As AI technology accelerates the core business and the blueprint for personal superintelligence unfolds, Meta's second-quarter revenue exceeded expectations with a 28% increase. However, weak third-quarter guidance and high AI capital expenditure dampened market sentiment.
During the recent Meta earnings call, CEO Mark Zuckerberg and CFO Susan Li sent a clear signal to the market: Meta's AI investments are not only generating real returns but are also shaping a "superintelligence" business landscape covering billions of people.
Regarding selling computing power, Zuckerberg stated that the company has indeed received numerous offers for its computing capacity, with "prices far higher than what we pay." He emphasized, "Simply selling all our computing power for short-term profit is foolish. We believe the margins on selling intelligence will remain significantly higher than directly selling computing power."
He also believes that current open-source models are not as powerful as cutting-edge models. As a "full-stack technology company," Meta must possess the ability to build its own models.
AI Monetization Shows Early Results, Ad Business Leads the Industry
The market's most watched revenue growth and AI monetization capabilities were strongly validated in the second quarter. The family of apps' advertising revenue reached $59.4 billion, a 27% increase year-over-year. "In dollar terms, our ad business reported faster year-over-year revenue growth than any other company's reported ad business," Zuckerberg said on the call. "So, these AI investments are paying off."
The improvement in recommendation systems and ad conversion rates from AI is the core engine of performance growth. Li pointed out that based on predictions and rankings from large language models, the global average price per ad rose 12% year-over-year. Currently, 9 million small businesses on Meta's platform use at least one AI-powered ad creative tool. Its AI-driven Advantage+ end-to-end solution continues to grow, with an annual revenue run rate exceeding $75 billion.
Facing the 'Computing Power Anxiety': Raising Capital Expenditure, Expanding Long-Term Capacity
As AI usage continues to increase, investment in computing infrastructure has become a market focus. Meta announced a raised capital expenditure forecast on the call, expecting full-year capital expenditure to be between $130 billion and $145 billion.
When discussing the shortage of computing power and the phenomenon of external parties seeking high prices to buy it, Zuckerberg demonstrated a long-term business perspective. He revealed that the company has received many offers for its computing power, with prices "far higher than what we pay," but he stressed, "Simply selling all our computing power for short-term profit is foolish."
In Zuckerberg's view, using computing power to build intelligence and compound its value is the true business flywheel. "We believe the margins on selling intelligence will remain significantly higher than directly selling computing power." To meet the massive demand for computing power, Meta is actively advancing infrastructure construction, including the recently announced partnership with BlackRock to develop a 1-gigawatt data center in Texas.
Li said the industry has historically underinvested in capacity for AI adoption, and overall industry capacity will remain tight for the foreseeable future. Therefore, Meta's current plan is "aimed at maximizing capacity in 2026 and 2027."
Zuckerberg's Ultimate Ambition: Putting Superintelligence Directly into People's Hands
Beyond the core advertising business, personal and commercial AI agents form the most significant part of Meta's future performance guidance and imagination. When discussing the long-term vision for AI, Zuckerberg used a striking statement: "We are the only major tech company whose primary goal is to put superintelligence directly into people's hands. We are not centralizing superintelligence; we are focused on widespread distribution."
He boldly predicted the broad prospects of personal agents: "If you look five years out, it's highly unlikely that there won't be billions of people who have a personal agent that knows your goals and works for you around the clock to achieve them."
Currently, Meta's business agents have been released globally on WhatsApp and Messenger, with over 1 million businesses using them weekly. In the future, businesses will only pay Meta when actual conversion results occur, which could create a new auction business model similar to the existing advertising system.
Facing the debate between open-source and closed-source models, Zuckerberg clearly stated that Meta does not blindly follow. He believes that current open-source models are not as powerful as cutting-edge models, and as a "full-stack tech company," Meta must have the ability to build its own models.
"When you look at what discerning customers and companies around the world want, they want to know they are in control of their own destiny... I think open source will be important, but building models will also be a key part of it."
Meta 2026 Second Quarter Earnings Call Transcript
Mark Zuckerberg, Founder, Chairman and CEO
Okay. Hello everyone, thank you for joining today. Our community and business had a strong quarter, with 3.6 billion people using at least one of our apps daily. We reached several milestones. Instagram's daily active users reached 2 billion, Threads' monthly active users exceeded 500 million, making it the fastest-growing conversational app ever. Facebook's daily active users have been over 2 billion for some time. WhatsApp just set a historic message record, with a peak of 30 million messages sent per second during the World Cup final. Additionally, Kunal Shah just joined us as the new head of WhatsApp; he founded one of India's most important payment companies and will be a great addition to the team.
