Meta 2026 Q2 财报：广告业务仍强，但公司正在变成重资产 AI 平台¶

报告日期：30 Jul 2026

报告期：截至 30 Jun 2026 的季度

市场数据口径：29 Jul 2026 美股收盘及盘后交易

引用口径：财务数据默认单位为十亿美元；电话会页码按官方 PDF 文件页序。电话会问答原文来自文末所列的官方 PDF，按投资问题重组但保留完整提问和回答；中文内容是本文解释。

结论¶

这不是一份核心业务恶化的财报。

Meta 的广告收入同比增长 27%，广告展示量增长 14%，平均广告价格增长 12%，说明用户参与度、广告加载和广告效果都仍然健康。收入同比增长 28%，也高于市场预期。AI 对推荐系统和广告排序的帮助已经可以在收入端观察到。（Q2 财报；电话会 PDF pp. 5–8）

但这也不是一份市场应该无条件认可的财报。

Meta 正在从一家轻资产、高利润率、持续回购股票的广告现金牛，转变为一家需要投入数据中心、GPU、电力、AI 人才和第三方算力的重资产 AI 平台。资本开支和相关运营费用已经发生，但个人 Agent、企业 Agent、模型 API、开发者工具和算力销售尚未形成可量化的新增收入和投资回报。

市场不认可的核心不是本季度 EPS，而是以下不对称：

已经确定的 CapEx、折旧、算力费用、人才成本和融资需求
vs.
仍未量化的个人 Agent、企业服务、API 和算力变现

因此，本次盘后下跌的方向是合理的：Q3 收入指引略低于预期，收入增速开始放缓，利润率和自由现金流受到明显压制，管理层也没有回答 2027 年 CapEx 以及 AI 项目何时产生可量化 ROIC。（Q2 财报；电话会 PDF pp. 10–11）

但不同抓取时点约 7%–10% 的盘后跌幅可能混入了对 headline EPS miss 的机械反应。剔除法律和裁员费用后，Meta 本季度运营利润同比仍增长约 9%，粗略调整后的 EPS 约为 7.35 美元，略高于市场预期。核心广告业务没有被破坏。（Q2 财报；Yahoo Finance；Investing.com 盘后记录）

当前需要判断的不是“Meta 的财报好不好”，而是：

核心广告业务产生的增量利润，是否足以在未来两到三年内覆盖 AI 基础设施带来的折旧、运营和融资成本。

这仍然没有得到证明。

本季度发生了什么¶

财务摘要¶

数据来源：Meta Q2 2026 财报、Q2 earnings presentation pp. 4, 9, 15。市场预期来自 AP 引述的 FactSet 数据。

这份财报最重要的分化是：

Revenue +28%
Operating cash flow +25%

但

GAAP operating income -8%
Free cash flow -91%

业务仍然可以增长和产生现金，但增长产生的现金正在迅速转化为数据中心和算力资产。

核心广告业务没有恶化¶

广告业务的主要指标仍然健康：

• 广告展示量同比增长 14%。
• 平均广告价格同比增长 12%。
• Instagram 全球使用时长同比双位数增长。
• Facebook 视频使用时长同比增长 9%，美国和加拿大增长超过 10%。
• Family daily active people 为 3.60b，同比增长 3%。
• Instagram 达到 2b 日活，Threads 超过 500m 月活。

数据来源：Q2 财报；使用时长、产品里程碑与 LLM 推荐系统来自电话会 PDF pp. 1, 5–6。

广告收入增长不是单纯依靠提价，也不是单纯依靠增加广告加载。展示量和价格同时增长，说明用户参与度、广告需求和广告效果都在改善。

Meta 还披露，LLM 已被用于理解 Instagram 上的全部公开 Reels 和 Feed 内容，并用于推荐、广告排序和内容治理。AI 对核心业务的作用并不只是管理层叙事，它已经体现在使用时长、广告展示量和广告价格上。

这是本季度最重要的正面事实。

EPS miss 有多严重¶

本季度费用包含（Q2 财报；电话会 PDF p. 5）：

• 2.40b 法律诉讼相关费用；
• 1.18b 与 May 2026 裁员有关的遣散费用。

合计为 3.58b。

如果将这两项费用从运营费用中剔除：

调整后 operating income
= 18.775b + 3.580b
= 22.355b

调整后 operating margin
= 22.355 / 60.801
= 36.8%

调整后运营利润同比增长约 9.4%，与管理层披露一致；其余调整后数字为本文根据同一财报口径计算。

如果简单按照本季度 16% 的税率回加税后影响：

调整后净利润
≈ 15.848b + 3.580b × (1 - 16%)
≈ 18.855b

调整后 EPS
≈ 18.855b / 2.566b
≈ 7.35

这个粗略计算高于 FactSet 约 7.19 美元的市场预期。（AP）

所以，6.18 美元的 headline EPS miss 并不能准确代表本季度的日常经营表现。

但也不能因此认为利润率问题完全是一次性的。即使剔除 3.58b 法律和裁员费用：

调整后费用
= 42.026b - 3.580b
= 38.446b

同比增长
= 38.446 / 27.075 - 1
≈ 42%

调整后费用仍然增长约 42%，明显快于收入的 28%；调整后运营利润率 36.8%，也显著低于去年同期的 43%。

管理层明确说明，持续费用增长来自（电话会 PDF p. 5）：

• AI 技术人才薪酬；
• 更高的折旧；
• 数据中心运营成本；
• 第三方云服务；
• 第三方 AI token 成本。

这些不是一次性费用，而是 AI 战略的经常性成本。

电话会：管理层回答了什么，回避了什么¶

财报数字解释了利润率和自由现金流为什么下降，电话会则提供了管理层对资本回报、算力供需和新业务的解释。

以下不按发言顺序摘抄，而是按投资问题重组。由于同一位分析师经常同时询问两个相关问题，本文不再割裂摘句，而是保留该轮完整英文提问和管理层回答，再给出中文解释与评论。

ROIC 与 2027 CapEx：列出了多条回报路径，但没有量化兑现顺序¶

完整原始问答¶

Brian Nowak, Morgan Stanley: Thanks for taking my questions. I have two, one for Mark, one for Susan. Mark, I appreciate all the color on the big pipeline for new products across consumer and business agents, API tools, and compute rental. There’s a lot of opportunities here.

