文 | 寻空的营销启示录
有一种感觉，大厂开始缩减市场预算，同时拿出更多的钱投资AI。
2019年，Alphabet、Meta、微软三家加起来的收入是3584亿美元，销售与营销费用465.5亿美元。折算下来，每100美元收入里有13美元花在推广上。13%，这也符合一般的营销推广比例逻辑。
2025年，同样这三家公司，收入8855亿美元，推广费663.4亿美元，每100美元收入对应7.5美元，相比6年前，下降了5.5个百分点。
六年时间，收入涨了147.1%，推广费只涨42.5%。
另一边，数据中心、芯片、模型研发和AI人才，正在获得越来越多的投入。
2025年，Alphabet的资本开支约为914亿美元，上一年约为525亿美元。Meta 2025年资本开支722亿美元，2026年预计1150亿至1350亿美元。
大厂们，一年买算力的钱，抵得上好几年卖产品的推广预算。
一边要求效率，一边愿意投入，钱的使用标准，正在跟着公司对未来增长的判断发生变化。
01 大厂赚得更多，市场投放却更谨慎
过去几年大厂赚得越来越多了，但投入市场推广的费用却有点涨不动了。
2019年到2025年，Alphabet、Meta、微软，三家合计营收增长147.1%；合计销售与营销费用只增长42.5%；加权费用率从13%降到7.5%。
拆开看，落差更大。
Alphabet 从11.4%降到7.1%，收入涨了148.9%，推广费只涨55.4%。Meta 从14.0%降到6.0%，收入涨184.3%，推广费涨21.4%。微软按财年算，从14.5%降到9.1%，收入涨123.9%，推广费涨40.9%。
同一口径下，微软截至2026年6月的财年占比是8.1%，还在往下走。
费用率下降，不等于费用在减少，三家的推广费六年里都在涨，Meta 涨了两成，Alphabet 涨了五成半，这里说的是每赚一块钱要拿多少出去做推广，在这个维度，下降的幅度是42.3%。
过去二十年，收入规模扩大之后，每多赚一美元，要多花同比例的钱去推广，这样的做法几乎是默认的，增长靠新用户，新用户靠触达，触达靠预算，预算随收入线性放大。
现在，情况彻底变了。
02 品牌和入口，让大厂有底气少花钱
大厂为什么有条件降低推广投入的相对强度？
一个直观原因，是品牌认知已经建立。搜索用Google，玩社交网络用Facebook和Instagram，办公用Office，这些东西，短期内推广预算降一降，对市场份额的影响也很有限。
不过，科技公司的优势还多了一层。
一家普通企业推广新品，常常要先购买触达机会。拥有庞大产品体系的平台，可以在已有产品中介绍新功能，也可以通过客户关系、服务升级和交叉销售，把新产品带到原有用户面前。
对拥有这些条件的企业来说，新功能的第一批用户，可能就在已有的账户、产品界面和客户名单里。品牌帮助用户理解，分发能力帮助企业触达。两种积累叠加，外部推广需要就没那么紧迫。
Google可以通过搜索框免费推广Gemini，Facebook可以在信息流中免费植入Muse的广告，字节跳动也可以在抖音中免费推广豆包。
当成熟业务能够复用这些积累，管理层就有条件对新增推广支出提出更高要求。已经建立的优势要继续维护，但每一笔新增预算，需要说明自己还能带来什么。
03 AI开始获得更大的支持
拥有成熟业务的大厂，仍然会担心未来。
今天的用户还在使用原有产品，明天可能直接向AI提问、交办工作，让它帮助选择商品。搜索、办公、购物的起点，都可能随之改变。
大厂最清楚入口有多值钱，也就很难安心等到新入口长成之后再参与。
模型研发、算力和人才，都需要提前投入。用户规模能做到多大，收入何时能够覆盖成本，还没有确定答案，但如果等到回报看得足够清楚，竞争位置也已经被别人占住了。
腾讯在2026年3月发布全年业绩时，马化腾就提到，核心业务的韧性和现金创造能力，为持续增加AI投资提供了资源，包括招聘顶尖人才和升级基础设施。
这段话把大厂的处境说得很明白。老业务还在赚钱，赚来的钱有了新的去处，对成熟业务，管理层可以要求它继续提高效率；对AI，则愿意拿出资金，承担一段时间内投入大、回报不明朗的压力。
这就是其他业务市场难受的地方，只要管理层把AI视为下一轮竞争的关键，其他项目申请预算时，就得和它一起接受比较。
服务商先感受到了这笔取舍。S4Capital 2025年营收7.548亿英镑，同比下降11%，净收入下降10.8%。创始人在年报里的解释是，科技客户占公司收入近一半，这些客户持续把资本开支优先投向AI产能扩张，营销这类运营支出排在后面。