We also released some powerful new models from Meta's superintelligence lab and launched new glasses. Overall, the scale and influence of our community across our apps are incredible, providing a powerful platform for us to deliver new innovations to billions of people. The opportunities before us are immense.
First, we are now at a stage where our AI investments are accelerating every major part of our core business. They are improving the experience of people using our apps, delivering better performance for advertisers, and helping our teams build new experiences and ship products faster.
Second, we are developing new personal agents that will form the basis of our next wave of products and revenue lines in the coming months and years.
Third, we see a large opportunity in selling to businesses, including API, business agents, potentially direct sales of computing power, and other services we build for large customers.
Today, I want to spend some time detailing how our investments are producing results and the long-term opportunities we see.
In accelerating the core business, we are seeing strong and promising results in several areas. On Instagram and Facebook, I am very optimistic about our work integrating large language models into our recommendation system. LLMs add a first-principles understanding of content and why it's compelling, as well as a deeper understanding of what people are interested in and their goals when using our apps.
This means we can show more relevant and engaging content that better reflects people's goals and interests. Our new Muse image and Muse video models will also greatly expand the range of content people can discover on our platform. There are already two huge data streams of content to leverage: first, content from your friends and people you follow; second, content from creators you don't follow. But now there will be a whole new, almost infinite universe of personalized content, which will make our services more useful and engaging for people.
For advertising, we are using LLMs to improve how our systems predict and rank the ads we show. We have expanded the range of context our systems can consider around a person's natural activities and advertising activities to determine ad relevance, resulting in significantly improved relevance and conversion rates on Facebook and Instagram. In dollar terms, our ad business reported faster year-over-year revenue growth than any other company's reported ad business. So, these AI investments are paying off.
We are also seeing strong demand for our new AI-powered creative tools. 9 million small businesses on our platform now use at least one of our AI ad creative tools, and we are rolling out new end-to-end creative solutions to help advertisers turn performance data into their creative decisions. Muse Image will further drive this. This model can analyze images, improve its own work, and generate better ad variations based on advertiser input. So far, we have received great feedback.
We also just launched Meta One, a new subscription service offering more tools and AI features within our apps. As demand grows, we will offer various tiers and pricing options.
I am also excited about how AI is helping our teams accelerate product development. Earlier this year, we released Instagram Instance. We also just launched Forum, a standalone group app, and Seller, a standalone marketplace app. I expect launching new apps will become much easier. So, we plan to develop more ideas and use our recommendation system to scale them to people who will find them interesting, as we did with Threads.
It has been over a year since Meta's superintelligence lab was launched, and our trajectory is solid. Last month, we released Muse Spark 1.1 and Muse Image. Since we rebuilt Meta AI and integrated Muse Spark, the number of people interacting with the assistant daily has increased by 60%, and it continues to grow rapidly week-over-week.
Muse Spark 1.1 is a powerful coding model, very efficient, and excels at computer use, tool use, and multimodal understanding. It is available through our new public API, and we will expand distribution through partner channels and more coding agents in the coming weeks. We are also building features to make it easier for businesses to adopt Muse Spark, an area we expect to continue focusing on.
The reason we are so focused on making Muse Spark perform well in terms of features is that we believe there is a very large opportunity in releasing several agents aligned with our mission and business.
The first is personal agents. Soon, we will have agents that can work for you around the clock, helping you achieve your goals, improve your life, health, relationships, finances, whatever you want. The first area where agents are truly taking off is coding. But engineers are more technical and willing to spend time making these agents work. So, to build a great personal agent, it needs to be a great consumer product, work out of the box, and be simple enough for billions of people to adopt and use. I am very excited about this, and we will have more information to share soon.
As we move towards a future where we all interact with multiple agents, WhatsApp and our other messaging services will become increasingly important. WhatsApp is already the primary interface for people interacting with Meta AI, and as we build a platform for more agents within our messaging apps, we will innovate on how to provide private and secure AI experiences. This quarter, we launched incognito mode on WhatsApp and the Meta AI app, allowing people to have private conversations with the assistant that even Meta cannot see. We also plan to make powerful privacy and security a fundamental component of the agents we are building.
