage over a quarter.
But these are far from making money. The potential merger that Tesla and SpaceX are assessing also makes its capital allocation more variable.
Google fell the hardest because both businesses were questioned at the same time.
Search advertising revenues increased by 12 per cent, below market expectations (about 15 per cent), and AI searches were less efficient than traditional searches. The income from cloud operations doubled at the cost of a single-season capital expenditure of $51.4 billion and free cash flows turned negative to $9.1 billion. Combined with losses on equity investments such as SpaceX, net profits declined by 24 per cent. The financial statements were all below expectations and the market voted directly.
Roy, an investor with interest in US stock, said that it would be unreasonable to fire all six as an "AI concept" from the start. The same goes for AI, some of which earns money that year, some of which wait years; some build infrastructure with customer money and others with their own profits.
Cash flows are negative, markets are tight.
Six companies experienced stock price increases and declines this quarter, corresponding to the situation of free cash flows. The so-called free cash flow is the real rest of the money that companies earn to build a plant, buy equipment, buy chips.
The three households with negative free cash flows, Google, Meta, Amazon, fell in October. The other three, Microsoft, Tesla, Apple, had stable share prices.
Roy stated that the market had been willing to bet on “bold money” for the first two years; now, it had to prove that it could afford to spend more.
Behind the change, this round of AI input is really too big. The combined capital expenditure of Microsoft, Amazon, Google and Meta this quarter was $202.9 billion, a single-season record of over $700 billion, 75 per cent more than last year, this year. The industry expects to reach $1.3 trillion next year.
To the extent that money is spent, of course the market is concerned: when will the money spent come back? Goldman Sachs' industry review of the first quarter of this year summarized the shift: the focus on AI has gone from “who voted most” to “how far away is the money going back to cash”.
The situation was different for the three “negative cash flows”.
Google and Amazon spending on data centres, chips and network equipment. Spending money, probably getting it back, just takes time. Google spent $51.4 billion this quarter, more than double the same year. The bottom line was a $561 billion backlog of orders from Valley singers, with management saying that more than half of them would become revenue in two years. The Amazon has the same logic: AWS has $578 billion in backlog orders and Open AI and Anthropic have each added hundreds of billions of dollars in long-term procurement.
It is worth mentioning that Microsoft also doubled its capital expenditure this quarter to $39.4 billion, although, with a much larger profit, the operating cash flow stood at $58 billion, leaving $18.6 billion to be deducted.
The situation in Meta is different. It also spent $34.9 billion on data centres, with a single-season capital expenditure of between $135 billion and $145 billion planned for the year, and a cloud manufacturer as a measure. The difference is that the cloud manufacturer's data centre is built for customers, and the customer is paid for; while Meta's data centre is built for its own use, whether the input can be recovered or how much the advertising can earn.
In 2022, Meta made a big bet in the universe, and in one year the stock price fell by 75 per cent, which was later slowed down by layoffs and budget cuts. Advertising is now more profitable than it was four years ago, but this bet is several times the size of the year.
It is even more noteworthy that Meta is moving from “financing the AI competition on its own profits” to “financing from the capital market”: long-term debt rose from $58.7 billion to $80.9 billion a year and the establishment of a class C unit without voting rights was announced. The markers have recently downgraded Meta's rating outlook from “stable” to “negative”, precisely because of the significant uncertainty surrounding capital expenditure and free-run cash flows.
And the other two, Tesla, went out of her way. Apple didn't pay much.
Tesla invests its main resources in Robotaxi and robots, but the larger plans have been reduced: Capital spending was contained throughout the year from over $25 billion to $20 billion, free cash flows returned to a positive $1.53 billion, with $30 billion in credit quota in their possession.
Apples are hardly involved in this round of inputs. It does not build large-scale data centres, and AI inputs are concentrated on progressive integration of existing products and services rather than large self-study training clusters, with capital expenditure of $9 billion per year, a 29 per cent decline over the same period, and a single-season cash flow of $28 billion.
But apples also come at a cost: the self-study of AI is not progressing well, and in January this year it was announced that it would cost about $1 billion a year to buy the Gemini model of Google to rebuild Siri; the supply chain has also been compromised, the AIS server has taken away a large amount of high bandwidth memory capacity, the cost of common storage chips has increased dramatically, and Mac Mini has raised $200.
The market response is clear: the return path is clear and the market is willing to wait; the path is unclear and the pressure is greatest; and the input is conservative or has begun to contract, instead it becomes a safe haven.
03. Rescheduled, cards were washed to midfield.
As long as it lasts, this wave of AI has completely washed the seats of the seven sisters and is not over.