这比单看科技公司的费用率更接近营销行业的实际处境。客户仍有钱，也愿意花钱，只是过去交给营销公司的那部分，现在有了更优先的用途。
国内公司的情况如何？
字节跳动今年9月拿到一笔约296亿美元的银团贷款，由近30家银行参与，规模从最初计划的200亿美元扩大到296亿美元。虽然名义上用于一般企业用途，但大部分报道认为，其中相当部分预计会支持AI芯片、数据中心等AI投入。
阿里8月份直接配售新股融资800亿港元，而且明确宣布100%用于AI，约60%扩充全球算力基础设施，约40%建设超大规模AI数据中心和升级云基础设施。
很显然，AI获得了更多资源。
04 广告公司的冲击才刚刚开始
S4Capital遇到的压力，恐怕只是开始。
我认识不少Agency的老板，他们有一个相近的感受，现在大厂还愿意大手笔投放的业务，几乎只剩下AI了。其他业务的推广预算也有，但一年不如一年。
分众传媒2025年年报里，楼宇媒体的互联网行业投放金额同比增长188.59%，收入占比从9.07%跳到25.15%，公司把它归因于AI大模型竞争加剧。
同一份年报里，作为基本盘的日用消费品投放下降16.22%。一个客户行业在翻倍，另一个在退。
代理公司的账本更直接。蓝色光标2025年营收686.93亿元，增长12.99%，全案广告代理收入35.39亿元，同比下降18.29%；同期AI驱动收入37.25亿元，增长210.42%，占总营收5.42%。
老业务在缩，新业务在涨，多出来的增量主要来自AI和出海投放。
甲方收入增长，整体营销预算却没有同步增长。在一个增长有限的总盘子里，AI推广拿走更多份额，其他业务能分到的比例自然下降。
这笔账算下来，乙方就知道，接下来该去找哪条业务线了，眼下做AI产品的品牌传播、内容和获客，确实还有机会。
但这也只能解决眼前的问题。
一位大厂高管告诉我，他们已经开始让AI直接投放，投放本来就是一个反复测试、观察反馈、调整的过程，而AI的自监督学习和强化学习机制一旦跑起来，长期效率会明显超过人。
Meta的Advantage+已经覆盖素材生成、预算分配、受众选择和出价调整。到2026年初，Meta又披露，正在测试面向广告主的AI业务助手，让客户直接向它询问优化建议。
平台希望客户在自己的系统里，完成更多从制作到投放的工作。
国内也有类似的变化。巨量千川在2026年9月更新的产品手册中介绍，仍在内测的AI成片推荐，可以根据商家已有素材、商品和直播间信息自动生成视频，商家筛选采纳后，素材就能加入投放计划。这已经把部分制作和投放衔接工作，放进了平台的产品功能里。
以前，甲方需要向乙方购买的一些服务，现在打开广告后台就能找到。
广告公司更难受的地方也在这里。甲方的投放预算即使没有减少，需要支付给乙方的制作费、操作费和服务费，也会先降下来。媒体消耗还在，广告公司能从中挣到的钱却变少了。
结语
13美元到7.5美元只是一个会计结果，它背后是一次决策。真正的变化发生在解释权上。
过去，增长归因于投放能力。多花一笔预算，多买一批用户，多一层曝光，增长可以被营销解释，预算就有理由跟着收入走。现在，增长越来越多地被归因于模型能力、算力规模和入口位置。
当一家公司能用模型解释增长，它就不需要再用媒体预算解释增长。
获客正在从市场部的事变成产品的事。入口内嵌，分发自带，用户第一次接触产品时，广告不再是必经之路，一次真实使用才是。习惯买不到，只能长出来。买不到的东西，预算就使不上劲。
广告预算不会消失，它会向能证明增量的地方聚集。新品类上市、新入口争夺、新用户教育，这些仍然要花钱，而且花得更急。真正被压缩的是另一类支出，只能证明自己被看见、不能证明自己带来增长的那种。
增长的解释权换了人，预算就跟着换人。
营销没有退场，它只是被要求先证明自己是增长，再谈是不是费用。
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。
A marketing revelation to find space
There was a feeling that large plants were starting to reduce their market budgets while at the same time investing more money in AI.