I am also very excited about the progress of our business agents. This quarter, we globally launched Meta business agents on WhatsApp and Messenger, and over 1 million businesses are already using them weekly to talk to customers or complete sales. We are now also launching business agents on Instagram. The interesting thing is, having an agent talk to your customers daily, it learns over time and can bring all those insights back to you. So, we are building more features to summarize all these conversations, digest what happened overnight, and present customer needs.
Soon, it will go further, including suggesting ways to grow your business, providing you with competitive intelligence, and real-time insights on what's working and what's not. Over time, we want to build this as a "one-stop business" service that can help you start and run an entire business using the Meta platform.
Regarding how we will monetize these, we have a mix of subscriptions and volume-based pricing, and I expect we will continue to evolve more of these products to be like our ad system, where businesses only pay when our help delivers results. Over time, this will allow us to run efficient auctions on our computing power, similar to what we do for advertisers today.
As AI usage continues to increase across our products and business, we will continue to invest aggressively in infrastructure to meet demand. Yesterday, as part of our Meta Compute initiative, we announced a new strategic partnership with BlackRock to develop a 1-gigawatt data center in El Paso, Texas.
Overall, we expect a large portion of our computing will be used for training our models, growing our core business, and delivering personal agents and new products. But we also expect to develop a large business serving large customers. We have already built the API. We are launching business agents. We have received numerous offers for computing power, at prices far higher than what we pay. We have more coding and productivity tools on our roadmap. We will share more about this soon.
As we approach personal superintelligence, we will also need hardware that allows you to interact with it seamlessly. Glasses are the ideal form factor because they can be with you all day and help you without taking you out of the moment. Our glasses are still one of the fastest-growing consumer electronics products ever, and we continue to expand the product line. We just released our own collection developed in partnership with EssilorLuxottica, including a model we designed with Kylie Jenner. They are the first glasses that come with Muse Spark out of the box, so they can understand what you see and give more helpful answers. Early sales are strong and exceeding our expectations. We will share more about our glasses product line at our Connect conference on September 23rd this year, so I encourage everyone to tune in.
Before I wrap up, I want to mention that I just published a column article explaining why I am so optimistic that we are building a positive future for everyone. At Meta, we have always been committed to building technology that puts power in people's hands, allowing them to connect with people they care about and shape the world the way they want. This is why we have always focused on making our products affordable and easy to use for everyone, a strategy that has greatly benefited our community and business.
As we enter the next chapter, the same philosophy guides our approach to AI. We are the only major tech company whose primary goal is to put superintelligence directly into people's hands. We are not centralizing superintelligence; we are focused on widespread distribution, empowering everyone to direct it towards what matters to them. This is how society has always progressed. I believe these are the right values to build a positive AI future.
If we help build this, then I think we will also continue to build a very strong business. That's all I wanted to cover in my opening remarks. AI is improving our core business. It is making our apps more relevant, delivering better results for businesses. We are starting to deliver more novel products, and we will have more progress on this soon. We are investing aggressively because the potential is huge, and we know there are many ways to create value here. As always, I appreciate you joining us on this journey. And now over to Susan.
Susan Li, Chief Financial Officer
Thank you, Mark, and good afternoon, everyone. Let's start with our segment results. Unless otherwise noted, all comparisons are year-over-year. Second-quarter Family of Apps total revenue was $60.4 billion, up 28% year-over-year. Second-quarter Family of Apps ad revenue was $59.4 billion, up 27% year-over-year, or 26% on a constant currency basis.
In the second quarter, the total number of ad impressions served across our services increased by 14%. Impression growth was strong across all regions, driven by growth in user engagement and user numbers, as well as optimization of ad load. The global average price per ad increased 12% year-over-year, driven by improved ad performance, improvement in the macro environment compared to the second quarter of last year, and a tailwind from foreign exchange. This was partially offset by strong impression growth, especially from regions and surfaces with lower monetization.
Family of Apps other revenue reached $1 billion for the first time, up 73% year-over-year, driven primarily by WhatsApp paid messaging and subscription revenue. In the Reality Labs segment, second-quarter revenue was $431 million, up 16% year-over-year, driven by strong growth in AI glasses revenue, partially offset by lower sales of Quest headsets.
Now turning to our consolidated results. Second-quarter total revenue was $60.8 billion, up 28% year-over-year, or 27% on a constant currency basis. Second-quarter total expenses were $42 billion, up 55% year-over-year, including $2.4 billion in legal-related expenses and $1.2 billion in severance costs related to the May 2026 layoffs.