My question is, as you look at these opportunities and the state of the current offerings, the compute capacity, which of them do you expect to be able to scale first sort of in ‘26 and ‘27 to sort of showcase quantifiable material ROIC for investors?

And then Susan, there’s been some public comments about capacity in ‘27 and doubling capacity and appreciate your color about CapEx. Any early comments on ‘27 CapEx, even the philosophy around sources of upside or sources of downward pressure, as you sort of think through different ways to finance this multi-year build and have ‘27 CapEx?

Mark Zuckerberg: I can take the first one. So in terms of the different opportunities and how we think about the compute, overall a substantial amount of the compute goes towards training models to be a leading lab and I think that’s an important investment.

But then the rest of it goes towards a set of different products and revenue opportunities, which spans from optimizing and improving our core business to building new consumer products that we’re releasing soon to the API, to the business agents work, to the developer tools work on the road map that I alluded to.

And then also the opportunity to sell compute directly where I mean, I mentioned that we have quite a number of offers at a meaningful premium over what we paid for the compute. The question and thinking about this is we believe that there will continue to be a significantly higher margin on selling intelligence rather than selling compute directly.

But we think that there’s a big opportunity obviously to sell compute as well. So we’re thinking through -- I think you asked what one area would likely scale the most. But I mean, I’m actually quite optimistic that we’re going to see meaningful growth in all of these areas. And I think we will have more to share soon on a number of them.

Susan Li: Brian, on your second question, we aren’t providing a specific outlook for 2027 CapEx at this time. Infrastructure planning remains highly dynamic and even this year, there are a range of outcomes embedded in our outlook.

I alluded in my main remarks to our focus on gearing our current infrastructure plans towards maximizing capacity in 2026 and ‘27 and giving us the flexibility to continue to grow in ‘28 and beyond but also giving us the ability to make server decisions when we come to our being able to evaluate our actual needs in ‘28 and beyond.

So we’re still working through what our capacities are going to be over the coming years. Generally, we believe near term capacity is more valuable than long term capacity and it remains a very dynamic planning process.

原文位置：官方电话会 PDF pp. 10–11。

中文解释¶

Brian 的问题不是让管理层再次列举 AI 机会，而是要求回答两个可以进入估值模型的问题：

1. 消费者产品、Business Agents、API、开发者工具和算力出租中，哪一项最可能在 2026–2027 年首先规模化，并展示具有重要性的可量化 ROIC；
2. 2027 年 CapEx 的大致方向、上行和下行因素，以及融资选择会如何影响建设节奏。

Mark 没有选择单一业务。他把算力分为训练 frontier model 和支持各类产品及收入机会两部分，并认为广告优化、消费者产品、API、Business Agents、开发者工具和算力销售都会增长。他强调出售 intelligence 的长期利润率高于直接出售 compute，但没有说明哪条路径会最先形成多大收入。

Susan 则明确拒绝给出 2027 年 CapEx 数字。Meta 当前优先最大化 2026–2027 年容量；对于 2028 年以后，公司先准备土地、电力和建设条件，等更接近需求发生时再决定是否购买服务器等高价设备。

评论¶

正面之处是 Meta 并非只有一个未经验证的 AI 产品可以消耗算力。现有广告系统、消费者产品、企业产品和对外出售算力共同构成多个回报路径；远期服务器采购仍保留调整空间。

负面之处是，Brian 要求的是“哪一项先规模化”和“quantifiable material ROIC”，管理层给出的却是“多个方向都会增长”。公司没有披露各业务收入、合同、增量利润率、算力分配、回收期或 2027 年支出上限，因此回答仍无法进入估值模型。

企业 GTM 与融资：Business Agents 可以复用旧能力，其余业务需要新肌肉¶

完整原始问答¶

Eric Sheridan, Goldman Sachs: Thanks for taking the question. Maybe two, if I can. When you frame up the enterprise opportunity, how much of that opportunity do you think is available to you today based on what you’ve built out in terms of go to market strategy as extensions of the advertising and the marketing business you have already versus go to market strategies that might have to be built to capitalize on the opportunity?

And then maybe Susan, if I could just squeeze a second one in. When you think about sources of capital for the business for the next couple years, you referenced the deal that was announced yesterday as an example of looking at ways to finance forward obligations.

We continue to get questions from investors about how to think about the mix of debt and equity and sources of capital. Philosophically, how are you guys thinking about wanting to be ambitious on the spend, but then marrying that with the need for capital? Thanks so much.

Mark Zuckerberg: Sure, I can talk about the first part. I think for enterprise, there’s going to be a combination of extending the current business, which is effectively selling to marketers and businesses that are basically customer facing, trying to reach customers and sell to them directly.

Obviously that’s the vast majority of the business across Facebook and Instagram. And we believe that there is an opportunity to extend this with business agents across messaging apps and other services to continue to interact with customers.

And just like the ad system, effectively, we will get paid when we deliver results for those businesses. So we view this as an extension of the sales and the partnerships that we have with many millions of advertisers and hundreds of millions of small businesses that use our platforms.

So that one I think should be quite a natural opportunity for us. And we’re focused on delivering it in a way where we’re not trying to kind of maximize the sale in the near term. We’re trying to maximize the results for people and build out a robust auction. And that’s what we’ve seen has served the business well over time.

There are other enterprise customers who I think we’re increasingly going to serve too. We’re building coding and developing and internal productivity tools partially because we need to build them ourselves and we need to make sure that we have tools that are tuned for ourselves.

And now that we have those, we feel like there’s a large opportunity to serve, whether that’s small businesses or larger businesses. That is a somewhat different muscle than we have historically had. And we will share more soon on how we’re planning to build that out.

