Meta刚宣布要加码押注AI、提高资本支出，股价应声下跌，信用风险指标创历史新高。
七巨头 vs. 四九三：一组被拉直的财务曲线
人均营收：Mag 7 在涨，S&P 493 没动
Apollo 首席经济学家 Torsten Slok 在报告中放出一组对照数据。自 2022 年起，「七巨头」（Alphabet、Amazon、Apple、Meta、Microsoft、Nvidia、Tesla）的人均营收从 55 万美元拉到接近 80 万美元。同期，标普 500 剔除这七家后剩下的 493 家公司，人均营收七年里一直趴在 55 万美元附近，画成图就是一条水平线。
493家公司的财报翻完，数字没变
利润率：近 30% vs. 10%，差距在拉大而不是收窄
第二组数据走势更直白。七巨头整体利润率已逼近 30%，S&P 493 仍停在 10% 上下。两条线不是在平行移动，是在分叉。Slok 没有用「意味着」或「预示」这类词，只写了一句：目前没有证据表明 AI 提升了这 493 家公司的生产效率。
AI 红利分配：从算力到利润的断裂链
全要素生产率微降 0.07%：宏观数据还没接住 AI
第三组数据来自旧金山联储。截至 2026 年第二季度，美国经生产要素利用率调整的全要素生产率同比下降 0.07%。没有加速，甚至微跌。AI 叙事最热的两年里，宏观生产率指标既没跳升，也没出现即将起飞的前兆信号。
把三张图叠起来看，一个事实浮出来：AI 的财务回报集中在构建算力基础设施的那七家。对撑起大盘的 493 家传统企业，Token 支出正在被重新写进成本科目，不再算作增长投资。Slok 给这组数据加了脚注：「AI 或许是一个强大的计算器，但商业里需要的智能，大多是判断力。」
"Token 支出"的身份转变：从增长引擎到成本中心
CFO 视角：账单厚了，收入行没出现
Slok 的原话可以直接贴进审计底稿：「主流企业正在把 Token 支出从增长引擎重新归类为成本中心，因为它在挤压利润率。」落到 CFO 桌上，意思是你上月续的 GPT-4o API 年费，下季度得从研发资本化挪到销售费用。MIT 南达实验室 2025 年 8 月的研究给出了一个数字，95% 的企业生成式 AI 项目没有可衡量的财务回报。CFO 看到这个比例，收缩预算几乎是条件反射。
VC 创业公司还在烧 Token，传统企业开始算 ROI
两边算的不是同一笔账。VC 投的 SaaS 公司把 Token 消耗当增长杠杆，烧钱换 ARR 是投资期叙事，亏损可以跟投资人讲。标普 493 里的制造业、零售、金融，CFO 面对的是季度盈利指引，没有「再给两个季度证明期」的余裕。同一个模型 API，在创业公司账单里叫基础设施投入，在传统企业损益表里叫未变现成本。
Meta 财报日的信号：加投 AI 反被当利空
上周 Meta 主动公布加码 AI 资本支出的计划。市场没有解读为跟上了节奏，而是又多了一张成本单。股价当日下跌，信用风险指标升至历史新高。一家市值超三千亿美元的公司，加投 AI 的消息被当作利润侵蚀信号消化。更狠的是，华尔街已经在给 AI 支出贴标签，它等于成本，除非你是那七家卖铲子的。
为什么「模型变强」不等于「利润变厚」
算力降价 ≠ 营收增速上升：Slok 的判断逻辑
推理成本每半年砍一刀，这件事本身不构成利润故事。Slok 的论证路径很直接：模型参数翻倍、token 价格跌到去年的三成，但 S&P 493 的人均营收七年没动过五十五万美元那个台阶。算力便宜了，买算力的人并没有因此多卖出一件货。他用的类比是计算器：算得再快，也替不了商务谈判桌上那句「这单能不能再压两个点」。AI 或许是一个强大的计算器，但商业里需要的智能，大多是判断力。
组织、信任、利益关系：AI 解不了的那层摩擦
一家中型制造商引入 AI 排产后，确实省了约 5% 的调度人力，但订单量没变、客户没多、竞争对手也没因此丢单。省下的五个人力成本被管理层重新分配到了别的部门预算里，利润行没动。Slok 把这类情况归为「组织摩擦」：部门间的利益博弈、供应商的长尾谈判、客户对替代方案的信任阈值，这些环节 AI 加速不了，只能绕过去。模型再聪明，也替不了销售在饭局上建立的那层信任。更狠的是，省下来的成本大概率被内部消化成新的预算科目，不会自动变成股东回报。
落地节奏差还是能力边界？Slok 持怀疑态度的理由
最自然的反驳是「非科技企业落地慢，再给两三年」。Slok 没接这个假设。他调了美国普查局每两周一次的大规模调查（覆盖 120 万家企业），发现拥有 250 名以上员工的大企业 AI 采用率正在回落，不是「还在爬坡」，而是试过了、算了 ROI、决定不续了。叠加 MIT 南达团队那组 95% 生成式 AI 投资零回报的数据，节奏差的解释站不住脚。更准确的说法是：多数传统企业的业务瓶颈不在计算速度，而在判断力、信任和利益分配，这三样当前一代模型解不了。