The year-over-year increase was primarily driven by higher employee compensation, infrastructure costs, legal-related costs, and third-party AI token costs. Excluding the previously mentioned severance costs, the growth in employee compensation was driven by the addition of technical personnel, particularly AI talent, over the past year. The growth in infrastructure costs was due to increased depreciation, data center operating costs, and higher third-party cloud spending.
At the end of the second quarter, we had over 75,000 employees, a decrease of 3% from the first quarter. This total includes approximately 8,000 employees affected by the May 2026 layoffs. We expect the majority of affected employees will no longer be counted in our headcount by the end of the third quarter of 2026.
Second-quarter GAAP operating income was $18.8 billion, down 8% year-over-year, with an operating margin of 31%. Excluding the second-quarter legal expenses and severance costs, our second-quarter operating income would have increased 9% year-over-year. The tax rate for the quarter was 16%. Net income was $15.8 billion, with earnings per share of $6.18.
Capital expenditure, including principal payments on finance leases, was $31.1 billion, driven by investments in servers, data centers, and network infrastructure. Free cash flow was $784 million. At the end of the quarter, we had $90.3 billion in cash and marketable securities and $83.7 billion in debt.
Now turning to business performance. Two main factors drive our revenue performance: our ability to deliver engaging experiences for our community, and our ability to effectively monetize that engagement over time.
On the first point, we continue to see significant benefits from our content recommendation initiatives. On Instagram, global time spent grew by double digits year-over-year in the second quarter, primarily driven by our improvements to feed and Reels recommendations. On Facebook, global video watch time grew 9% year-over-year, and over 10% in the US and Canada, driven by improvements in ranking.
We are finding that LLMs are increasingly capable of providing ranking and recommendation gains. First, they make our existing systems smarter by understanding the actual content and generating better training data. Second, LLM-powered agents also help engineering development by evaluating content quality, detecting trends, and testing ranking changes.
Earlier this year, we reached a milestone where every public Reels and feed post on Instagram is automatically processed by an LLM and analyzed across multiple dimensions, from topic to tone. We are also working to bring more of Facebook's surfaces under this. These signals can then be passed to downstream applications for ranking, recommendation, and content policy enforcement, which is a key building block for greater personalization.
This quarter, we also began using our Muse model family for content understanding, including signals like video topic classification and summarization, and we are seeing positive early results. Finally, our recommendations are also becoming more personalized, giving people more direct control over what they see while presenting more fresh content.
On Reels, we released our largest single ranking improvement, combining faster inference with a new architecture that leverages deeper user history to improve predictions. This led to a 15 basis point increase in inbound sessions on Instagram, particularly strong in terms of shares and watch time, both strong indicators of improved content-user matching. We are now bringing this to the feed, and early results look similar.
We are also bringing new content to people faster. Our investments in more real-time infrastructure and new video modeling improvements now allow our largest ranking model to identify high-quality new Reels at the time of creation. On the Instagram feed, over half of recommended content is now less than a day old, more than double the amount from a year ago.
We are also giving people more direct control over what they see. Today, Instagram users can access the Youralgo page, which allows users to write natural language prompts to adjust their recommendations. Similarly, on Facebook, we launched Shape Your Feed. Early results show that over 80% of users who use this feature have retained it.
Looking ahead, we are executing on a long-term effort to develop the next generation of recommendation systems. This includes building foundational models designed to support both organic content and ad recommendations, and developing LLM-native recommendation systems. We reached our first research milestone in the first half of this year, continuously pre-training a large-scale model on recommendation data, and observed healthy scaling laws in the process. We are encouraged by this milestone and expect to continue making progress in the second half of the year.
Turning to the second driver of our revenue performance, which is improving monetization efficiency. The first part of this work is optimizing ad levels within organic engagement. Here, we continue to enhance our systems to show ads at the best time and place. In the second quarter, we also expanded ad availability on new surfaces, including completing the global ad rollout on Threads. On WhatsApp, we introduced support for more ad objectives and ad business performance metrics and statuses, and continue to make progress towards global launch.
Turning to the second part of improving monetization efficiency, which is improving performance for businesses using our services. Within our ad system, we are achieving performance improvements as we deploy more sophisticated and predictive models. This quarter, we launched Meta's Generative Recommender, a paradigm shift in how our ad system works. Instead of scoring each possible ad individually, we now use an LLM to jointly reason about ad content and user preferences and predict the best ad for each person. This makes our ad matching smarter and more precise, leading to compound performance improvements for advertisers.
