这一周的财报季，微软、谷歌、亚马逊三大云巨头相继放榜，出现了多年来罕见的同框画面：云厂商的算力被客户排着队抢购，订单簿已经写到了2028年；而同一家公司的报表上，却赫然挂着多年未见的负自由现金流。
需求旺盛到“售罄”，现金紧张到“转负”，两组信号恰好说明这轮AI基建竞赛已经进入一个新阶段。
云业务全线加速，订单排到三年后
先看亚马逊。作为全球最大云厂商，AWS三季度收入459.8亿美元，同比增长39.3%，创下近五年最快增速，年化收入规模达到1840亿美元，增速已连续六个季度抬升。支撑这一数字的，是账面上5780亿美元的积压订单，同比呈三位数增长，较二季度末的4960亿美元又增加了800多亿美元。CEO Andy Jassy此前接受访谈时说得直白：2026年的产能已经售罄，2027年大部分被预订，2028年也有相当一部分拿到了客户承诺。
紧随其后的微软，数字同样超出预期。2027财年第一季度，公司营收922.8亿美元，同比增长19%，净利润366.3亿美元，增长约32%。本季微软首次单独披露Azure的绝对收入：单季331.65亿美元，同比增长48%，增速在千亿美元量级上不降反升，商业剩余履约义务（RPO）同比增长73%至6770亿美元。微软还透露单季新增了24个数据中心，这种交付节奏在历史上并不多见。
相比之下，谷歌云的增速最为陡峭。Alphabet三季度谷歌云收入307.1亿美元，同比增长超过一倍，积压订单达到5610亿美元。更难得的是盈利质量：谷歌云营业利润率达到36.2%的历史新高，营业利润超过上年同期的三倍。但Sundar Pichai的另一句话更值得注意：客户现在想锁定2027年的算力，“极其困难”。谷歌甚至一度租用外部算力，包括按月付费使用SpaceX的算力资源，来填补自己的供给缺口。
三家公司给出的口径不同，指向却是同一件事：AI算力的需求曲线远比供给曲线陡峭。云厂商过去担心的是产能建成后无人使用，如今担心的是产能建成后接不住订单。积压订单不是意向，是已签署的合同，它们把未来数年的收入提前锁定，也把交付压力提前摆上了台面。
资本开支竞赛下，现金流第一次告急
要把这些订单接下来，代价同样不小。亚马逊把2026年资本开支指引维持在约2200亿美元，仅三季度物业及设备支出就达到773亿美元，超出市场预期三成以上；Alphabet全年资本开支指引为1950亿至2050亿美元；Meta这家以广告为基本盘的公司，是本轮AI基建最激进的买家之一，将资本开支下限上调至1350亿美元，并明确2027年将继续积极投资；微软在财报后随即启动了至少250亿美元的发债计划，获4.6倍认购，这家软件巨头借此近八年来首次重返美国投资级债券市场。
开支先行、收入后置，现金流量表的压力随之而来。亚马逊过去12个月自由现金流净流出423.6亿美元，而一年前同期还是净流入；Alphabet自由现金流为负90.7亿美元，连续第二个季度为负；Meta净流出21.2亿美元，是2012年以来的首次；四家中唯有微软仍保有186亿美元的正自由现金流。三位“现金牛”在同一季度集体失血，这在云计算诞生以来还是头一回。
有意思的是，投资者并未对负自由现金流一票否决。微软CFO Amy Hood表示将在“时机合适时”继续融资；Meta则设立了无投票权的C类普通股，为后续AI投入预留了理论上约7000亿美元的融资空间。管理层显然已经预判到，这轮基建周期的资金需求会远超经营现金流的承载能力，债务与股权工具正在成为标配。信用市场也在做出选择：微软的AAA债券订单蜂拥而至，而评级更低的科技债发行成本明显抬升，Meta的五年期信用违约掉期成本一度升至该合约有记录以来的最高水平，资金正在向确定性集中。
冰火两重天的市场判决
从10月21日盘后Meta率先放榜，到28日微软、谷歌同日放榜，再到29日亚马逊压轴，订单与现金流的故事讲完后，市场逐一给出了定价，而市场的判决相当悬殊。微软业绩全面超预期，盘后一度上涨6%，虽然次日随科技股整体回调小幅收跌0.16%，但整体平稳；Meta营收利润双超预期，AI购物智能体Muse上线六周、北美用户破千万是最大亮点，盘后一度涨超6%，但次日受巨额法律罚单和现金流转负拖累，收跌约7%，成为四家中唯一业绩超预期仍遭抛售的。
Alphabet则是另一番景象。营收、每股收益、广告收入三项核心指标集体不及预期，搜索业务增速从一季度的19%一路放缓至12%，管理层解释，汇率是最大单一拖累，叠加世界杯广告支出在二季度集中释放，AI模式查询激增但变现率尚低；再加上连续为负的自由现金流和87亿美元的非现金股权投资亏损，股价次日重挫12.64%，创下上市以来最大单日跌幅，并拖累纳指当日下跌1.38%。亚马逊尽管AWS再超预期，但线上零售仅增长2%、四季度指引中值偏低，加上持续为负的自由现金流，盘后一度跌超7%。
涨跌之间的分野，指向同一个结论：市场不再为“AI投入”本身付钱，只为“AI回报”付钱。市场的计较并非苛责：本轮AI收入里，相当一部分来自云厂商与AI公司之间的交叉订单，含金量需要打折看待；而GPU与数据中心的折旧年限若被高估，未来的利润表还可能承受额外压力。云业务增速高不再是护身符，投资者开始逐项检查：订单能否转化为收入，现金消耗能否被利润覆盖，传统业务的基本盘是否被AI侵蚀。谷歌的遭遇正说明，当云业务的爆发无法对冲搜索广告放缓的担忧时，市场会毫不犹豫地重新定价。
上游的超级周期：扩产浪潮下的冰与火