But I think that there is just a very large opportunity on this. And the way that I think about this is there’s obviously been a bunch of news about the compute side. But I think that the enterprise opportunity is kind of the sum of all of these different things. It’s not just the selling compute, also the API services, the productivity services, the kind of business agents for other parts of the business beyond marketing are all parts of the overall offering.

And I think that there’s just a very, very large opportunity there. So we’re quite focused on that. That’s going to be somewhat of a new muscle that we build as a company, but I think it’s a very important one that we build so that way we can make sure that we can maximize the opportunity ahead of us.

Susan Li: Eric, on your second question in terms of sources of capital, this is something that we have been looking at thoughtfully as we think about financial planning for the future.

Obviously, our strong operating cash flow certainly has put us in a position of strength as it pertains to funding our infrastructure build out. But we’ve also been evolving our capital structure in recent years to include a greater mix of debt as we work to bring down our cost of capital.

And we have generally found it prudent to continue adding cost efficient long duration sources of capital as we make investments in initiatives that themselves have long time horizons, especially AI infrastructure projects.

We’ve also broadened our aperture there to include partnerships like the one we announced with BlackRock yesterday. And we’ll continue to be thoughtful about evaluating the appropriate different sources of capital over time as we evaluate future projects.

原文位置：官方电话会 PDF pp. 11–12。

中文解释¶

Mark 将企业业务分为两类：

• Business Agents 是广告和消息业务的自然延伸。Meta 已经服务数百万广告客户和数亿小企业，可以继续按给企业带来的结果收费，并建立类似广告系统的结果拍卖；
• coding、development、productivity、API 和更广泛的企业内部工具需要新的销售与交付能力。Meta 因内部需求先做出这些工具，但向企业客户销售并不是公司过去的核心能力。

Susan 的融资回答则说明，经营现金流仍是主要资金来源，但公司近年来主动增加债务以降低资本成本，并会用更长期限的资金匹配 AI 基础设施的长期回报周期。BlackRock 合作是外部资本补充自身资产负债表的一个例子。

数据证据¶

本季度已有超过 1m 家企业每周使用 Business Agents。Meta 还介绍了巴西租车公司 Movida 的案例：客户可以在 WhatsApp 对话中完成选车、定价和付款；该公司报告 WhatsApp 日均订单同比增加 44%，其中 85% 的对话由 AI Agent 独立解决。官方电话会 PDF pp. 3, 7–8

长期债务在半年内增加约 24.9b，上半年停止股票回购，同时 CapEx 达到 50.9b。（Q2 财报）

评论¶

Business Agents 与 Meta 的消息入口、广告客户关系和效果付费能力相邻，可能是新增 AI 业务中商业化路径最自然的一项。Movida 案例也说明 Agent 可以完成交易闭环，而不只是客服问答。

但一个案例不能证明产品整体的留存、付费和利润率。coding 和 productivity tools 还需要企业销售、采购、权限安全、审计、服务等级和客户成功能力。在缺少客户及收入数据时，不应给予与 Azure 或 Microsoft 365 相同的估值。

融资能力降低了短期资金约束，却不会消除资本成本。更多债务和项目合作反而说明管理层预计高投入将持续较长时间；上半年股票回购已经让位于基础设施建设。

个人 Agent：机会很大，但产品尚未交付¶

完整原始问答¶

Mark Shmulik, Bernstein: Yes, thanks for taking the question. Mark, everyone’s got a story of someone coming back from Silicon Valley, deep writing code, building agents with AI, and then they head home, and they kind of tell their parents they’re using AI wrong, as just kind of like a glorified search tool.

Consumer behavior is always pretty difficult to predict. But kind of reading your op-ed on AI for everyone, how do you think about whether consumer adoption can close this AI utility gap? Like are we on the cusp of something kind of breaking through or do we just need to be a bit more patient? Thanks.

Mark Zuckerberg: Well, I think that some things have already broken through. And I think one of the interesting things about AI compared to other technologies is that every year or so, and the cycle may accelerate, but every year or so you get new capabilities that create new possible product lines.

So there’s obviously the AI assistant market for consumers where that’s what we’re doing with Meta AI. And then there’s the other products that competitors have in that space. And then in the last year, I think the coding agent market has really grown very quickly and that’s the first real agentic market.

And I think there’s a number of reasons why coding is first, right. You have technical customers who are willing to do the work to make it work. Coding is an inherently digital and closed loop activity. So it’s sort of, in some ways it is somewhat of a, kind of an easier system to train on. But our bet is that -- or one of the bets here that we’re making is that we think that consumer personal agents is going to end up being an extremely important and massive market.

I think that it’s extremely unlikely if you look out five years from now, for example, that -- or whatever period of time you want, that you don’t have billions of people with a personal agent that understands your goals, and that is just working on your behalf 24/7 to achieve your goals.

And whatever the domain is that you care about, whether it’s helping you with your health or your hobbies or your personal finances or your productivity and running your home better, or improving and enhancing your relationships, helping with your career -- just all these different things.

This is like a very, very deep set of use cases. But like you said, I think that building for consumers is a little bit different than building for developers. If you’re building for consumers, especially if you’re trying to build something that isn’t used by millions of people, but is used by billions of people, it needs to just work, right? It needs to be simple.

And I think a lot of what we see with these agents today, a lot of the kind of proto-agents is, it takes a bunch of fiddling. You have to get into a terminal to get it set up. You get a use case going. It kind of seems magical, but then maybe it breaks down over time. And I think that the companies that can deliver the personal agents that just work, I think that this is going to be one of the next major opportunities in AI.

And we are very excited about that in addition to what we’re doing with Meta AI, in addition to what we’re doing with business agents, in addition to all these other things, while also developing the models that we think are going to continue to advance new capabilities to create new product lines as well.

So this is why we’re very optimistic. Now, I get we’re going to ship this at some point soon and that’s going to be very exciting. And we haven’t done that yet. So this is -- there’s kind of only so much that I can -- that I can say on an earnings call about this.

But I think this is a very large opportunity and one that I think is almost inevitable that someone does. And I think it really plays to Meta’s strengths as a company. We build consumer products that reach billions of people.