We deployed the first generative model into our ad retrieval system and saw a significant improvement in ad performance. An early pilot using an LLM to better understand user preferences resulted in a 1% improvement in in-app event conversion rates on Instagram. In the second quarter, we also advanced user understanding models to analyze advertising and organic activity, improving both the user experience and advertiser performance.
Combined with our GEM model for ad ranking and sequence learning, these advancements led to an 8.3% increase in ad clicks on Facebook and a 15.7% improvement in conversion rates. We are also leveraging AI to empower businesses to more easily manage their campaigns, develop ad creatives, and interact with customers. Our AI-driven Advantage+ end-to-end solution continues to grow, with an annual revenue run rate exceeding $75 billion. We are working to deepen adoption, as advertisers using multiple tools see compound performance improvements.
I will share an example of how Advantage+ has significantly simplified and improved the efficiency of performance marketing for SMBs. (inaudible) an online clothing brand in India, previously set up Facebook and Instagram ads manually for each campaign. After adopting the Advantage+ sales campaign and layering on Advantage+ audience, placement, and budget optimization, they saw a 13% lift in purchases and a 16% improvement in ad-to-cart conversion rates.
Adoption of our Gen AI ad creative tools continues to expand, with over 9 million small businesses using at least one AI creative tool. The image generation feature now allows advertisers to create more creatives at scale from existing content, including a new feature to create images from video assets, whose adoption rate more than doubled this quarter.
We also launched a new end-to-end creative solution, providing advertisers with AI infrastructure to turn real-time performance signals into their next creative decision while maintaining brand identity and tone. We built the integration from day one with agencies, so teams can diagnose, generate, and scale high-performing creatives without leaving their existing workflows.
Looking ahead, with the launch of Muse Image, we expect to further enhance advertisers' ability to generate high-quality, brand-consistent creatives at scale. With Meta business agents, businesses can better serve customers through our messaging apps, responding to inquiries, recommending products, and handling support around the clock.
Earlier this month, we also launched the Meta business agent platform, which provides businesses with the infrastructure to build, customize, and deploy their business agents at scale on WhatsApp. The platform offers built-in enterprise-level controls, guardrails, and measurement tools for large businesses, allowing them to define rules and deliver personalized experiences within the messaging app customers already use.
Movita, one of the largest car rental companies in Brazil with nearly 400 locations, deployed a business agent on WhatsApp that handles the entire booking process from vehicle selection, pricing to payment in a single conversation. Repeat customers can complete a booking in just three messages. Within a month, Movita reported a 44% year-over-year increase in daily bookings through WhatsApp, with 85% of conversations in that channel fully resolved by the AI agent without human assistance.
We are also building other ways to monetize our ecosystem. Two other revenue sources are subscriptions and monetizing our competitive models through API. Meta One is an evolution of our subscription portfolio, designed to create more value for everyday users, businesses, and creators by giving them access to more features and AI tools to create, connect, and stand out. We are excited to bring this to more users and continue to build enhanced tools for our subscribers.
We recently also launched a high-intelligence model API at a competitive price and are encouraged by initial results. We recently made Muse Spark available to US developers, expanding its distribution and making it easier for developers to adopt the model. We expect to roll out the model API to more distribution channels soon, offer it in more countries, and open it up to businesses.
Our approach to building capacity is strongly influenced by several key factors. First, the broad environment for building infrastructure is dynamic and uncertain in both the short and long term. The industry has historically underinvested in capacity for this wave of AI adoption, making existing capacity, including our own, extremely valuable. In the long run, the supply chain needs to be built out to support the capacity we and others anticipate will be needed for AI experiences.
Second, we have high confidence in our ability to use capacity to scale and build our existing experiences and continue investing in foundational models that will create significant new opportunities. Therefore, our current plans are aimed at maximizing capacity in 2026 and 2027. When we have had incremental capacity in the past, it has proven very valuable in scaling experiences, and we believe this will hold true in this timeframe.
In the long run, the exact growth curve of usage is harder to predict, but we believe our distribution advantage will give us the opportunity to provide valuable AI products to everyone, whether it's our 3.6 billion users or millions of businesses. This should hold true regardless of whether our models are at the frontier, but we believe being at the frontier will unlock new markets and opportunities that may require additional computing power.
