Meta's Q2 revenue forecast disappointed, fueling investor concerns about the company's unprecedented AI spending. The social media giant projected Q3 revenue between $61 billion and $64 billion, with the midpoint falling below the average analyst estimate of $63.2 billion, according to Bloomberg-compiled data. Details: Meta's Q3 outlook fell short of expectations, and it raised the lower end of its full-year capital expenditure forecast. Following the earnings release, Meta CEO Mark Zuckerberg and CFO Susan Li held an analyst call to address business questions. Below is a transcript of that call.
Morgan Stanley Analyst Brian Nowak: My first question is for Mark. In your earlier remarks, you detailed new products like consumer and enterprise agents, API tools, and compute leasing—a very large product pipeline. These areas also hold many opportunities. Looking at these opportunities, and considering the current development of products and compute supply, which businesses do you think are most likely to scale first? Which areas might see breakthroughs around 2026 or 2027, allowing you to demonstrate a quantifiable, meaningful return on invested capital (ROIC) to investors? My second question is for Susan. Management has previously discussed capacity plans for 2027 and doubling capacity. I appreciate your detailed explanation of capital expenditure. Can you share any preliminary thoughts on capital expenditure for 2027, even just the overall thinking? For example, what factors could drive capital expenditure higher? What factors could create downside pressure? Also, as you advance this multi-year, large-scale construction plan, is management considering different financing methods? What are the company's capital expenditure arrangements for 2027?
Mark Zuckerberg: I'll take the first question about these different business opportunities and how we allocate compute resources. Overall, a significant portion of compute will go to model training to maintain our competitiveness as a leading AI lab. I believe this is a very important investment. The remaining compute will be deployed across multiple product and revenue opportunities, including continuously optimizing and improving our core business, launching new consumer products, supporting the API business, the enterprise agent business, the developer tools I mentioned, and directly offering compute services. I also noted that many customers are already willing to buy our compute at prices far above our cost. When facing these opportunities, our judgment is that providing AI-powered services offers much higher margins than directly selling compute, and this advantage will persist. Of course, we also believe there is a massive market opportunity in directly providing compute services. Your question about which business is most likely to scale first—I'm actually quite optimistic about the development prospects of all these businesses. I think they all have the potential for significant growth. We expect to have news to share on the progress of many of these soon.
Susan Li: Regarding your second question, we are not providing specific guidance for 2027 capital expenditure at this time. We are still continuously evaluating our infrastructure plans for the coming years, so there is considerable room for adjustment. Even for this year, the capital expenditure range we provided encompasses various scenarios. As I mentioned, our current infrastructure planning is focused on two main goals: first, to meet compute demand as much as possible for 2026 and 2027; second, to retain sufficient flexibility to support growth in 2028 and beyond, allowing us to decide on server deployment once we have a clearer picture of actual future demand. The company is still evaluating the exact amount of compute we will need in the coming years. Overall, we believe that compute capacity we can bring online sooner is more important to us than capacity a few years out. Therefore, our overall infrastructure planning is continuously being adjusted and optimized.
Goldman Sachs Analyst Eric Sheridan: My first question is for Mark. You've discussed the enterprise opportunity. Given your existing go-to-market infrastructure, including channels built from the ads and marketing business, how much of that opportunity can you already address? How much still requires building new go-to-market models and sales structures to capture? My second question is for Susan, a follow-up on capital. Looking ahead at the company's funding sources, you mentioned the announced deal (Meta partnering with BlackRock to build a $14 billion AI data center in Texas) as a way to front-fund future capital needs. Investors also ask us about Meta's future capital structure—the mix of debt and equity, and overall funding allocation. Over the longer term, I understand you want to maintain aggressive investment while meeting large capital needs. What is the company's overall funding philosophy and capital planning?