We’re great at once we get something working, scaling it to a large number of people, and we’re good at building the infrastructure to be able to support these intensive applications. So, I feel very good about this. I kind of understand that we need to deliver it for our community and that’s what we’re very focused on.

原文位置：官方电话会 PDF pp. 12–14。

中文解释¶

Mark 认为 coding agent 已经成为第一个快速增长的 Agent 市场，因为用户技术能力更强、任务完全数字化并且反馈闭环清晰。个人 Agent 可能是下一个巨大市场：它会持续理解用户目标，并在健康、家庭、财务、生产力、关系和职业等方面工作。

但消费者产品必须开箱即用、简单且长期稳定，不能依靠终端配置，也不能运行一段时间后失效。Meta 的消费者产品经验、数十亿用户分发和基础设施能力适合这个机会，但 Mark 明确承认相关产品尚未交付。

评论¶

Meta 的消息入口、社交关系、用户规模和消费者分发能力确实适合个人 Agent。产品若成立，公司不需要从零购买流量。

但管理层自己承认产品尚未完成。个人 Agent 目前属于产品期权，不属于已验证收入，不能用于解释 2026 年已经发生的折旧和资本开支。

为什么一边购买算力，一边考虑出售算力¶

完整原始问答¶

Douglas Anmuth, JPMorgan: Great thanks for taking the questions. One for Susan, one for Mark. Susan, you talked about how LLMs are increasingly capable of delivering ranking and recommendation gains. Can you just talk more about the roadmap here and how far along you are in just leveraging better models and more compute?

And then, Mark, just in terms of the number of offers to monetize your compute externally, you also at the same time are purchasing capacity from a number of third parties.

So just hope you could help us understand some of the differences here? Is it just timing and stop gap issues? Or is it training versus inference, and leveraging the chips that are best suited for each?

Susan Li: Thanks, Doug. I’ll take that first question about where we are on the recommendations roadmap. First of all, we certainly see further headroom to continue improving recommendations over the rest of the year and into 2027. And we expect that will help us drive additional gains on both engagement on Facebook and Instagram. A couple of things I would highlight.

First, we’ll continue to make recommendations even more personalized and relevant to user interest. And that’s in part by advancing our recommendation models and architectures to further capture user interest more precisely and respond faster to what people care about in a given moment.

Our AI investments are going to play a significant role in delivering on this vision including the expansion of LLM based content understanding to develop a deeper understanding of posts and creators that people value, to capture user interests more precisely and respond more quickly to what they care about in the moment and using AI to surface high-quality, fresh and trending content and reduce the share of low-quality content.

Second, we’re continuing to improve our data infrastructure to allow our models to train on more data and leverage that data more effectively.

We’re adding more detail to how we describe content that users have engaged with in the past and enriching past user interaction sequences with more granular content. That allows our models to more precisely learn which engagements are more or less valuable to users.

We’re continuing to scale up both the length of user interaction sequences we use during training as well as the complexity of our model architectures across Facebook and Instagram to take advantage of the larger data sets. And then we’ve already made significant strides, leveraging LLMs for content understanding.

We’re going to further incorporate them into both our recommendations and content policy enforcement stacks given their capability to more deeply understand content. And I would say, broadly, I think we’re very optimistic about that body of work.

We’ve also invested in using LLM-based agentic approaches to transforming our recommendations system. And we grew the number of launches from our ranking agents this half. That also helps make our engineers more productive and that’s another path that we’re very excited about as well.

Mark Zuckerberg: I can answer the second part of that. I mean I think your question was about how we think about offers that we’re getting to sell the compute, but then we’re also buying the compute. I mean look, the high-level observation is that there’s just nowhere near enough compute for all the demand.

So that is why we see that like basically, we are getting a large number of offers for the compute that we have, but also we have a lot of internal uses that we think are going to be quite valuable.

Now in terms of running the business, obviously a common trade-off that we need to make is around how much do you monetize something today versus develop future assets for the future? And I think that it’s always a portfolio, right? It’s not like -- you don’t want to only do long-term things and not kind of prove the markets out that exist in the near term.

But I also think it would be foolish to basically just sell all of the compute and take a short-term profit. But when you have the opportunity to build intelligence on top of it which will be a multiple and that compounds the value of the compute on top of that.

So I think the answer is what we’re doing which is to basically use a lot of our capital to build out compute, having confidence that we have the ability to monetize the compute directly when that makes sense, but also knowing that we have quite a number of different use cases to monetize the intelligence on top of the compute including the enterprise cases that we talked about and including some of the consumer cases that we talked about, and including just the core business which is not even necessarily new products that we haven’t talked about, but just in terms of using that to be able to further add intelligence and improve the ranking and recommendations and ads in the core services.

So I think all of that is true, and it creates this dynamic where there is a lead time where we’re investing in building out these data centers now. They come online at some point in the future. You obviously are not getting value out of them until they’re online.

But that’s -- we basically see a very large demand for all of this, and want to go maximize the opportunity to build out all of these different businesses.

原文位置：官方电话会 PDF pp. 14–16。

中文解释¶

Doug 的第一个问题确认现有广告和推荐系统是否仍能吸收更多模型与算力。Susan 的回答是肯定的：Meta 会继续提高推荐模型对用户兴趣和内容的理解，增加训练所用交互序列的长度和粒度，把 LLM 用于推荐及内容治理，并用 ranking agents 提高工程效率。

第二个问题指出 Meta 同时买入和出售算力的表面矛盾。Mark 没有按训练、推理或芯片类型拆分，而是从总体供需回答：行业算力远远不足，Meta 同时面对有吸引力的外部报价和大量内部用途。

公司需要在近期变现和建设长期智能资产之间做组合配置。直接出售算力可以产生近期利润并验证市场；在算力上训练模型、改善广告和建设 Agent，则可能形成更高利润率并复合放大算力价值。因此 Meta 不会出售全部算力。

评论¶

正面之处是外部报价为阶段性富余容量提供了变现出口，现有推荐与广告系统也能立即使用更多算力。部分数据中心若先于新产品投产，公司仍可能出租算力，降低闲置风险。

负面之处是外部报价证明的是行业算力紧缺，不是 Meta AI 产品需求。公司没有披露报价方、合同期限、容量、收入和利润率，也没有回答训练与推理、不同芯片或时间错配分别占多少。算力外售更适合作为闲置风险缓冲，而不是当前估值中的主要新增收入。

高性能、低成本与开放模型：Meta 为什么两端都做¶

完整原始问答¶

Ross Sandler, Barclays: Yes, hey Mark. Sticking with the comments on the AI Lab. So Muse Spark 1.1 is pretty close to the Pareto Frontier, but it’s at the lower end of the intelligence spectrum or lower cost end, I should say. And it sounds like from your last answer, you think it’s important to compete at both the lower cost and the more expensive performant end.