Therefore, our long-term capacity strategy is designed to give us the flexibility to continue growing our computing power in 2028 and beyond by laying the data center and network foundation to accommodate future server decisions. The long-term nature of these assets inherently provides flexibility, allowing us to adjust our investments according to the pace of AI adoption.
Additionally, we are making strategic investments in areas like custom chips internally, which will provide long-term strategic flexibility and supply chain leverage. This will help us achieve better returns on these long-term investments. Finally, we believe overall industry capacity will remain tight for the foreseeable future.
As we have said before, we firmly believe that the models, consumer experiences, and enterprise products we are building will be the best and highest-return use of our infrastructure. These enterprise products have the potential to take many forms, as Mark mentioned, tools, our API, or directly monetizing computing power, given the significant market demand. We expect to remain flexible with these opportunities, which will help us have the computing power we need when we need it while maintaining strategic flexibility and giving us multiple avenues to generate returns on invested capital that exceed expectations, thus more effectively funding our building.
When funding these infrastructure investments, the strength of our balance sheet allows us to attract capital from a wide range of markets to supplement the cash flow generated by our business. Our announcement yesterday with BlackRock is an example of the partnerships we can build to complement our approach to building infrastructure capacity.
Now turning to our financial outlook. We expect third-quarter 2026 total revenue to be between $61 billion and $64 billion. Our guidance assumes that, based on current exchange rates, foreign currency will be a headwind of approximately 1% to total revenue growth year-over-year.
Regarding expenses and outlook. We are raising the lower end of our expense outlook to include the $2.4 billion in legal-related expenses recognized in the second quarter. We now expect full-year 2026 total expenses to be between $165 billion and $169 billion. We continue to expect operating income this year to be higher than operating income in 2025.
We expect 2026 capital expenditure, including principal payments on finance leases, to be between $130 billion and $145 billion, narrowed from our previous outlook of $125 billion to $145 billion. Provided there are no changes to our tax environment, we expect our tax rate for the remaining quarters of 2026 to be between 15% and 17%, up from our previous outlook of 13% to 16%.
Finally, we continue to watch active legal and regulatory matters that could have a significant impact on our business and financial results. For example, we continue to see scrutiny related to youth issues in multiple markets, and there are several youth-related trials in the US this year that could ultimately result in significant losses.
In closing, our business momentum continued in the second quarter, with strong execution on core advertising and engagement initiatives. We also advanced our efforts to bring personal superintelligence to everyone, releasing exciting models and expecting to continue this momentum with new products for the remainder of the year.
With that, Krista, let's begin the Q&A session.
Question and Answer Session
Operator
Thank you. We will now begin the question-and-answer session. (Operator instructions) Your first question comes from Brian Nowak of Morgan Stanley. Please go ahead.
Brian Nowak, Analyst
Thank you for taking my questions. I have two, one for Mark and one for Susan. Mark, thank you for the rich information about the new product pipeline of consumer and business agents, API tools, and compute leasing. There are many opportunities here. My question is, as you look at these opportunities and the current state of products and computing power, which ones do you expect to scale first in '26 and '27 to show investors quantifiable, material returns on invested capital? And then Susan, regarding some public comments about '27 capacity and doubling capacity, thank you for the clarification on CapEx. Any early thoughts on '27 CapEx? Even just thoughts on upside sources or downside pressures to help us think about the different ways to finance this multi-year build.
Mark Zuckerberg, Founder, Chairman and CEO
I can take the first question. Regarding the different opportunities and how we think about computing power, by and large, a significant portion of computing power is used for training. As a leading lab training models, I think this is an important investment. But the rest is used for a range of different product and revenue opportunities, which encompasses everything from optimizing and improving our core business, to our upcoming new consumer products, to the API, business agent work, the developer tools work on the roadmap I mentioned, and the opportunity to directly sell computing power, which I mentioned we have received numerous offers for, at prices far higher than what we pay.
The key to thinking about this is that we believe the margins on selling intelligence will remain significantly higher than directly selling computing power. But we also think there is a significant opportunity in selling computing power. So, we are thinking—I think you asked which area might scale the largest, but honestly, I am quite optimistic that all of these areas will achieve meaningful growth. I think we will share more information on several of them soon.