Mark Zuckerberg: I think the enterprise business will come from two directions. One is an extension of our existing business. We currently serve advertisers and businesses, helping consumer-facing merchants reach users and sell directly to them. This is the foundation of most of Facebook and Instagram's business. We believe we can extend this model with enterprise agents, allowing businesses to continuously interact with customers through messaging apps and other product scenarios. The logic is similar to the ad system: as long as we help businesses achieve results, we can generate revenue. We see this as an extension of our existing sales relationships and commercial partnerships. We already have relationships with millions of advertisers and hundreds of millions of small businesses using our platform, making this a very natural direction. In advancing this, our focus isn't maximizing short-term sales volume, but helping business users achieve the best results and building a mature, effective bidding system. Long-term, our ad business has proven this model works well for driving growth. Beyond this, we will also serve more types of enterprise clients. For example, we are developing code tools and internal productivity tools, partly because we need them ourselves and need to ensure they meet our internal requirements. When these capabilities mature, we believe they represent a significant service opportunity for both small and large enterprises. This differs from our past business models and requires building new capabilities. We'll share more about building this business system soon. Overall, I see a massive opportunity in the enterprise market. There's been much discussion about the compute business, but I believe the enterprise opportunity comes from a combination of directions, not just selling compute. Specifically, it includes compute services, API services, productivity tools, and enterprise agent services for non-marketing scenarios. These together form our overall enterprise solution. I believe there is a very large market here, and we are investing heavily. For the company, this is a new capability to build. But I think it's a crucial capability we must develop to better capture the huge future market opportunity.
Susan Li: Regarding your second question on funding sources, we are carefully evaluating the relevant arrangements in our future financial planning. First, our strong operating cash flow provides a solid foundation for funding our infrastructure buildout. In recent years, we have been optimizing our capital structure, gradually increasing the proportion of debt financing to lower our overall cost of capital. Overall, for investments in projects with long-term cycles, especially AI infrastructure, we believe consistently introducing lower-cost, longer-duration funding sources is a more prudent approach. Additionally, we are continuously expanding our financing methods, including exploring models like the partnership with BlackRock announced yesterday. In the future, when evaluating new investment projects, we will continue to carefully consider different combinations of funding sources to select the most appropriate financing plan.
Bernstein Research Analyst Mark Shmulik: Mark, many people have a similar experience: someone comes back from Silicon Valley, spends a lot of time coding and developing AI agents, and then tells their parents they're using AI wrong—many people just use it as a "better search engine." Consumer behavior is always hard to predict. Considering your recent article "The AI Future is for Everyone," I'd like to ask how you see consumer usage habits evolving. Will they gradually discover more value in AI, solving the problem of AI being powerful but user adoption being limited? Are we at a stage where an AI application breakthrough is imminent, or does this still require more time and patience?
Mark Zuckerberg: Actually, breakthroughs are already happening in some areas. AI technology has an interesting characteristic compared to other technologies: its development cycle may accelerate. Roughly every year, we see new capabilities emerge, which in turn create new product directions. Currently, consumer-facing AI assistants are one clear direction. Meta AI is exploring this, and competitors have similar products. In the past year, I believe the AI coding agent market has developed very rapidly, becoming the first agent application market to truly scale. AI coding agents led the way for several reasons. First, they target technical users who are willing to invest time to try and refine products. Second, coding is a highly digital task with clear feedback, making it relatively easier to train AI for. A key judgment for us is that consumer personal agents will eventually become a very important, massive market. Looking five years out, I think it's almost certain that billions of people will have their own personal agents. These agents will understand your goals and act on your behalf 24/7, whether in areas like health management, hobbies, personal finance, improving home life efficiency, enhancing relationships, or helping your career development. The use cases are very rich and deep. As you mentioned, developing products for consumers is different from developers. For a consumer product, especially one targeting billions, it must be simple, reliable, and "work out of the box." Many agents we see, or early forms of agents, still require significant setup and tuning. Users might need to enter a terminal environment to configure and run a task. Sometimes the experience seems magical, but performance can degrade with deeper use. I believe the companies that can truly build personal agents that "work seamlessly without complex operations" will usher in the next very important wave of opportunity in AI. Beyond Meta AI and our enterprise agent business, we will continue developing next-generation models that enable new capabilities. Because these directions coexist, we are very confident about the future. We will launch related products soon, but I can't officially announce them yet. I can share limited information on an earnings call. But I believe this is a huge opportunity, and someone will achieve it. This direction also aligns well with Meta's strengths. Meta has long excelled at building consumer products serving billions of users. Once we validate a product works, we are very good at rapidly scaling it to a massive user base. We also have the capability to build the infrastructure to support these high-intensity applications. So, I am very confident. Ultimately, we need to deliver the product to users, which is one of our key focuses now.