模型能力到利润的断裂链路
数据没骗人，只是没人想听
上一段把 Sloc 对"落地节奏差"的怀疑讲清楚了。怀疑归怀疑，财报季是检验怀疑的硬时刻。未来四个季度，493 家公司要把 AI 支出从叙事里拽出来，放到损益表上对质。
财报季该盯什么：AI 支出的新读法
资本支出项下 AI 占比：从"可选项"变成"必答题"
AI预算终于被CFO单独拎出来审了
过去两季，"我们在布局 AI"还能当定性描述塞进 10-K 的风险披露页。这一季不行了。卖方研究员开始逐行拆 capex 和 opex：哪部分是 GPU 集群折旧，哪部分是 API 调用费，哪部分是内部数据标注人力。拆完，问题就具体了。你花了 3000 万美元买推理算力，对应的 AI 归因营收是多少。如果答案是零，这笔钱在 CFO 的表格里就是一条纯粹的侵蚀项。
对 493 里的传统企业，还有一个更实在的取舍。它们不需要每天追最新前沿模型，需要的是一个稳定接进 ERP 和 CRM 的推理层。问题不再是"选哪个模型"，而是"这条流水线到底改变了哪个决策节点"。答不上来，Token 账单就是成本，不是投资。
投资者关注点迁移：从"AI 叙事"到"AI 损益"
两年前的财报电话会，CFO 说一句"我们正加大 AI 投入"，股价就吃一波叙事溢价。现在分析师翻到 P&L 页，问的是另一套问题。AI 专项收入多少，AI 专项成本多少，两者差额在扩大还是收窄。
卖方研究框架也跟着换。TAM 幻灯片少了，AI 边际影响敏感性分析多了。对利润率本来就在 10% 附近徘徊的 493 公司来说，每多烧一块钱 Token 而收不回一块钱营收，净利率就从 10.0% 滑向 9.7%。按年营收几百亿的盘子算，这个差值一年下来吃掉几千万利润。你买的旗舰款推理服务，瞬间变成平替套餐。
一个还没人回答的问题：判断力能不能被定价
Sloc 那句"AI 或许是一个强大的计算器，但商业里需要的智能，大多是判断力"，落到资产负债表上就变成一道还没人做过的题。判断力能不能被摊销。
一个 CFO 能给 Token 定价、给 GPU 小时定价、给 SaaS 席位定价。但"客户对你做错一次决策后仍然选择续约"的信任值多少，进不了表。组织内部的问责链条、利益相关方之间的博弈，也不存在一个折旧模型。
所以这一季财报里真正该追问的不是"AI 花了多少"，而是"AI 改变了哪个本来不会改变的决策"。如果答案是"没有，只是做同一批决策快了一点"，支出性质就锁死了：运营成本，不是增长投资。
这一季可以做的三件事。看 493 公司财报时，把 AI capex 加 opex 和 AI 归因收入做成一列差值，连续四个季度差值为负且不收窄，叙事溢价就该撤了。做传统企业业务的，下次董事会前备一个数：AI 改变了决策 X，省了或赚了 Y。拿不出这个数，坦白讲，你买的是计算器，不是智能。这套判断框架在 AI 能力还停在"加速执行"阶段时成立；下一代模型如果真正移动了决策边界本身，成本中心的归类需要推倒重来。
参考文献
- Torsten Sløk, Apollo Global Management, "AI has delivered zero financial payoff outside the tech industry," Oct 2026. finance.yahoo.com/economy/art…
- 格隆汇, "阿波罗首席经济学家：AI尚未给科技业以外的美国企业带来回报," 2026-10-25. caifuhao.eastmoney.com/news/202610…
- PANews/TradingView, "阿波罗首席经济学家：AI尚未给科技业以外的美国企业带来回报," 2026-10-25. es.tradingview.com/news/panews…
- IT之家/搜狐, "阿波罗首席经济学家：除科技七巨头外，AI还没帮美国企业赚到真金白银," 2026-10-26. m.sohu.com/a/109281900…
延伸入口
- 原文归档：tobemagic.github.io/ai-magician…
- 公众号：计算机魔术师
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。
Meta has just announced an increase in the bet on AI and higher capital spending, and stock prices have fallen and credit risk indicators have reached record highs.