In 2019, the combined revenues of the Alphabet, Meta and Microsoft were $35.84 billion, and sales and marketing costs were $46.55 billion. This translates into $13 per $100 income spent on outreach. 13%, which also follows the general marketing promotion ratio logic.
In 2025, the same three companies received $885.5 billion in revenues and $6.634 billion in extension costs, equivalent to $7.5 per $100, a decrease of 5.5 percentage points compared to six years ago.
In six years, income increased by 147.1 per cent, while the cost of extension increased by 42.5 per cent.
On the other hand, data centres, chips, model development and AI talent are gaining increasing input.
In 2025, Alphabet ' s capital expenditure was approximately $91.4 billion, compared with $52.5 billion the previous year. Meta spent $72.2 billion on capital in 2025 and projected $115 billion to $135 billion in 2026.
Large factories, which buy money for a year, offset the promotion budget that has sold their products for several years.
While demanding efficiency, willing to invest, the standards for the use of money are changing as companies judge future growth.
Large factories make more money and markets are more prudent.
Large factories have been earning more and more over the past few years, but the cost of market extension has increased somewhat.
Between 2019 and 2025, Alphabet, Meta and Microsoft received 147.1 per cent of the total, while total sales and marketing costs increased only 42.5 per cent, and the weighted cost rate decreased from 13 per cent to 7.5 per cent.
Open it, it's bigger.
Alphabet fell from 11.4 per cent to 7.1 per cent, and income increased by 148.9 per cent, and promotion fees increased by only 55.4 per cent. Meta fell from 14.0 per cent to 6.0 per cent, income rose by 184.3 per cent and promotion fees increased by 21.4 per cent. Microsoft dropped from 14.5 per cent to 9.1 per cent per fiscal year, and income increased by 123.9 per cent, and promotion costs increased by 40.9 per cent.
At the same calibre, Microsoft ' s share as of June 2026 was 8.1 per cent.
The cost rate is falling, not equal to the cost, the cost of extension for the three families has been rising for six years, Meta has increased by 20 percent, Alphabet by 50-50 percent, and here is the amount of money spent for promotion, which is 42.3 per cent.
Over the past two decades, every dollar that increases the scale of income has to be used to scale up at the same rate. This approach is almost tacit. Growth depends on new users, new users touch and touch budgets, and budgets are scaled up along income lines.
The situation has now changed completely.
2 brands and entrances, which give the big factory the power to spend less.
Why should large plants be in a position to reduce the relative intensity of extension inputs?
An intuitive reason is that brand recognition has been established. Search for Google, social networking Facebook and Instagram and Office, which have a short-term reduction in the promotion budget, also have a limited impact on market share.
However, there is an additional layer of advantage for technology companies.
A common enterprise promotes new products, often first purchasing touch opportunities. Platforms with large product systems can introduce new functions in existing products or bring new products to their original users through customer relationships, service upgrades and cross-marketing.
For enterprises with these conditions, the first users of the new functionality may be in existing accounts, product interfaces and customer lists. Brands help users understand and distribute skills to help businesses reach. The need for external outreach is less urgent than the accumulation of both.
Google can promote Gemini free of charge through search boxes, Facebook free of charge through the Muse ad and byte beat free of charge in Douyin.
When mature operations can re-utilise these accumulations, management is in a position to place higher demands on additional extension expenditures. The advantages that had been established needed to be maintained, but each new budget needed to be clarified what it could bring.
03 AI began to gain greater support
Large factories with mature businesses still fear the future.
Today's users are still using the original product, and it is possible that tomorrow they will question AI directly and turn it over to help select commodities. Searches, offices and the starting point for shopping can change.
Large factories know best how valuable the entrance is, which makes it difficult to wait until the new entrance has grown.
Modelling, numeracy and talent all require early investment. The extent of user size, the timing of income to cover costs, has not yet been determined, but the competitive position has been taken over by others when returns become clear enough.
When Tencent performed throughout the year in March 2026 publish, Maerteng mentioned that the resilience and cash-generating capacity of the core business provided resources for a sustained increase in AI investments, including recruitment of top talent and upgrading of infrastructure.
This is a clear statement of the situation of the factory. The old business is still making money, the money it earns is new, and for mature operations management can ask it to continue to improve efficiency; for AI, it is willing to contribute to the pressure of a period of heavy investment and uncertain returns.
This is where other business markets suffer, and so long as management sees AI as the key to the next round of competition, other projects will have to be compared when they apply for budgets.