Back to the beginning of this round of AI. ChatGPTpublish, 30 November 2022. At the end of the year, seats were organized according to market value: Apple ($2 trillion), Microsoft ($1.8 trillion), Google ($1.1 trillion), Amazon (over $850 billion), Tesla (less than $400 billion), Ingweida ($360 billion) and Meta ($300 billion). In Weeda is number six.
In May 2023, for the first time, the market value book of British Wida reached $1 trillion. Today, two and a half years later, it once reached the highest global market value.
As of 30 October this year, the new seats are: Apple ($5.3 trillion), Ingweida ($4.93 trillion), Microsoft ($3.88 trillion), Google ($3.44 trillion), Amazon ($2.6 trillion), Tesla ($1.58 trillion) and Meta ($1.57 trillion). In total, the seven countries expanded from $7 trillion to $23 trillion, and the weight of the 500 quota increased from 22 per cent to 33 per cent.
The price of the chip sold by the British West is 13 times higher, from sixth to top of the list, which was once three times higher than Tesla's. Three years ago, the two companies had the same market value. Meta has increased nearly fivefold, but still ranks are at the bottom. In the process, Tesla fell to the last.
Roy stated that three clues to the rewriting of the rules of the game could be seen in the financial statements of the season.
First, who's collecting and who's borrowing.
The largest share of the huge expenditures of the six technology giants went to Inverda. The capital expenditure for the previous fiscal year was only about $6.1 billion. As long as the industry continues to struggle, its income is guaranteed. However, the same is true of the fact that its status is essentially based on the expenses of several other households, and it is under pressure as soon as it begins to contract downstream.
On the other hand, some cloud manufacturers, which used to build machines on a profit basis, are now paying debts, issuing shares and borrowing money. Google netted $28.7 billion during the quarter and planned to issue $84.75 billion in equities and $48.1 billion in single-season debt in Amazonia.
More macro-data are derived from Reuters surveys: Microsoft, Alphabet, Amazon, Meta and Oracle five technology enterprises, with a record cash flow of $54.29 billion in operating activities in the first nine months of 2026, but capital expenditure doubled to $556.9 billion, a first shortfall. The combined free cash flow of five companies shifted from a surplus of $143.5 billion over the same period of the previous year to a deficit of $14 billion.
Roy stressed that the market had taken note of the fear of revolving finance: chip manufacturers and cloud manufacturers used their credit to provide loans or guarantees to downstream customers, and that the flow of those funds back into chip and calculator purchases tended to create a false picture of demand.
Second, depreciation has only just begun.
The purchase of plant and chips is not a one-time cost. Depreciation is applied as soon as the equipment is purchased and the profit statement is distributed on a quarterly basis. Microsoft ' s single-season depreciation amortization was 44 per cent higher than that of the same year and would only accelerate later; the Amazon had reached $79.3 billion over the past 12 months, more than its net profit for the same period a year earlier ($76.5 billion). This means that the equipment is ageing and devalued faster than it makes money. Morgan Chase, for its part, predicted that free cash flows in the Amazon in 2027 could slide to negative $128 billion.
The reason behind this is that the AI chips and servers are being updated too quickly. High-end GPUs, which cost $10 billion, are thought to be available for 10 years, but may be five or even three years later, owing to poor performance, at risk of asset impairment.
Investor Michael Burry has recently publicly accused the giants of intentionally stretching the depreciation years of the AI chip, and it is estimated that there may be a cumulative gain of approximately $176 billion between 2026 and 2028.
At this point, the least invested apples and the already tight Tesla are relatively easy.
Third, stopping is harder than keeping burning.
If the business returns of AI were to materialize slower than expected, companies able to brake and adjust their pace would be able to keep profits; otherwise, they would be held back by costs. Nomura's cross-assets strategist summed up the dilemma as a “difficult”: continuing to accelerate, free-flow pressure, slowing down, and the fall in demand for AI.
The cloud factory is under a long-term contract, the autopilot input of Tesla cannot be stopped, and Meta ' s AI referral system is the foundation of the advertising business and cannot be stopped.
Apples are the only ones with sufficient room. Its input is small enough to be precise when the technical routes and business models are clearer.
The next validation point is the financial report of November publish by Yvette. If its data centre revenues and guidelines continue to accelerate, it means that the downstreams are still being added; if the guidelines are revised downwards, it means that the downstreams are starting to contract.
This season has thus declared the end of an era: the “arms race” phase of AI's capital, which officially gave way to the “counting” phase. In the first half of the game, who is bold and who is betting; in the second half, who is in the middle of the game, who is in the middle of the game and who is able to produce a paper of output that convinces the market when depreciation begins to devour profits and debt costs are high.
Auto-fetched and translated by AI Atlas. All rights belong to the original author.