Susan Li, Chief Financial Officer
Brian, regarding your second question, we are not currently providing a specific outlook for 2027 CapEx. Infrastructure planning remains highly dynamic, and even within this year, our outlook includes a range of possible outcomes. I mentioned in my prepared remarks a key focus of our current infrastructure plan, which is to work towards maximizing capacity in 2026 and 2027, and to give us the flexibility to continue growing in '28 and beyond, while also allowing us to make server decisions as we assess actual demand in '28 and beyond. So, we are still studying our capacity needs for the coming years. Broadly speaking, we think near-term capacity is more valuable than long-term capacity, and this remains a very dynamic planning process.
Operator
Your next question comes from Eric Sheridan of Goldman Sachs. Please go ahead.
Eric Sheridan, Analyst
Thank you for taking my questions. If I may, I have two. When you define the enterprise opportunity, how much of it do you think is currently accessible based on your established go-to-market strategy (as an extension of your existing ad and marketing business), and how much requires building a new go-to-market strategy to capture? And then, Susan, if I can squeeze in a second one. When you think about the sources of funding for the business over the next few years, you mentioned the transaction announced yesterday as an example of exploring ways to finance future obligations. We constantly get questions from investors about how you think about the mix of debt and equity and sources of capital. Philosophically, how do you balance your ambition on spending with the need for capital? Thank you very much.
Mark Zuckerberg, Founder, Chairman and CEO
Okay. I can take the first part. For the enterprise, I think it will be a combination of an extension of existing business and building new business. The existing business is essentially selling to marketers and businesses that are basically customer-facing, trying to reach customers and sell directly to them. Obviously, this accounts for the vast majority of the Facebook and Instagram business. We believe there is an opportunity to continue engaging with customers through messaging apps and other services, using business agents, and, like the ad system, effectively getting paid when we achieve results for businesses.
So, we see this as a natural extension of our existing sales and partnerships with millions of advertisers and hundreds of millions of small businesses using our platform. So, I think this should be a fairly natural opportunity for us, and we are focused on delivering it in a way that doesn't try to maximize sales in the short term, but rather maximizes outcomes for users and builds a robust auction mechanism, which we have seen serves the business well in the long run.
There are also other enterprise customers, and I think we will increasingly serve them as well. We are building coding and internal productivity tools, partly because we need to build them for ourselves. We need to ensure we have the right tools tuned for our own needs. Now that we have them, we believe there is a huge opportunity to serve, whether it's small businesses or large enterprises. This is different from what we have historically been good at, and we will share more information soon on how we will build this. But I think the opportunity here is very large.
My view on this is that obviously there is a lot of news about computing, but I think the enterprise opportunity is the sum of all different things. It's not just about selling computing power; it includes API services, productivity services, and business agents for non-marketing parts of the business. These are all part of the overall product. I think there is a very, very large opportunity there. So, we are very focused on this. This will be a new capability the company needs to build, but I think it's a very important capability to ensure we can maximize the opportunity before us.
Susan Li, Chief Financial Officer
Eric, regarding your second question on sources of funding, this is something we have been thinking carefully about as we plan our finances for the future. Obviously, our strong operating cash flow gives us an advantage in funding our infrastructure expansion. But in recent years, we have also been developing our capital structure, including increasing the proportion of debt, to try to lower our cost of capital. We have found it prudent to continue adding cost-effective long-term sources of funding as we invest in initiatives with long-term horizons, especially AI infrastructure projects.
We have also broadened our funding channels, including partnerships like the one we announced with BlackRock yesterday. We will continue to carefully evaluate different sources of funding as we assess projects in the future.
Operator
Your next question comes from Mark Shmulik of Bernstein. Please go ahead.
Mark Shmulik, Analyst
Yes, thank you for taking my questions. Mark, everyone has a story of someone coming back from Silicon Valley, deep coding, building agents with AI, and then they go home and tell their parents they are using AI wrong; it's like an advanced search tool. Consumer behavior is always hard to predict, but reading your column about "AI for everyone," how do you see consumer adoption bridging this AI utility gap? Are we on the verge of some breakthrough, or do we just need to be more patient? Thank you.
Mark Zuckerberg, Founder, Chairman and CEO
Well, I think some things have already broken through. What's interesting about AI compared to other technologies is that roughly every year (and this cycle might accelerate), every year you get new capabilities that create new possible product lines. So, obviously, there is a market for consumer-facing AI assistants, which is what we are doing with Meta AI, and what competitors are doing with their products in that space.