Service providers first felt the trade-off. S4Capital received Pound754.8 million in 2025, a decrease of 11 per cent over the same period and a decrease of 10.8 per cent in net income. The founders explained in the annual newspaper that technology clients account for nearly half of the company ' s revenues and that they continue to prioritize capital expenditure towards AI expansion, and that such operating expenses are lagging behind in marketing.
This is closer to the actual situation in the marketing industry than simply looking at the cost of technology companies. Clients still have money and are willing to spend it, just the part that was handed over to marketing companies, and now have more priority uses.
What about domestic companies?
In September this year, byte was awarded a consortium loan of approximately $29.6 billion, with the participation of nearly 30 banks, which expanded from the originally planned $20 billion to $29.6 billion. While nominally used for general enterprise purposes, most reports suggest that a significant portion of them are expected to support AI inputs such as AI chips, data centres, etc.
In August, Ali directly sold a new share of HK$ 80 billion and announced 100 per cent for AI, about 60 per cent for the expansion of the global computing infrastructure, and about 40 per cent for the construction of a mega-AI data centre and the upgrading of cloud infrastructure.
Apparently, AI has received additional resources.
Advertising company shock just started.
The pressure on S4Capital is probably only the beginning.
I know a lot of Agency's bosses, and they've got a similar feeling, and now the big factory is willing to put a lot of money into the business, almost all of them AI. Other operations also have an extension budget, but less than one year.
In the annual newspaper for 2025, the Internet industry in the building media registered an increase of 188.59 per cent over the same period. The share of income jumped from 9.07 per cent to 25.15 per cent, which companies attributed to increased competition in AZLLM.
In the same annual paper, the delivery of consumer goods for daily use as basic plates fell by 16.22 per cent. One customer business is doubling, the other is retreating.
Agent's books are more direct. The Blue Curb received $6,863 million in 2025, an increase of 12.99 per cent, and the total advertising agent received $3.539 billion, a decrease of 18.29 per cent over the same period; and AI-driven income of $3.725 billion, an increase of 210.42 per cent, or 5.42 per cent, over the same period.
The old business is shrinking, new business is on the rise, and the increase is mainly due to AI and shipping.
Revenue growth on the one hand has not been accompanied by parallel increases in the overall marketing budget. In a compact with limited growth, AI has taken more shares, and the share that other operations can allocate is naturally declining.
With this amount, B would know which line of business to follow, and now branding, content and buyers of AI products, there are real opportunities.
But it can only solve the immediate problems.
A large plant manager told me that they had begun to put AI directly, which was a process of repeated testing, observation of feedback, adjustment, and that the long-term effectiveness of AI ' s self-supervised learning and enhanced learning mechanisms would clearly exceed that of people.
Advantage+ in Meta already covers asset generation, budget allocation, audience selection and bid adjustment. By early 2026, Meta had also disclosed that it was testing the AI Business Assistant for advertisers to ask them directly about optimization proposals.
The Platform expects clients to do more in their own systems, from production to delivery.
Similar changes have occurred in the country. In a product manual updated in September 2026, a large amount of Chikawa explained that the still-successful AI recommendation could automatically generate video based on asset, goods and live inter-recipient information already available in the business, and that asset could join the project once the commercial screening had been adopted. This has integrated some of the production and delivery work into the Platform ' s product function.
Previously, some of the services that A needed to buy from B could be found by opening the backstage of the advertisement.
Advertising companies are worse off here. Even if the budget allocated for A is not reduced, the costs of production, operation and services to B would have to be paid first. Media consumption is still in place, and the amount of money that advertising companies can earn from it is diminishing.
Concluding remarks
US$ 13 to US$ 7.5 was an accounting result, behind which was a decision. The real change took place in the right of interpretation.
In the past, growth was due to the ability to deliver. With an extra budget, an additional number of users and an additional layer of exposure, growth can be explained by marketing, and the budget has reason to follow income. Growth is now increasingly attributed to modelling capabilities, calculus size and entry points.
When a company can use models to explain growth, it does not need to use media budgets to explain growth.
The winner is moving from a marketing department to a product. The entrance is embedded, the belt is distributed, and when users first come into contact with the product, advertising is no longer a necessity, but a real use. You can't buy it, you can only grow it. The budget can't make a difference without buying.
The advertising budget will not disappear, and it will gather where it can prove the increase. New product listings, competitions for new entry points and education for new users continue to be costly and more expensive. What is really being squeezed is another type of expenditure that can only prove that it is seen and that it cannot prove that it brings about growth.
When the power to interpret growth changes, the budget changes.
Marketing did not go back, but it was asked to prove its growth before talking about costs.
Auto-fetched and translated by AI Atlas. All rights belong to the original author.