云巨头的扩产冲动，顺着产业链向上传导，最先受益的是存储厂商。SK海力士三季度营收104.6万亿韩元，首次单季突破100万亿，营业利润83.3万亿韩元，营业利润率高达80%，HBM4开始放量出货；公司预计存储短缺将持续到2030年，2027至2028年产能已有超过一半被长约锁定，并宣布了约100万亿韩元的股东回报计划。
三星电子同样创下历史：单季营收195.3万亿韩元，营业利润107.4万亿韩元，成为全球首家单季营业利润突破100万亿韩元的科技公司，DRAM均价环比上涨20%，HBM4销售额环比增长三倍。三星预计供给紧张将延续至2028年，同时将2026年资本开支上调约六成至84.4万亿韩元。两家的表态互相印证：存储环节的议价权已经完全回到卖方手中。
不过，同处这条产业链上的英特尔，日子并不好过。
三季度营收165.1亿美元、同比增长21%，数据中心与AI业务营收大增78%，创下15年来最快增速，每股收益0.52美元超出预期，但以色列晶圆厂10月的爆炸事故、PC市场疲软与内存涨价三重压力叠加，四季度营收指引仅为150亿至160亿美元，每股收益指引0.25美元，远低于市场预期，盘后股价一度暴跌逾20%。
同一条产业链，有人利润丰厚，有人连交付都成问题，短缺经济下的分化正在加剧。这轮景气同样漫过了中国产业链：国产DRAM厂商长鑫科技预计前三季度营收约2650亿至2700亿元，同比增长约七倍，归母净利润1310亿至1325亿元，同比大幅扭亏。
分水岭：从谁的模型强，到谁的电够用
把这一周的信号合起来看，行业的竞争逻辑正在发生位移。瓶颈已经从芯片本身，转移到电力、土地、冷却、并网审批和施工周期上，这些环节的扩张以年计，扩产速度天然追不上模型迭代的速度。Andy Jassy说2028年已锁定的需求“令人震惊”，Sundar Pichai说每家公司都希望手里的算力翻倍，本质上是同一句话：算力在可见的未来都是稀缺品。
面对短缺，各家的应对路径开始分化。亚马逊选择继续加码资本开支，用预建产能换取订单转化；微软双管齐下，一边自建数据中心，一边借助债券市场补充弹药；谷歌的取舍最有意味，Sundar Pichai明确表示谷歌云正在为DeepMind的前沿模型研发“让路”，主动放弃部分短期收入，同时把自研TPU推向外部客户，将芯片能力变现；Meta则在现金流转负的压力下，提前推动Muse的商业化，用订阅和交易抽佣让AI业务尽快自我造血。
对下游客户而言，短缺意味着三到五年的长约锁定和更高的价格，中小模型公司只能承受现货波动，行业准入门槛正在从技术能力变成资源获取能力。对上游而言，存储双巨头的纪录利润和英特尔的交付困境同时出现，说明产能兑现能力正在取代产能规划能力，成为新的竞争维度。
财报周留下的核心问题，不是AI需求是否真实，订单簿已经回答了这一点，而是回报周期能否被市场接受。
当三巨头同时出现负自由现金流，而股价反应截然不同时，市场的评分标准已经写明：谁能在现金消耗与收入兑现之间维持可信的平衡，谁就能继续拿到这场比赛的入场券。当然，这一判断也有隐含前提：倘若2027年新增供给集中释放，而广告等传统现金流同时走弱，已签下的长约能否扛住重新议价，将是卖方市场能否延续的试金石。2027年之前，算力大概率仍是卖方市场；2027年之后的故事，取决于今天的巨额投入能否按时变成报表上的收入。（本文首发钛媒体APP，作者 | 硅谷Tech-news，编辑 | 焦燕）
（本文数据来自各公司财报、业绩电话会及公开市场信息，截至2026年10月30日。）
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。
This week, Microsoft, Google and Amazon cloud giants have been on the list, showing the same picture that has been rare for many years: The ability of the cloud factory to calculate is being looted by its customers in line and the order book has been written up to 2028, while the same company's statements are in the form of negative free cash flows that have not been seen for many years.
Demand is so high that it is “sold out” and cash is so tight that it is “turning back” and two sets of signals are just a sign that the AI-based competition is entering a new phase.
Cloud operations are accelerating. Orders are scheduled for three years.
