文 | 一千二百字
微软和谷歌今天同步进行三季度业绩放榜，自然年Q3在微软对应的是它2027财年Q1。本周亚马逊也将发财报，美国五大云厂商中主要的三兄弟集中放榜，自然吸引外界在AI语境下的业绩对比分析。这三家公司在最新全球科技公司市值排行榜中，也是紧挨着分列第三、四、五位，排在英伟达和苹果之后。
在发财报前，微软被曝刚启动了一轮密集的削减成本计划，引起业界震动。先是微软Azure部门总裁要求员工针对云服务器租赁业务拿出降低20%成本的方案，后被微软否认了该报道；随后多家海外主流媒体报道，10月份微软人力资源部门在内部推出自愿离职计划，可能涉及5-6万名员工，占比约四分之一，这将是公司史上最大的一次自愿离职方案。该方案涉及的词是buyout，可以理解为“买断”。在今年7月微软刚针对老员工推出的自愿退休项目中，8750名符合资格的员工有超30%接受了买断方案。10月份这次的离职意愿比例及其引起的震荡尚无法估量。
这些改变都与AI相关，这里将微软和谷歌放在一起，做一个资本性支出和自由现金流的对比分析，看看AI吞金兽究竟对云计算大厂施加了多大压力。
四分之三资本开支花在“芯片”上
柱状图分别叠加了这两家公司的资本性支出和自由现金流，相加可以视为经营现金流总和。Q3季度谷歌的资本性支出高达514亿美元，同比增幅114%是近几个季度最高值。这也导致它单季度自由现金流亏空继续扩大，为负91亿美元。而上一季度负59亿美元是谷歌自2004年上市以来20多年首次出现单季度自由现金流为负。
微软Q3资本性支出为452亿美元，同比增长约30%，已经连续3个季度低于谷歌的对应支出。根据微软财报PPT，本次的资本性支出中有四分之三属于短期资产，主要是CPU和GPU及配套，以支持Azure、Microsoft365及研发需求。上一季度，这一占比是三分之二。也就是说，AI芯片这种短期消耗型资产正占据微软越来越大比重。这是它承压的一方面。
本来，芯片这种实体硬件在会计统计中可被视为固定资产，但在大模型语境下，AI芯片的整体折旧速度快于传统服务器硬件。名义上按5-6年折旧，但实际商业有效寿命已经被AI迭代压缩到3-4年，即真实商业价值换代衰减周期比账面折旧期更短。企业还要继续加大采购或自研力度，以填补需求。
由于保持了一个相对克制的资本性支出增幅（部分原因是预期交付的第三方芯片采购出现了延迟交货），且经营现金流增长稳定，Q3季度微软手中的自由现金流还剩下186亿美元，基本上是全球主要云厂商中唯一一个仍能保证连续几个季度自由现金流为正数的公司，其他几家中外云厂商已纷纷进入“赤字轨道”。
亚马逊方面，Q3业绩还没出，从之前几个季度看，今年Q2的TTM连续12个月资本性支出规模已经达到2021-2022年对应季度的3倍左右，今年Q2自由现金流也在近几年首次出现负值。亚马逊高管在分析师会议上说，即使资本性支出如此增长，仍然不够覆盖今年的算力需求。也就是内外部订单接踵而至，AI基础设施不得不继续扩容。
这样，对大模型生态的比拼让本以轻资产模式取胜的谷歌变得越来越“重”；让已在云时代完成一次成功翻盘的微软，不得不通过大规模的降本与人员精简应对；让“云计算一哥”亚马逊拼命自研训练与推理芯片，以降低英伟达依赖。在这三家公司中，目前对英伟达芯片的依赖度从高到低的排序应该是微软-亚马逊-谷歌，而自研芯片内部占比增速以亚马逊更为突出。
投资回报的赛跑
在“AI吞金兽”是既定趋势的认知下，企业的投资回报效率与回报路径就成为外界衡量资本性支出合理性的标准，目前主要有几个维度：比如AI年化收入/资本性支出；对比云增速和资本性支出增速的差值；看算力利用率和新购置芯片上架效率等。在AI尚未贡献明确利润时，市场似乎又重新从看重利润回到看重增速的估值逻辑阶段？实际没有这么简单。
在微软Q3季度923亿美元总营收中，有639亿美元由微软云贡献，同比增速在30%上下，其中Azure同比增长48%，是它近年最快同比增速，Azure收入达332亿美元。由于基础设施层Azure的毛利率远低于像Microsoft 365商业云、GitHub与开发者云服务、Dynamics 365与商业应用云等SaaS、PaaS层的毛利率，所以Azure收入占比的不断提升正在拉低微软云的整体毛利率，Q3季度微软云毛利率降至66%，去年同期为68%。