JPMorgan Analyst Doug Anmuth: I have two questions. First, Susan, you mentioned that large language models (LLMs) are increasingly powerful for improving ranking and recommendation. Can you elaborate on your technical roadmap and development plans in this area? What stage are you at in using more advanced models and more compute to improve ranking and recommendation? My second question is for Mark, regarding the company's work offering compute services to external customers. You mentioned many customers are already requesting compute. But the company is also purchasing compute from multiple third-party suppliers. We'd like to understand the difference. Is this a temporary situation due to timing and phasing needs? Or is it because model training and inference tasks have different requirements, needing chips better suited for each scenario?
Susan Li: I'll take the first question on the future plan for recommendation systems. First, we believe there is still room for improvement in our recommendation systems in the second half of this year and into 2027. We expect these improvements to help Facebook and Instagram continue to enhance user engagement. I want to highlight two areas. First, we will continue to make recommendations more personalized, more precisely matching user interests. Specifically, we will continuously upgrade our recommendation models and model architectures, enabling the system to better understand user interests and respond more quickly to what users are focused on at a given moment. AI technology investment will play a key role in achieving this. One direction is further expanding content understanding capabilities based on LLMs. With LLMs, we can better understand the posts and creators users like, more precisely identify user interests, and more promptly recommend content aligned with users' current focus. Simultaneously, we will use AI tools to help discover high-quality, fresh, and trending content, reducing the proportion of low-quality content. Second, we will continue to improve our data infrastructure, allowing models to train on more data and use it more effectively. Specifically, we will describe content users have interacted with in the past more granularly and use richer content information to refine user historical interaction records. This will help the model better judge which interactions are more valuable to the user. We will also expand the scope of user historical interactions referenced during model training and increase the complexity of Facebook and Instagram recommendation models to fully leverage the value of larger datasets. We have made significant progress in using LLMs for content understanding. In the future, we will further apply LLMs to our recommendation systems and content governance systems, giving them deeper content understanding capabilities. Overall, we are very optimistic about the prospects of this work. Additionally, we are exploring the use of LLM-based agent technology to transform our recommendation systems. In the first half of this year, we launched more projects using AI agents to optimize recommendation ranking. This not only helps improve recommendation system capabilities but also increases engineer productivity. This is another development direction we are very excited about.
Mark Zuckerberg: I'll answer the second question. The essence of your question is this: we are receiving requests from external customers to buy compute, while also purchasing compute ourselves. At a high level, current market demand far exceeds available compute supply. That's why we receive many requests to buy compute. Simultaneously, we have many internal use cases where we believe this compute investment will create significant future value. Operationally, we always consider: should a portion of resources be monetized now, or invested in longer-term capability building? I believe this is essentially a portfolio choice. A company cannot only make long-term investments and completely ignore existing market opportunities; but it would also be unwise to sell all compute for short-term gains. If we can build our own AI capabilities on top of this compute, those capabilities will generate higher value and amplify the long-term return on compute investment. So, our current approach is: invest significant capital in building compute infrastructure, while believing we have the capability to commercialize directly by offering compute services externally at the right time. But we also know there are many monetization opportunities in building AI capabilities on this compute, including the enterprise business we mentioned, consumer-facing applications, and our core business. When I say core business, I don't necessarily mean undisclosed new products, but using AI to further enhance existing services, like improving content understanding in Facebook and Instagram, optimizing ranking, content recommendations, and the ad system. I believe all these factors create the current situation: we are building data centers now, but these facilities have a construction period and only start creating value after coming online. What we see is massive future demand for compute services, enterprise applications, consumer products, and AI upgrades within the core business. Therefore, we want to seize the opportunity and drive these business directions as much as possible.
BofA Merrill Lynch Analyst Justin Post: Mark, about a year ago, you formed the senior management team for the company's AI lab. Can you share your thoughts on the team's development? Looking ahead, will investors see a significant acceleration in the pace of product releases for models, chips, and other AI-related areas? Long-term, what durable competitive advantages do you see the company's AI lab building?