First look at the Amazon. As the world ' s largest cloud manufacturer, AWS earned $45.98 billion in the third quarter, an increase of 39.3 per cent over the same period, the fastest increase in the last five years, with annualized revenues of $184 billion, rising in six consecutive quarters. This figure is supported by an inventory backlog of $578 billion, an increase of more than $80 billion over the three-digit figure of $496.0 billion at the end of the second quarter. CEO Andy Jassy had previously been interviewed to make it clear that the capacity in 2026 had been sold out, that most of 2027 had been booked and that a significant portion of 2028 had received a customer commitment.
Following Microsoft, the figures also exceeded expectations. In the first quarter of 2027, companies received $9.228 billion, an increase of 19 per cent over the same period, and net profits of $36.630 billion, an increase of about 32 per cent. For the first time in the season, Microsoft separately disclosed the absolute income of Azure: $33,165 million in a single season, an increase of 48 per cent over the same period, an increase in volume of $100 billion, and an increase of 73 per cent to $67.7 billion in commercial residual performance obligations. Microsoft also revealed the addition of 24 new data centres in a single season, a pace of delivery that has been rare throughout history.
In contrast, Google clouds are growing at the steepest speed. Alphabet earned $30.71 billion in the third quarter, more than double the increase over the same period, with a backlog of orders reaching $56.1 billion. What is more difficult is the quality of profits: Google Clouds has a record high profit margin of 36.2 per cent, which is three times higher than the same period in the previous year. But the other sentence of Sundar Pichai is even more noteworthy: clients now want to lock in the 2027 calculus, “extremely difficult”. Google even rented external calculations, including the monthly use of SpaceX ' s computing resources to fill its supply gap.
The three companies gave different calibres, but pointed to the same thing: the AI calculus demand curve was much steeper than the supply curve. Cloud manufacturers were concerned that they would not use their capacity after it was built, and now that they would not be able to take orders. The backlog of orders is not intended, but signed contracts that pre-lock revenues in the coming years and pre-position delivery pressures.
The first time the cash flow was in a rush in a capital expenditure competition.