在预期指引上，微软管理层称2027财年自由现金流仍将保持正数，同时预期自然年Q4季度的资本性支出将超过500亿美元。本来Q3季度的预期就应超过这一数字，但英伟达等第三方芯片延迟交付打乱了部署节奏。这也是外部芯片依赖性严重的弊端。“随着我们持续提升基础设施效率，每一美元资本开支未来有望支撑更多营收、带来更高回报。”微软管理层概括地说。
谷歌方面，近期市场的一项利好消息是，OpenAI被曝有计划批量采购谷歌的自研芯片TPU。这意味着有望直接把谷歌TPU从内部自用+中小客户租赁的小体量业务，推升至百亿美元级的独立增长曲线。
Q3季度谷歌母公司Alphabet营收1276亿美元。其中云收入为307亿美元，同比增速103%是微软Azure增速的两倍多，也创下自身记录，云营收规模小幅落后于微软Azure。其中，外部企业客户集中订阅采购谷歌GCP的AI算力集群用于大模型训练和推理部署，是推动云营收增长的核心因素；而TPU产品销售成为业绩重要增量，也让谷歌云加速与OpenAI采购TPU传闻形成互证。
但财报披露后谷歌盘后股价下跌10%左右，按照10月28日的市值计算相当于直接砍去了3800亿美元市值。高增长没有换来市场的充分信心，可能与它继续扩大的自由现金流赤字和长期债务规模有一定关系，以及由投资收益从正到负导致的净利润同比缩减25%、EPS 2.12美元大幅不及约3.0美元预期有关。
从目前建立的AI投资回报路径上看，亚马逊AWS是从“算力房东”到“模型转租人”的模式转变，靠自研芯片逐步压低这一过程的成本；微软是靠与OpenAI大模型深度捆绑，给自家高毛利率的一系列SaaS生产力工具附加更高价值和订阅价格，把AI卖给已有客户，而压力来自折旧费用；谷歌是这三家中算力自主化程度最高的，但TPU规模化客户需求还没有集中释放，变现路径限于云的增长和广告效率提升。
当前市场反应说明，尽管取得局部高光表现，这些公司仍没有走出外界对持续高投入下的回报效率担忧的心理周期，承压仍在。今天甲骨文股价上涨10.73%，是因为在投资者日公司给市场吃了一颗定心丸，宣称2027至2030财年的整个AI投资周期内新增外部融资总额不超过200亿美元，意味着放弃大肆举债填补CapEx空缺，市场对克制的投资行为予以股价奖励。接下来腾讯、阿里、百度等国内云厂商将陆续发布三季报，类似的处境和矛盾将吸引外界持续关注。
本文由 AI宝库自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
1,200 words.
Microsoft and Google are running a three-quarter performance roll today, and Natural Year Q3 corresponds to Microsoft's 2027 fiscal year Q1. This week, the Amazon will also make a financial statement, and the leading three brothers among the five cloud manufacturers in the United States will be on the list, naturally attracting a comparative analysis of performance in the AI context. The three companies, in the latest global market value list for technology companies, are also ranked third, fourth and fifth, behind the British and Apples.