Could you just talk about that a little bit? And then the leadership of your lab has also talked about going back into open source, kind of where you were a couple of years ago. So how does that fit into the strategy and all this discussion around monetization of AI products and models? Thank you very much.

Mark Zuckerberg: Yes. So I mean just to get into the kind of scientific process on this a little bit, basically, training large models, at each stage of training a large model, you see novel behaviors.

So basically, the way that you want to do this is you build up from training smaller models to building larger models. And the models that we’ve released so far are based on -- they're a certain scale in climbing up the scaling ladder and we’re continuing to scale larger models.

So I think for Muse Spark 1.0 and Muse Spark 1.1, I think that they are very impressive models for the scale of the model, the kind of stage of development of the lab. And I feel quite good about them.

Now we want to have models that are more advanced as well and that’s why we’re scaling larger models, and that’s why we’re building out a bunch of the research infrastructure around that.

But at the same time we also want to have good, more efficient models that will be a lot of what we serve to consumers at scale, right? If you’re serving billions of people, you want the ability to have more advanced models for things that are very hard problems and you want the ability to serve the vast majority of prompts from simpler and more efficient models.

So I think both are going to be very important. The efficiency really matters, but I also think that we want to be able to solve the hardest problems for businesses and customers around the world. So we care about both of them.

On open source, I think we have always felt like open source was an important part of the ecosystem, and it’s good for the world, and it creates its own feedback loops that are positive for us around getting the community invested in our infrastructure stack and our work and contributing improvements. But we’ve always basically said that we were going to do a mix of open and closed. And that continues to be true.

Now in ramping up Meta Superintelligence Labs, in some ways, actually counterintuitively, it takes some more work to do open source models because you’re not necessarily -- if you’re doing something as a closed system that you’re only building for your own use cases, it can be a little more jagged.

Whereas if you release it as open, and it’s going to be used for a lot of things, you want to make it more well-rounded. And I just wanted to make sure that the MSL team was uninhibited in building out the most intelligent models that we could. And we expect that we will get back to releasing some open source models at some point soon.

But like we’ve always said, we’re not dogmatic about this. We think open source is important. We want to contribute to that ecosystem. We plan to do a combination of open and closed models.

原文位置：官方电话会 PDF pp. 17–18。

中文解释¶

Mark 将更强的大模型和更低成本的高效模型视为不同用途。最困难的企业和消费者问题需要先进模型，但数十亿用户的大多数请求必须由更简单、更高效的模型处理，否则推理经济性无法成立。

开放模型可以吸引社区采用 Meta 的基础设施栈并贡献改进，但为了让模型适用于更多场景，开放发布需要额外的完整性和打磨工作。Meta Superintelligence Labs 初期先确保研究团队不受发布要求限制，未来仍计划恢复部分开放模型，同时保留封闭模型。

评论¶

这个回答解释了模型组合与推理成本之间的关系，也说明 Meta 的 open-source 策略从“默认开放”转向“按战略选择开放与封闭”。

但电话会仍没有披露 Muse 的训练成本、不同模型的单位推理成本、API 收入、推理毛利率或开放生态带来的可量化回报，模型投资依然难以通过现金流估值。

模型主权与 2026–2027 年容量：最后一轮追问给出的完整答案¶

完整原始问答¶

Kenneth Gawrelski, Wells Fargo: Thank you. Two, if I may, please. I just want to maybe, Mark, just touch on the last point again, around open weight models. There’s been a lot of discussion, you’ve weighed in on the topic.

Why or why not does that change Meta’s view of developing closed proprietary frontier models? Is there an opportunity if open weight models proliferate that you don’t have to develop your own frontier model? So that’s question one.

And the second one, a clarification, if I may, for Susan. You noted that in your prepared remarks that you plan to maximize ‘26 and ‘27 capacity, is that a demand or a supply comment? Meaning, are you suggesting that Meta plans for internal use of all the capacity built through ‘27? Or are you saying that you will evaluate ‘28 and beyond builds based on demand for Meta products and services? Thank you.

Mark Zuckerberg: I can take the open source question. Let’s see. So basically the question is, do we think that because there are some open weight models that we can just rely on those. I mean right now the open source models are not as strong as the frontier models. So no is the basic answer.

And then there’s also just always the perpetual both policy debate and question around other companies’ actions and whether that’s actually a thing that we can -- that a company like Meta can rely on. And I think that that’s very tricky.

So I think on both fronts, we believe we’re going to be able to do better work, and we think that there’s some risk in that reliance, I don’t believe that that is the right thing to do.

I think that we’re a company that -- if you look at Meta from -- take a step back on this. A lot of people view the surface layer of we build some social media apps and we have an ad business. We are really a full stack technology company.

We built our own data centers, our own infrastructure, our own chips, our own low-level software. When we got started -- like my background in engineering, like I wrote a lot of the systems code. A lot of the reason why Facebook worked was because it actually -- it just worked, right?

Like it literally worked when other social networks did not work fast and efficiently. And I think we just have the ability to build things that can be more personalized, more optimized, more efficient.

Some qualitative experiences are just not even possible for others to build because we go all the way down the stack. And it just seems to me pretty clear that having kind of sovereignty over building your own models is going to be an important part of that stack going forward which is why it is important for Meta.