And the next thing we can do is pay for these orders. The Amazon maintained its 2026 capital expenditure guidelines at approximately $22 billion, reaching $77.3 billion in property and equipment expenditures in the three quarters alone, more than 30 per cent above market expectations; Alphabet ' s annual capital expenditure was directed at $15-5 billion to $20.5 billion throughout the year; Meta, a company based on advertising, was one of the most aggressive buyers of AI capital construction in the current round, revised capital spending floors upwards to $135 billion and confirmed that active investment would continue in 2027; and Microsoft launched a debt plan of at least $25 billion shortly after the financial report, with 4.6 times more subscriptions, thereby returning to the United States investment-grade bond market for the first time in nearly eight years.
Expenditures are incurred ahead of income and the pressure on the cash flow statement follows. The net outflow of free cash flows from the Amazon over the past 12 months was $42.36 billion, compared to the same period a year ago; the net flow from Alphabet was negative $9.07 billion, the second consecutive quarter negative; the net outflow from Meta was $2.12 billion, the first since 2012; and the positive cash flow from Microsoft remained at $18.6 billion in four households. Three “cash cows” lost their blood collectively in the same quarter, the first since cloud computing.
Interestingly, investors did not veto negative free cash flows. Microsoft CFO Amy Hood indicated that it would continue financing “when the time is right”, while Meta established a non-voting category C general unit, leaving theoretically some $700 billion in financing space for subsequent AI inputs. It is clear that management has predicted that the capital requirements of the capital cycle will far exceed the carrying capacity of the cash flow and that debt and equity instruments are being framed. The credit market is also making choices: Microsoft's AAA bond orders are in full swing, and lower-rated technology debt issuance costs are noticeably rising, Meta's five-year credit default swap costs were once the highest ever recorded in the contract, and funds are being pooled for certainty.
It's a two-way street decision.
From 21 October, when Meta took the lead, to 28, Microsoft, Google, and 29, Amazon, the order and cash flow story was over, the market gave pricing each time, and the market judgement was quite uneven. Microsoft's overall performance exceeded expectations by 6 per cent at a later stage, and although the next day a small fall of 0.16 per cent followed the overall rate of the Technology Unit, the overall picture was stable; Meta's profit was double-predicted, with six weeks on-line at AI Q42 ZMuse and tens of thousands of users in North America being the biggest flashpoint, with an increase of over 6 per cent at a later stage, but the next day was marred by huge legal fines and cash flows, with a drop of about 7 per cent, which was the only expected to be sold out of four households.
Alphabet is another scene. Collectively, the three core indicators of revenue generation, revenue per share and advertising revenue fell short of expectations, the rate of growth in search operations slowed from 19 per cent in the first quarter to 12 per cent, management explained that the exchange rate was the largest single drag, combined with the concentration of World Cup advertising expenditures in the second quarter, the surge in AI model queries and the low liquidity rate; together with a continuing negative free cash flow and $8.7 billion in non-cash equity investment losses, the stock price fell by 12.64 per cent the following day, hitting the largest single-day decline since the listing and slowing down by 1.38 per cent the same day. Although AWS was more than expected, online retailing grew by only 2 per cent, the four-quarter guidelines were lower in median terms, and continued negative free cash flows were followed by a drop of over 7 per cent.
The gap between the rise and the fall points to the same conclusion: the market no longer pays for “AI inputs” per se, but only for “AI returns”. Market comparisons are not onerous: a significant portion of the current AI income comes from cross-orders between cloud manufacturers and AI, the gold content of which needs to be viewed at a discount; and future profit tables may be under additional pressure if the depreciation years of GPU and data centres are overestimated. The high speed of cloud operations is no longer a talisman, and investors begin to check on a case-by-case basis whether orders can be converted to income, whether cash consumption can be covered by profits and whether the base disk of traditional operations is eroded by AI. Google’s experience shows that the market does not hesitate to re-pricing when the eruption of cloud operations does not allow it to respond to concerns about the slowing down of flushing advertising.