Prior to financial disclosure, Microsoft had just been exposed and had launched an intensive cost-cutting round, which had triggered an industry shock. The report was denied by Microsoft's Managing Director, who asked for a 20 per cent cost reduction programme for cloud-server leasing, and subsequently by Microsoft, who reported in mainstream overseas media that in October the Microsoft Human Resources Department had introduced a voluntary separation plan internally, which could involve 50,000-6,000 employees, or about a quarter of that, which would be the largest voluntary separation programme in company history. The terms of the programme are “buyout”, which can be understood as “buyout”. Of the voluntary retirement programmes just launched by Microsoft in July this year for older employees, over 30 per cent of the 875 eligible employees accepted the buyout programme. The rate of willingness to leave in October and the shock it caused are not yet measurable.
These changes are all related to AI, where Microsoft and Google are put together to do a comparative analysis of capital expenditure and free cash flows to see how much pressure the Ai swallows on the cloud computing plant.
Three-quarters of capital spending on chips.
The column chart incorporates the capital expenditures and free cash flows of the two companies, respectively, and can be considered to be the sum of operating cash flows. The Q3 quarterly capital expenditure of Google amounted to $51.4 billion, an increase of 114 per cent over the same period, the highest in recent quarters. This has also led to the continued expansion of its single-quarter free-flow deficit to negative $9.1 billion. In contrast, negative $5.9 billion in the previous quarter was the first single-quarter free cash flow negative in Google since it was listed in 2004.
The capital expenditure of Microsoft Q3 was $45.2 billion, an increase of about 30 per cent over the same period, and had been lower than Google ' s counterpart expenditure for three consecutive quarters. According to Microsoft Press PPT, three-quarters of this capital expenditure is short-term, mainly CPU and GPU and associated, to support Azure, Microsoft 365 and R & D needs. This ratio was two thirds in the previous quarter. In other words, short-term expendable assets such as AI chips are taking on a growing share of Microsoft. That's one side of it.
Initially, such physical hardware as chips could be considered as fixed assets in accounting statistics, but in the LLM context, the overall depreciation of the chips was faster than that of conventional server hardware. Nominal depreciation is 5-6 years, but the actual commercial active life has been reduced to 3-4 years by AIT, i.e., the real commercial value-for-declining cycle is shorter than the book depreciation period. Enterprises also need to continue to increase their procurement or self-study efforts to meet demand.
As a result of a relatively restrained increase in capital expenditure (in part as a result of delays in the delivery of expected third-party chips) and a steady increase in operating cash flows, the free cash flow left $18.6 billion in the Q3 season in Microsoft's hands is essentially the only company among the world's leading cloud manufacturers that can still guarantee a positive free cash flow for several consecutive quarters, and several others outside their homes have entered the “deficit track”.
In the case of the Amazon, Q3 performance has not yet been achieved, and this year Q2 Q2 capital TM expenditure has been on a 12-month scale that is three times as large as the corresponding quarter of 2021-2022, and Q2 free cash flows have been negative for the first time in recent years. At the meeting of analysts, senior Amazonian executives said that even with such growth in capital spending, this year ' s calculus needs were not adequately covered. That is, internal and external orders followed, and the AI infrastructure had to continue to expand.
As a result, Google, which has made the LLM ecological equations a light-asset victory, has become increasingly “heavy”; Microsoft, which has completed a successful roll-over in the era of clouds, has had to streamline its response through large-scale downfalls with people; and Amazon, who has worked so hard to develop training and reasoning chips of his own to reduce the dependence of Ingweida. Among the three companies, the current high-to-low order of reliance on the British Weeda chip should be Microsoft-Amason-Google, while the self-study chip's internal share is more prominent than the increase in the Amazon.
A race for return on investment.