But is also important -- is also why other people care about open source and why open source matters overall.

Because like other companies, even if they don’t have the ability to build these models, do not want to have to just rely on a small number of closed labs. Like I think that there is a very important place in the world for there to be open source models.

And to be clear, that doesn’t take away from the API opportunity or any of the things that I’m talking about because someone still needs to run the models and run inference on them and be able to kind of run those models efficiently and have the compute to do that is going to continue to be a source of business advantage. So I don’t really think that these things are necessarily at odds.

But when I look at like what a lot of discerning customers and companies are going to want around the world, they’re going to want to know that they have control of their destiny and that they can trust the models that they’re using and that their data is safe and that they’re not sending it to competitors and all that.

So I think open source is going to be important. But I think for us, also building the models is going to be a critical part of it.

It also -- for what it’s worth, and I know that like the models all get evaluated on like a common set of evals and they get chalked up to like, okay, some models within like a couple of points of another model or whatever.

But they all really do have different combinations of skills, kind of like people, right? It’s like people have spikes in certain areas and have different personalities and are better and worse at different things.

And if you’re trying to build personal superintelligence for people or you’re trying to build a business agent for small businesses, that needs to have potentially some different skills than what the other labs are tuning. And just like being able to do the full stack work on our -- on like Instagram recommendations on our ad system is how we’ve gotten the results there over time.

I believe that building the kind of full stack model is going to be a lot of the advantage over time in how we build personal superintelligence agents, business agents, all these different use cases for all of the different customers that we want to serve.

So in addition to the distribution that we have and the ability to reach all these people and scale products that work, I think this is a lot of the durable advantage is that you build something that is kind of specific to that use case and excellent in those use cases.

And I mean look, I get that this is a big investment and it’s a big bet. We see the technology working. We’re happy with the trajectory of the lab.

I’m excited about the products that are coming, and we believe that this is going to be a big thing. So I mean I get that this is sort of a big bet across the industry. My personal bet is that the people who invest in this are going to be rewarded and feel very good over time.

Susan Li: Ken, I’ll just quickly answer your second question. When we refer to focusing on 2026 and ‘27 capacity, there are really two factors.

One is we are today, and expect to be in the sort of foreseeable future, supply constrained, and that really includes our core business, too, where there are -- we still have numerous ROI positive places that we would put compute toward if we had it.

And the second, of course, is just the uncertainty over long-term constraints on building capacity, and we talked about some of the sort of the need to build out more of the supply chain earlier in my comments.

So I think beyond ‘27, when we look into ‘28 and further than that, the world is going to evolve a lot, our own internal demand will evolve.

We’ll have turned over a lot of cards by then. And so when we think about planning today for ‘28, we’re really focusing on flexibility. That’s just giving us kind of the ability to have land and power.

But to really make the actual decisions about buying chips and other big-ticket items further in the future. So for now I think we know we have a lot of good use cases for capacity in ‘26 and ‘27, and that’s really what we’re building toward.

原文位置：官方电话会 PDF pp. 18–21。PDF 脚注说明，Susan 在电话会现场把此处说成了 demand constrained，正式逐字稿已更正为 supply constrained。

中文解释¶

Ken 将两项最重要的不确定性直接连在一起：

1. 如果开放权重模型继续进步，Meta 是否可以停止承担 frontier model 研发成本；
2. 最大化 2026–2027 年容量究竟意味着内部需求足够强，还是只是供应建设受限。

Mark 的回答是否定依赖外部模型。他认为当前开放模型仍弱于 frontier model，外部公司将来是否继续开放也受政策与商业选择影响。Meta 把自己定义为全栈技术公司，需要控制数据中心、芯片、底层软件和模型，才能针对个人 Agent、Business Agents、广告与推荐系统进行全栈优化。开放模型和 API 并不冲突，因为模型仍需要算力、推理服务和高效运行。

Susan 则把容量规划拆成两个因素。第一，Meta 当前及可预见未来仍受供给约束，连核心业务内部也有许多 ROIC 为正、但缺少算力而无法实施的项目。第二，长期供应链和基础设施建设存在不确定性。因此公司为 2028 年以后先准备土地和电力，但把芯片等高价设备的最终采购决定留到更接近需求发生时。

评论¶

这轮回答解释了 Meta 为什么不能简单外包模型，也解释了 2026–2027 年建设并非只服务尚未发布的个人 Agent：核心广告和推荐业务仍有可投放的正回报算力需求。

但“ROI positive”并不自动等于高于 Meta 的资本成本，也不说明回报足以覆盖当前建设规模。管理层仍未披露这些内部项目的边际收入、节省成本、所需算力或回收期。

模型主权的战略逻辑是完整的，经济回报仍未量化。电话会没有给出：

• Muse 模型训练成本；
• API 收入和推理毛利率；
• 自研芯片相对外购 GPU 的成本改善；
• frontier model 对广告收入的增量贡献；
• 开源生态的可量化回报；
• 2027 CapEx、峰值年份或中期 FCF 目标。

电话会的核心信息¶

电话会因此强化了而不是消除了市场的核心分歧：

管理层证明了算力存在很多潜在用途，
但没有证明这些用途产生回报的速度和规模。

为什么市场不认可¶

Q3 指引略低于预期，同时收入增速放缓¶

Meta 给出的 Q3 2026 收入指引为 61b–64b，中点为 62.5b。市场预期约为 63.14b。（Q2 财报；AP/FactSet）

以 Q3 2025 收入 51.24b 为基准（Q3 2025 财报）：

即使最终达到指引中点，收入增速也会从 Q2 的 28% 降至约 22%。指引还包含约 1% 的汇率逆风，但方向仍然是收入增长放缓。

对普通公司来说，22% 的收入增速很好；但对正在将 CapEx 接近翻倍的 Meta 来说，市场期待的是 AI 投资带来收入加速，而不是费用和资本开支加速、收入增速下降。