Seven giants.
Income per capita: Mag 7 rising, S&P 493.
Apollo Chief Economist Torsten Slok released a set of comparative data in his report. Since 2022, the per capita income of the "Seventies" (Alphabet, Amazon, Apple, Meta, Microsoft, Nvidia, Tesla) has risen from $550,000 to nearly $800,000. During the same period, the BDP 500 removed the remaining 493 companies, which had been in the vicinity of $550,000 for seven years, drawing a horizontal line.
493 companies have turned their books and the numbers haven't changed.
Profit margin: close to 30% vs. 10%, gap widening rather than narrowing
The second group is more straightforward. The total profit of the seven giants is approaching 30%, and S&P 493 is still at 10% or less. The two lines are not moving in parallel, but at the fork. Slok did not use words such as "means" or "forecasts" but simply wrote: there is no evidence that AI has improved the productivity of these 493 companies.
AI Dividend distribution: the break-up chain from arithmetic to profit
A slight decline in total factor productivity 0.07%: macro data not captured AI
The third group is from the San Francisco Fed. By the second quarter of 2026, full factor productivity adjusted for factor utilization in the United States had declined by 0.07 per cent over the same period. No acceleration, not even a drop. AI In the hottest two years of narrative, macroproductivity indicators have neither jumped nor showed signs of impending take-off.
The three maps are stacked together, and one fact emerges: the financial returns of AI are concentrated in the seven countries that build the computing infrastructure. Token ' s expenditure is being rewritten into a cost line for the 493 traditional enterprises that are holding up a large plate and no longer count as growth investment. Slok added a footnote to this set of data: "AI may be a powerful calculator, but the intelligence that business needs, mostly judgment."
"Token Expenditure" identity shift: from growth engine to cost center
CFO perspective: bill is thick, revenue is not present.
The words of Slok can be posted directly in the audit base: “Major businesses are recategorizing Token expenditures from growth engines to cost centres because it is crushing profit margins. "It's down to the CFO table, which means your GPT-4o API annual fee last month, which has to be moved from R & D capitalization to sales costs next quarter." The MIT Nanda Laboratory study in August 2025 gave a figure and 95 per cent of the business-generated AI projects had no measurable financial returns. CFO sees this ratio, the contraction of the budget is almost conditional reflection.
VC, the start-up company's still burning, Token.
It's not the same account. The VC SaaS company uses token consumption as a growth lever, burning money for ARR is a narrative of the investment period, and losses can be told to investors. The CFO is faced with a quarterly profit guide and has no "two additional quarterly certification periods". The same model, API, is called infrastructure input in the account of the start-up firm and unrealized cost in the income statement of traditional enterprises.
Meta Press letter number: Add AI against space
Last week Meta volunteered to publish plans to add AI capital expenditure. The market was not interpreted as keeping pace, but as an additional cost sheet. Stock prices fell that day and credit risk indicators rose to record highs. A company with a market value of over $300 billion added to AI is being used as a profit erosion signal to digest. Even worse, Wall Street is already labeling AI spending, which is the cost unless you're the seven shovel sellers.
Why is a stronger model not equal to a more profitable model?
Calculating Price Reductions ≠ Battalion Growth Rate: Slok Logic
The cost of the reasoning is cut every six months, which in itself does not constitute a profit story. Slok's demonstration path is straightforward: model parameters double, token's price dropped to 30% last year, but S&P 493's per capita income did not move the $550,000 step in seven years. It's cheap, and those who buy it don't sell an extra piece. The analogy he uses is a calculator: one that counts faster than the one that says, "This one can press two points." AI may be a powerful calculator, but the intelligence that business needs, mostly judgment.
Organization, trust, interest: the friction that AI can't solve
A medium-sized manufacturer that introduced AI after the birth did save about 5% of the dispatchers, but the orders remained the same, the customers were not too many, and the competitors did not lose their orders. Five of the saved human costs were redistributed by management to other departmental budgets, and profits remained static. Slok classed such situations as "organizational frictions": inter-departmental interest games, long-term negotiations among suppliers, and customer trust thresholds for alternatives, which AI could not accelerate, but only bypass. The model is smarter than the level of trust built in the restaurant for sale. Worse still, the cost savings are internalized into a new budget line and do not automatically translate into shareholder returns.
Poor landing rhythm or capacity boundaries? Slok, skeptical reasons.