In recognition of the established trend, the return efficiency and return path of an enterprise's investment became the external measure of the reasonableness of capital expenditure, with several main dimensions: AAI annualized income/capitalized expenditure; contrasting the differentials between cloud and capital-based expenditure; calculus utilization; and efficiency on top of newly purchased chips. When AI has not yet made a clear profit, does the market seem to return to the valuation logic phase of focusing on profits? It's not that simple.
Of the $92.3 billion received by Microsoft Q in the third quarter, $63.9 billion was contributed by Microsoft clouds, at a rate of 30 per cent or less, with Azure growing by 48 per cent over the same period, the fastest in recent years, with it earning $33.2 billion. As the Māori rate for Azure in the infrastructure layer is far below the Māori rate for SaaS, PaaS, such as Microsoft 365 Business Cloud, GitHub and Developer Cloud Services, Dynamics 365 and Business Applications Clouds, the increasing share of Azure income is lowering the overall Māori rate of Microsoft Clouds, which fell to 66 per cent in the Q3 quarter, compared to 68 per cent in the same period last year.
In anticipation, Microsoft management stated that the free cash flow for the 2027 fiscal year would remain positive, while capital expenditure for the Q4 quarter of the natural year was expected to exceed $50 billion. This figure should have been exceeded by the Q3 season, but delays in the delivery of third-party chips, such as In Weida, disrupted the pace of deployment. This is also a serious drawback of external chips. “As we continue to improve infrastructure efficiency, each dollar of capital expenditure is expected to support more revenue and higher returns in the future.” Microsoft management summed up.
On Google, one of the good news of the recent market is that Open AI was exposed to a planned bulk purchase of Google's self-study chip TPU. This means that it is expected that the Google TPU will move directly from its own internal + small and medium-sized client-leased small volume business to an independent growth curve of billions of dollars.
Q3 quarterly, the Google parent company Alphabet received $1276 million. Of this, cloud revenues amounted to $30.7 billion, more than twice as fast as Microsoft Azure ' s growth rate as compared to 103 per cent in the same year. Among these, the use of the LLM training and reasoned deployment of the AI Calculator cluster for the procurement of GoogleGCP by external business clients is central to the growth of cloud-blowing; and the sale of TPU products has become an important increase in performance, which has allowed Google Clouds to accelerate the exchange of information with Open AI procurement TPU.
However, after the disclosure of the financial statements, the post-Google stock price fell by about 10 per cent, amounting to a direct cut of $380 billion based on the market value as at 28 October. The lack of full confidence in high growth in exchange for markets may be linked to its continued expansion of the free-flow deficit and the size of its long-term debt, as well as to a 25 per cent decline in net profits from positive to negative investment returns and a significant shortfall of approximately $3.0 million in the expected EPS $2.12.
In terms of the currently established AI investment return path, the Amazon AWS model shift from a “calculating landlord” to a “model subtenner” model, with self-study chips gradually reducing the cost of the process; Microsoft is limited to cloud growth and advertising efficiency in the transformation path by placing a deep tie to OpenAi LLM, adding higher value and subscription prices to a series of SaaS productivity tools with high rates of home-based Māori, selling AI to existing customers, with pressure from depreciation costs; Google is the highest for all three households with the highest degree of computing autonomy, but TPU-sized customer demand has not been released centrally.
The current market response suggests that, despite local highs, these companies have not emerged from the psychological cycle of external concerns about the efficiency of the return under continued high levels of investment and are still under pressure. Today ' s increase of 10.73 per cent in Oracle stock prices is due to the fact that investors ' day took a pill for the market, claiming that additional external financing would not exceed $20 billion over the full AI investment cycle for the fiscal year 2027-2030, implying abdication of massive debt to fill the CapEx vacancy, and that the market offered equity incentives for restrained investment behaviour. This will be followed by Tencent, Ali, Baidu, etc., which will be followed by the publish trilogy, and similar situations and contradictions will attract sustained outside attention.
Auto-fetched and translated by AI Gem Box. All rights belong to the original author.