CapEx 高位不是暂时现象¶

Meta 2025 年全年 CapEx 为 72.22b。最新 2026 年指引为 130b–145b。（2025 全年财报；Q2 财报）

Q1 时的指引是 125b–145b，本季度调整为 130b–145b。上限没有继续提高，但下限上升，说明高额投入越来越像已经确定的资本承诺，而不是可以随时取消的预算。（Q1 财报；Q2 财报）

管理层在电话会上表示，目前的基础设施计划以最大化 2026 和 2027 年容量为目标；对于 2028 年以后，才会主要保留土地、电力和网络的选择权，并推迟服务器采购决定。（电话会 PDF pp. 8, 10–11, 20–21）

这意味着 2026 和 2027 年仍然处于确定性较高的投入期。市场无法确认 2026 是 CapEx 峰值，也无法确认自由现金流会在 2027 年恢复。

自由现金流和资本回报模式发生变化¶

本季度经营现金流为 31.86b，CapEx 为 31.08b，自由现金流只剩 0.78b。（Q2 财报；earnings presentation p. 15）

资产负债表和资本配置也在变化：

• 长期债务从 2025 年末的 58.74b 增加到 83.66b，增加约 24.92b；
• 2026 年上半年发行长期债务净额约 24.91b；
• 2026 年上半年没有回购股票，去年同期回购约 22.92b；
• 现金及可交易证券为 90.26b，但扣除长期债务后，净现金只剩约 6.60b；
• 公司开始通过 BlackRock 等外部合作伙伴共同融资和建设长期 AI 基础设施。

资产负债表、现金流及回购数据来自Q2 财报；融资思路和 BlackRock 合作来自电话会 PDF pp. 8, 12。

过去的资本循环更接近：

广告收入 -> 经营现金流 -> 股票回购

现在正在变成：

广告收入 -> 经营现金流 -> AI 基础设施
                         -> 增加债务
                         -> 引入外部资本

这个变化不代表公司没有能力投资，而是意味着股东获得现金回报的时间被推后，而且未来价值更加依赖资本回报率。

新 AI 收入仍然缺乏规模证据¶

Meta 展示了多个可能的 AI 收入来源：

• Meta One 订阅；
• Muse Spark API；
• Business Agents；
• 个人 Agent；
• coding 和 productivity tools；
• 直接出售算力；
• AI 眼镜；
• AI 对推荐和广告系统的改进。

其中最有现实基础的是核心广告优化，其次是 WhatsApp 付费消息和 Business Agents。

本季度 FoA other revenue 首次达到 1.01b，同比增长 73%；超过 1m 家企业每周使用 Business Agents；在 Movida 的 WhatsApp 案例中，85% 的对话由 AI Agent 独立解决。这些都是积极信号。（电话会 PDF pp. 3, 5, 7–8）

但相对于 59.36b 的季度广告收入，这些新增收入仍然很小。

分析师在电话会上直接询问：2026 和 2027 年哪一个业务最早可以展示可量化、具有重要性的 ROIC？

管理层没有给出单一业务、收入规模、利润率或时间表，只表示多个领域都有望增长，未来会分享更多信息。（电话会 PDF p. 10）

个人 Agent 尚未发布；企业 productivity 和 coding tools 需要建立 Meta 过去不具备的企业销售能力；直接出售算力虽然存在外部报价，但管理层认为在算力上构建智能的利润率更高，因此不会为了短期利润大量出售算力。（电话会 PDF pp. 12–16）

这形成了当前最核心的不确定性：

Meta 有很多可用算力的方式，
但没有说明哪一种方式可以最先形成足够大的高利润收入。

市场并不是反对所有 AI CapEx¶

Microsoft 与 Meta 在同一天发布财报。两家公司都在大幅增加 AI 投资，但市场反应相反：

来源：Axios 对同日财报和初始盘后反应的比较。

两者的区别不是 Microsoft 不花钱，而是 Azure 和企业软件已经可以展示客户需求、收入和利润增长；Meta 的新增 AI 收入仍然主要是产品路线图。

所以市场不是简单地说“CapEx 越低越好”，而是在判断：

新增 AI 收入和利润
是否已经足够覆盖
新增折旧、运维、人才和融资成本。

Meta 暂时无法证明这一点。

管理层指引的隐藏含义¶

管理层继续预计 2026 年 operating income 高于 2025 年。（Q2 财报；电话会 PDF p. 9）

表面上这是一个积极承诺，但需要结合基数理解；以下数字来自2025 全年财报和Q2 财报：

2025 operating income = 83.276b
2025 H1 operating income = 37.997b
2025 H2 operating income = 45.279b

2026 H1 operating income = 41.647b

要让 2026 全年 operating income 刚好超过 2025，2026 H2 只需要略高于：

83.276b - 41.647b = 41.629b

这比 2025 H2 的 45.279b 低约 8%。

换句话说，“2026 运营利润高于 2025”并不意味着下半年利润会加速增长。即使 2026 下半年运营利润同比下降约 8%，公司仍然可以完成这项指引。

市场因此不会仅凭这一句话相信利润率已经见底。

法律费用是否真的一次性¶

本季度 2.40b 法律费用可以在分析经营效率时回加，但不能完全当作与公司无关的一次性事故。

Meta 在财报中继续提醒（Q2 财报；电话会 PDF p. 9）：

• 多个市场正在加强对青少年相关问题的审查；
• 2026 年美国安排了多项青少年相关诉讼；
• 这些案件最终可能产生重大损失。

因此，单笔 2.40b 费用可能是一次性的，但法律和监管风险具有重复发生的可能。估值时可以将其从核心运营利润中剔除，但不能把未来法律成本永久设为零。

我的判断¶

市场判断正确的部分¶

市场正确地识别了以下风险：

• Q3 收入增速预计从 28% 放缓至约 22%；
• 经常性 AI 费用增速明显高于收入；
• 2026 和 2027 仍处于高强度基础设施投入期；
• 自由现金流暂时接近归零；
• 股票回购让位于 CapEx 和融资；
• 新 AI 收入缺少规模、利润率和时间表；
• 管理层没有提供 2027 CapEx 和可量化 ROIC。