The most natural counter-argument is "the slowness of the fall of non-technology enterprises, two or three more years." Slok didn't take that assumption. He had a biweekly mass survey (covering 1.2 million enterprises) conducted by the United States Census Bureau, and found that the AI adoption rate for large enterprises with more than 250 employees was falling, not "still climbing the slopes," but rather trying, forgetting, ROI, decided not to continue. With 95% of the MIT Nanda team generating AI zero-return data, the poor-paced explanation is untenable. More precisely, most of the business bottlenecks in traditional enterprises are not the speed of calculation, but the three current generation models of judgement, trust and distribution of benefits cannot be solved.
Modelling capacity to profit break-up links.
It's not a lie. It's just nobody wants to hear it.
The previous paragraph made clear Sloc's suspicions about the "fatal rhythm." The financial season is a hard time to test the suspicion. In the next four quarters, 493 companies are pulling AI expenditures out of narratives and placing them on the income statement.
What should the financial season focus on?
AI ratio: from "optional" to "necessary"
The AI budget was finally taken out by the CFO.
In the last two seasons, "we're layouting AI" can also be inserted when a qualitative description is inserted into the 10-K risk disclosure page. Not this season. The seller ' s researcher started breaking capex and opex by line: which part is GPU cluster depreciation, which part is API call fee, and which part is internal data showing manpower. When it's done, the problem is specific. You spent $30 million on reasoning, and the corresponding AI credit account was paid. If the answer is zero, the money in the CFO table is a pure edible.
There is also a more substantial trade-off for traditional enterprises in the 493 miles. They do not need to follow up on the latest frontier models every day, they need a layer of reasoning that stabilizes into ERP and CRM. The question is no longer which model to choose, but which decision nodes the current line has changed. I can't, Token bill is cost, not investment.
Investors focus on migration: from "AI narrative" to "AI gains and losses"
Two years ago at the press conference, the CFO said, "We're increasing the AI input," and the stock price was a wave of narrative premiums. Now the analyst turns to the P&L page and asks another set of questions. The difference between the size of AI's earmarked income and AI's earmarked cost is widening or narrowing.
The seller's research framework was changed. TAM slides are missing, and AI has a lot of marginal impact sensitivity analysis. For the 493 company, which had a profit margin of 10 per cent, the net interest rate slipped from 10.0 per cent to 9.7 per cent for every dollar that it burned token and failed to collect back a dollar. On the basis of the tens of billions of dishes collected annually, the margin comes down a year to eat tens of millions of profits. You bought the flagship reasoning service, and it suddenly turned into a tablet.
One question that remains unanswered: can judgement be priced?
The "AI may be a powerful calculator, but the smartness that business needs, mostly judgment" fell on the balance sheet and became an uncharted question. Whether judgment can be amortized.
A CFO can price Token, GPU hours, SaaS seats. But the value of the customer's trust in "still opting for renewal after making the wrong decision." The chain of accountability within the organization, the playing of stakeholders, and the absence of a depreciation model.
So the real question in this season's financial paper is not "Ai spent much," but "Ai changed which decision would not have changed." If the answer is "no, just a little faster than the same decision", the nature of the expenditure is locked: operating costs, not growth investments.
Three things we can do this season. When looking at the company's financial statements, the AI capex plus opex and AI attribute income is treated as a series of margin values, which are negative and unconstrained for four consecutive quarters, and the narrative premium should be withdrawn. For traditional business, the next board of directors has a number: AI changed decision X, saved or earned Y. I can't get that number. Frankly, you bought a calculator, not a smart one. This diagnostic framework was put in place at the time of the "Accelerating Implementation" phase; if the next generation model really moves the decision boundary itself, the classification of cost centres needs to be reversed.
References
- Torsten Sløk, Apollo Global Management, "AI has delivered zero financial payoff outside the tech industry," Oct 2026. finance.yahoo.com/economy/art…
- Grunkel, "Chief Economist of Apollo: AI has not yet paid off for American enterprises outside the science and technology industry," 2026-10-25. caifuhao.eastmoney.com/news/20360...
- PaNews/ TradingView, "Leconomist Apollo: AI has not yet paid off for American enterprises outside the science and technology industry," 2026-10-25.
- IT House/Fox Search, "Chief Economist of Apollo: Apart from the seven giants of technology, AI has not yet helped American companies to earn real money and silver," 2026-10-26.m.sohu.com/a/1092800...
Extension entrance
- Original archive: tobemagic.github.io/ai-magician...
- WeChat Official Account: Computer magician
Auto-fetched and translated by AI Atlas. All rights belong to the original author.