Mark Zuckerberg: I'm very satisfied with the current direction and progress. We have released several models that have shown very strong capabilities in the early scaling phase. As I mentioned, we are developing next-generation models that are larger and more capable, and we are very excited about that. I believe this includes improvements in model intelligence and data capabilities. No matter which product area we enter, I believe it creates a virtuous cycle: gaining feedback by understanding user behavior and how user groups use the product, which in turn helps us improve the product. The better we utilize this feedback mechanism, the better the product becomes. In AI, many people focus on intelligence itself. However, data and knowledge are also very important to truly provide valuable services to users. If we are talking about personal super intelligence, it needs to deeply understand a person's life, goals, and environment. This comprehensive and accurate understanding of the user will be a very critical part. Different companies will have different advantages in different markets and business segments. I believe one reason we can explore multiple directions simultaneously is that AI technology is versatile. As we improve our AI capabilities, they can be applied across many different scenarios. At the same time, we need to accumulate data and feedback in these scenarios to form a continuous improvement loop, as this is a key source of long-term competitive advantage. Meta has proven that once we find an effective product form and use case, we are very good at scaling it to billions of users. We may even be one of the best companies in the world at scaling mature products to a massive global user base. I am very confident about the personal agent direction. I am equally bullish on the enterprise agent direction. Especially in the enterprise agent space, we already have a service foundation serving millions of advertisers and small businesses, giving us an advantage in further promoting and expanding related services. These capabilities, along with the data and feedback loops we are building, will become our long-term, durable competitive advantages. Of course, in research, we also need to build a good R&D culture. This is the most fundamental but most important thing: ensuring the team is managed rationally and efficiently, can develop stably long-term, minimizes internal friction, and allows capabilities to accumulate over time. I believe if we can do these things well, although it's hard to explain with a simple mathematical formula, this is the pattern for long-term corporate success. This is what we are working on.
Barclays Analyst Ross Sandler: I have a follow-up on the AI lab. The Muse Spark 1.1 model seems very close to the optimal balance of performance and cost, operating at the lower end of the intelligence spectrum, or more accurately, the lower cost end. From your answer, you seem to believe the company needs to compete both in lower-cost directions and in higher-cost, higher-performance model directions. Can you elaborate on this? Additionally, the head of the company's AI team mentioned the company might return to an open-source approach, similar to its direction a few years ago. In the current discussions around AI product and model commercialization, how does management see the open-source strategy fitting into the overall plan?
Mark Zuckerberg: Let me start from the model development process itself. Essentially, at every stage of training large models, we observe new capabilities emerging. The typical path is starting with smaller models and gradually scaling to larger ones. The models we have released are at a certain scale, and we are continuing along the path of model scaling. For Muse Spark 1 and Muse Spark 1.1, I believe they are excellent models given their current scale, stage of development, and the current construction phase of the AI research team. I am very satisfied with their performance. Of course, we also want to build more capable models. That's why we are continuously scaling up models and building the corresponding research infrastructure. At the same time, we want high-quality, more efficient models, as they will play a crucial role in future large-scale consumer services. To serve billions of users, we need both more advanced models for very complex, difficult problems and simpler, more efficient models for the vast majority of daily requests. I believe both directions are very important. Model efficiency is critical, but we also want to solve the most complex problems facing global enterprises and users. So, we will focus on both. Regarding open-source, we have always believed it is a very important part of the ecosystem and has positive implications for the entire industry. Open-source also creates its own virtuous cycle for us: the community becomes more deeply involved in our infrastructure and related work, helping us continuously improve and refine. Our consistent view is to adopt a combination of open-source and closed-source approaches, and this will not change. In building Meta Superintelligence Labs (MSL), one point that might be counterintuitive is that developing open-source models actually requires more work from us. A model used only internally for the company's own applications can be optimized for specific needs. But a model released publicly and used widely across different scenarios needs to be more comprehensive and balanced. I want to ensure the MSL team can fully leverage its capabilities to build the most advanced models we can achieve, without being constrained by other factors. We expect to release a series of open-source models again at some point in the future. As we have always emphasized, we are not fixed on one model. We believe open-source is very important and want to continue contributing to the open-source ecosystem. In the future, we plan to pursue both open-source and closed-source models.
Wells Fargo Analyst Ken Gawrelski: My first question is a follow-up on Mark's comments about open-weight models (models with open parameters). There is much discussion on this topic, and management has shared views. If open-weight models become more prevalent in the future, could this change Meta's strategy of developing proprietary, closed-source frontier models? In other words, would the widespread development of open-weight models negate the need for Meta to develop its own frontier models? My second question is for Susan to clarify. In your earlier remarks, you mentioned the company plans to maximize compute capacity for 2026 and 2027. Is this more of a demand-side judgment or a supply-side plan? In other words, do you mean Meta plans to use all the compute capacity it brings online in 2027 for internal demand? Or will the company decide on the scale of further infrastructure buildout based on actual demand for Meta products and services in 2028 and beyond?