这足以支持估值折价。

市场可能反应过度的部分¶

市场可能低估了以下事实：

• 广告收入仍增长 27%，而且展示量和价格同时增长；
• AI 推荐已经改善 Instagram 和 Facebook 的使用时长；
• 经营现金流仍同比增长 25%；
• 剔除法律和裁员费用后，运营利润仍同比增长约 9%；
• 粗略调整后的 EPS 约为 7.35，并不低于市场预期；
• Meta 拥有 3.6b 日活用户和数百万广告客户，是分发个人 Agent 和 Business Agents 的现实优势；
• 如果短期算力出现富余，公司称可以按高于成本的价格对外销售，算力并非完全没有退出路径。

因此，这份财报不足以证明核心投资逻辑被破坏。

更准确的判断是：

Meta 的广告护城河仍然存在，但公司的资本回报模式正在改变，AI 期权需要用更高的资本成本和更低的确定性进行估值。

多空逻辑¶

看多逻辑¶

• AI 推荐继续提高使用时长、转化率和广告单价，使广告收入维持 20% 以上增长。
• WhatsApp、Business Agents、订阅和 API 从较小基数快速增长。
• Meta 的消费者分发和广告客户关系，使其可以低获客成本推广个人和企业 Agent。
• 2026–2027 建设完成后，CapEx 增速下降，经营现金流继续增长，自由现金流迅速恢复。
• 富余算力可以出售，或通过第三方合作降低资产负担。
• 盘后下跌已经部分反映 CapEx 和利润率风险。

看空逻辑¶

• 广告收入增速继续回落，但 AI 费用、折旧和第三方云成本保持高增长。
• 个人 Agent 和企业工具迟迟无法产生规模收入。
• Meta 为维持 frontier model 能力持续增加训练投入，但模型没有形成差异化产品。
• 2027 CapEx 继续上升，市场始终看不到自由现金流恢复时间。
• 企业销售、云服务和开发者工具不是 Meta 的既有能力，新增业务兑现慢于预期。
• 法律、青少年保护和监管成本重复发生。
• 债务和外部融资持续增加，回购长期无法恢复。

后续验证指标¶

后续财报不应该只看 revenue 和 EPS，需要同时跟踪以下指标。

核心广告是否继续承担投资¶

• 广告收入增速；
• 广告展示量增速；
• 平均广告价格增速；
• Instagram 和 Facebook 使用时长；
• FoA operating margin。

如果广告收入仍保持 20% 以上增长，但 FoA operating income 持续下降，说明新增 AI 费用正在超过核心业务回报。

AI 新业务是否开始形成独立收入¶

• FoA other revenue；
• WhatsApp paid messaging；
• Business Agents 活跃企业数和付费方式；
• Meta One 订阅数量；
• Muse API 使用量和收入；
• coding/productivity tools 的企业客户；
• 对外算力销售收入。

管理层只有开始披露收入、客户数、usage 或利润率，AI 新业务才从叙事变成可估值资产。

CapEx 是否进入可控阶段¶

• 2027 CapEx 指引；
• quarterly CapEx；
• 折旧增速；
• 第三方云和 token 成本；
• operating cash flow 与 CapEx 的比例；
• 自由现金流；
• 新债务和外部融资。

最好的组合是：

广告和 AI 新收入保持增长
+ CapEx 增速下降
+ operating margin 企稳
+ FCF 恢复

最危险的组合是：

收入增速放缓
+ CapEx 继续上调
+ 折旧和 token 成本加速
+ 新 AI 收入仍不披露

可证伪条件¶

当前判断是“核心业务仍强，但 AI 资本回报尚未证明”。以下事实会改变这个判断。

支持看多、降低不确定性的证据：

• Q3 或后续收入持续达到指引上沿；
• 调整后 operating margin 连续改善；
• Business Agents、API、订阅或算力销售形成数十亿美元季度收入；
• 2027 CapEx 不再显著增加；
• 自由现金流恢复，同时广告增速没有明显下降。

支持看空、证明资本回报恶化的证据：

• 广告收入增速降至 15% 以下；
• 广告量价增长同时放缓；
• 2027 CapEx 再次大幅上调；
• FoA operating income 持续下降；
• 债务继续快速增加且回购长期为零；
• 新 AI 产品仍然只有用户数，没有付费收入和利润率。

最终观点¶

这次财报后的正确问题不是“跌了这么多是否可以买入”，也不是“EPS miss 是否代表公司变差”。

应该问：

如果今天重新拿现金评估 Meta，
我愿意用什么价格购买一个仍在高速增长的广告现金牛，
同时承担一个每年 130b–145b CapEx、
回报周期和新增收入仍不明确的 AI 建设计划？

本季度能够确认的是：

1. 广告业务仍然强，AI 对核心业务已经有回报。
2. EPS miss 主要由法律和裁员费用造成，headline 比实际经营情况差。
3. 即使剔除一次性费用，成本仍然增长过快，利润率压力是真实的。
4. CapEx、债务和外部融资已经发生，自由现金流和回购受到挤压。
5. 新 AI 收入仍不足以验证这笔投资的 ROIC。

因此，我不会把这份财报定义为核心逻辑破坏，也不会仅因为盘后下跌就把它定义为机会。

当前更合适的定位是：

Meta 的核心广告资产仍然优质，但在 AI 投资回报被量化以前，需要对利润率、自由现金流和资本配置给予更高的风险折价。

资料¶

• Meta Q2 2026 results
• Meta Q2 2026 earnings presentation
• Meta Q2 2026 earnings call transcript
• Meta Q1 2026 results
• Meta Q4 and full year 2025 results
• Meta Q3 2025 results
• AP: Meta shares drop as Q2 profit declines
• Axios: Meta and Microsoft report ballooning AI expenses
• Investing.com: Q2 earnings call and initial market reaction
• Yahoo Finance: META historical data and market reaction。29 Jul 2026 收盘价为 585.61 美元；盘后数据因抓取时间不同约下跌 7%–10%。

Comments