Supercycle upstream: ice and fire in the wave of increased production
The expansion of cloud giants is driven upwards along the industrial chain, benefiting first and foremost from storage manufacturers. SK Hercules collected 1046 trillion won for the third quarter, for the first time exceeding 10 trillion won in a single season, profited 833,000 trillion won, profited up to 80 per cent, and HBM4 started releasing goods; the company expects the storage shortage to continue until 2030, with more than half of its capacity from 2027 to 2028 being locked in and announced a return plan for about 10 trillion won.
Samsung Electronics has also made history: 195.3 trillion won for a single season, 107.4 trillion won for business, becoming the world's first single-season technology company with a profit exceeding 10 trillion won, 20 per cent increase in the DRAM average price and three times increase in the HBM4 sales. Samsung expects supply constraints to continue until 2028, while increasing capital spending by about 60 to 84.4 trillion won in 2026. The statements of the two families were mutually corroborating: the bargaining power of the storage chain had been fully returned to the seller.
However, in Intel, on the same chain, life is not easy.
While the three-quarter harvest of $16.51 billion, an increase of 21 per cent over the same period, the data centre and the AI operations battalion experienced a significant 78 per cent increase, the fastest increase in 15 years, with a return of $0.52 per share exceeding expectations, the October bombing of the Israeli Crystal Circle plant, the weakness of the PC market and the triple pressure of the memory price increase were compounded by a four-quarter line of $15 billion to $16 billion, with each share of proceeds being directed at $0.25 per share, far below market expectations, with post-clip prices falling by over 20 per cent at a time.
In the same industrial chain, there are lucrative people, and even delivery is a problem, and fragmentation under the shortage economy is increasing. This landscape is also spreading over China’s industrial chain: the national production of DRAM’s plant’s director, Zhong Tsing, is expected to earn between 265.0 billion and 270 billion yuan in the first three quarters, a sevenfold increase over the same period, resulting in a net profit of between $131 billion and $132.5 billion for the mother, which is a significant loss.
Watershed: From whose models are strong to whose electricity is adequate
By combining this week's signals, the industry's competitive logic is changing. Bottlenecks have shifted from the chips themselves to electricity, land, cooling, netting approval and construction cycles, expanding year by year and increasing production at a natural rate that does not keep pace with model iterative. Andy Jassy said that the needs that had been targeted in 2028 were “shocked” and Sundar Pichai said that every company wanted to double its calculus, essentially the same sentence: it was scarce in the visible future.
Faced with shortages, the response paths began to divide. The Amazon chose to continue increasing capital spending in exchange for conversion of orders with pre-construction capacity; Microsoft, on the one hand, built its own data centre while supplementing ammunition with the bond market; and Google's trade-offs were most significant in that Sundar Pichai made it clear that Google Clouds was developing a “facilitation” for DeepMind's forward model, voluntarily renouncing some short-term income while pushing the self-study TPU to external customers to convert the chip capacity; and Meta, under pressure of cash flow, advanced the commercialization of Muse by using subscriptions and transaction brokers to make AI's own blood as soon as possible.
For downstream customers, shortages mean long-term lock-in and higher prices for three to five years, small and medium-sized model companies can only withstand spot fluctuations, and industry entry thresholds are shifting from technological capacity to resource acquisition capacity. Upstream, the combination of record profits from the storage of the two giants and the delivery dilemma of Intel suggests that the capacity to deliver is replacing capacity for capacity planning as a new competitive dimension.
The central question left by Financial Week is not whether AI demand is real, which the order book has answered, but whether the return cycle is acceptable to the market.
When the triads simultaneously show negative free cash flows, and stock prices react differently, the market rating criteria already state that whoever can maintain a credible balance between cash consumption and revenue realization can continue to get tickets to the competition. Of course, there is also an implicit premise in this judgement that if the new supply is released in 2027 and traditional cash flows, such as advertising, are weakened at the same time, the long bargain that has been signed will be a litmus test for the sustainability of the seller ' s market. Until 2027, the calculus rate remained in the seller ' s market; the story after 2027 depends on whether today ' s huge investment is translated into income on time. (This article starts with titanium media app, by | Silicon Valley Tech-news, editor | Joo Yon)
(The data are from the company ' s financial statements, performance conferences and open market information as at 30 October 2026.
Auto-fetched and translated by AI Atlas. All rights belong to the original author.