Mark Zuckerberg: I'll answer the question about open-source. The core question is: because open-weight models now exist, can we rely on them directly in the future without developing our own? Currently, the answer is no, because open-weight models are not yet as capable as the most advanced frontier models. Furthermore, there is a long-standing policy debate and uncertainty about the behavior of other companies—specifically, whether a company can truly rely on model capabilities provided by another institution long-term. I think this is quite complex. From both perspectives, I have always believed that Meta is capable of building better models, and we also believe there is risk in over-relying on external models. I personally don't think that's the right path. If you re-examine Meta, what many people see might just be the surface: we built some social apps and have an ad business. But Meta is truly a full-stack technology company: we build our own data centers, develop our own infrastructure, design our own chips, and construct our own underlying software systems. In the company's early days, I was an engineer myself and wrote a lot of the underlying system code. Facebook's success was largely due to it being stable and efficient enough—at times when other social networks couldn't be fast and smooth, we were. I believe we have the ability to build more personalized, optimized, and efficient products. Some experiences are difficult for other companies to replicate because our technical capabilities cover the entire technology stack. For me, it's clear that future self-development and control of our own model capabilities will be a very important part of the entire technology system. This is why it's so important for Meta. It's also why other companies value open-source models, and why the open-source ecosystem itself is significant. Many companies, even if they can't train these models themselves, don't want to be limited to relying on a few closed AI labs in the future. Therefore, I believe there must be space for open models. To be clear, this doesn't diminish the API service or other business opportunities we mentioned. Because whether models are open or closed, someone needs to run them, perform inference, and deploy them efficiently. Having the compute and infrastructure capabilities to run models will remain an important business advantage. I don't see these as inherently conflicting directions. But if you consider the growing number of sophisticated enterprises and customers globally, they will want to ensure they can control their future direction, trust the models they use, secure their data, and not hand over important data to potential competitors. From this perspective, I believe open-source models will play a significant role; simultaneously, Meta's self-developed model capabilities will be a key component. Also, while all models are typically compared on a unified benchmark, which might show a few percentage points difference, the actual capability sets of different models are different. It's a bit like people: everyone has their strengths and characteristics, performing better on different tasks. If you are building personal super intelligence for individual users, or business agents for small enterprises, these systems may need capabilities different from models from other AI labs. Just as we could optimize the entire flow of the Instagram recommendation system and ad system in the past, this full-stack capability has allowed us to achieve good long-term results. I believe that building this full-stack model capability in the future will also be a significant advantage for us in constructing personal super agents, enterprise agents, and other application scenarios for different customers. Combined with our user reach and our ability to scale mature products to a massive user base, I believe these will form a key source of our long-term competitive advantage. I believe Meta's true strength lies in its ability to build highly tailored systems for specific application scenarios and make them excel in those scenarios. This will be a very large investment and a very important strategic choice. But we have already seen this technology working. We are satisfied with the trajectory of the AI lab team and are excited about the upcoming products. We believe this will become a very important development direction. Similarly, I understand this is a major investment and bet for the entire industry. But from my personal judgment, I believe the companies willing to invest in this area will be rewarded and satisfied with their choices in the long run.
Susan Li: I'll briefly answer your second question. When I mentioned "focusing on maximizing compute capacity for 2026 and 2027," I had two considerations. First, we are currently, and will be for the foreseeable future, in a state of compute demand exceeding supply. This is true not only for our AI business but also for our core business—if we had more compute, there are still many worthwhile scenarios where increased compute investment could create further value. Second, there is the uncertainty surrounding future buildout capabilities. As I mentioned, we need to further mature the entire supply chain to support even larger-scale infrastructure in the future. When we look out to 2028 and beyond, we believe the industry environment will change significantly, and our internal needs will also evolve. By then, we will have more information to make clearer judgments. Therefore, when planning for 2028 at this stage, our focus is on maintaining flexibility. We want to secure foundational resources like land and power in advance, but for purchasing chips and other major capital commitments, we want to defer decisions as much as possible to a later point, based on actual conditions at that time. Currently, we are very clear that there is still substantial, well-defined demand for compute in 2026 and 2027, which is the primary focus of our current infrastructure buildout.
Comments