Then, in the past year, I think the coding agent market has grown very fast, and it's the first real agent market. I think there are several reasons why coding took off first, right? You have a technical customer base willing to spend time making it work. Coding is inherently a digital, closed-loop activity. So, in some ways, it's a bit of an easier system to train.
But our bet—or one of the bets we are making here—is that we believe consumer personal agents will eventually become an extremely important and massive market. I think if you look five years out, or whatever timeframe you want, it's highly unlikely that there won't be billions of people who have a personal agent that knows your goals and works for you around the clock to achieve them, regardless of the domain you care about, whether it's helping you improve your health, hobbies, personal finances, making your home life more efficient, or improving and enhancing your relationships, helping your career, helping all these different things.
This is a very, very deep set of use cases. But as you said, I think building for consumers is slightly different from building for developers. If you build for consumers, especially if you are trying to build something for billions of people, not just millions, it needs to work out of the box, right? It needs to be simple. I think the problem with many of the agents, many of the prototype agents we see today, is that they require a lot of debugging; you need to get into the terminal to set them up. You start a use case, and it looks a bit magical, but it might break down over time.
I think companies that can deliver a "out-of-the-box" personal agent will be one of the next major opportunities in the AI space. We are very excited about this, in addition to our work on Meta AI, business agents, and all these other things, and we are simultaneously developing models we believe will continue to advance new capabilities to create new product lines.
So, this is why we are very optimistic. Now, we will release this product sometime in the near future, and that will be very exciting. We haven't done it yet. So, there's only so much I can say about this on an earnings call. But I think this is a very large opportunity, and I think it's almost inevitable that someone will do it. I think it really plays to Meta's strengths as a company—we build consumer products that reach billions of people. Once we get something working, we are very good at scaling it to a large user base; we are good at building the infrastructure to support these intensive applications. So, I feel very good about this. I understand we need to deliver it for our community, and that is what we are very focused on.
Operator
Your next question comes from Doug Anmuth of JPMorgan. Please go ahead.
Douglas Anmuth, Analyst
Great. Thank you for taking my questions. One for Susan, one for Mark. Susan, you talked about the increasing capability of LLMs in providing ranking and recommendation gains. Could you elaborate a bit more on the roadmap here and how far along we are in leveraging better models and more compute? And then, Mark, regarding the number of offers to monetize your computing power externally, you are also buying capacity from multiple third parties. So, I hope you can help us understand some of the differences? Is it just a matter of timing and transition, or is it the difference between training and inference, and using the best chips for each task? Thank you.
Susan Li, Chief Financial Officer
Thank you, Doug. Let me start with your question about the progress on the recommendation roadmap. First, we certainly see room for further improvement in recommendations in the remainder of this year and into 2027. We expect this to help us drive more engagement growth on Facebook and Instagram. I want to highlight a few points.
First, we will continue to make recommendations more personalized and relevant to user interests, in part by advancing our recommendation models and architecture to more precisely capture user interests and respond more quickly to what users care about at a given moment. Our AI investments will play a significant role in achieving this vision, including expanding LLM-based content understanding to more deeply understand the posts and creators users value, more precisely capture user interests, and respond more quickly to what users care about in the moment, as well as using AI to surface high-quality, fresh, and trending content, and reduce the share of low-quality content.
Second, we continue to improve our data infrastructure to allow our models to train on more data and use that data more effectively. We have added more detail to the description of content users have interacted with in the past and enriched the sequence of past user interactions with more granular content. This allows our models to more precisely learn which interactions are more valuable or less valuable to the user. We continue to increase the length of user interaction sequences used in training, as well as the complexity of our model architectures on Facebook and Instagram, to leverage larger datasets.
Then, we have already made significant progress in using LLMs for content understanding, and given their ability to understand content more deeply, we will further integrate them into our recommendation and content policy enforcement stack. And, I would say, overall, we are very optimistic about this set of work.
We are also investing in agent-based methods using LLMs to transform our recommendation system, and the number of releases from our ranking agents has increased in the first half of this year. This has also helped improve the productivity of our engineers, which is another path we are very excited about.
Mark Zuckerberg, Founder, Chairman and CEO
Okay. I can take the second part. I think your question is about how we view the offers we receive to sell computing power, while at the same time we are also buying computing power. I mean, the high-level observation is that there simply isn't enough computing power to meet all demand. That's why we see—basically, we receive a lot of offers for the computing power we have, but we also have many internal uses we consider very valuable.
Now, in running the business, obviously a common trade-off we need to make is how much you monetize it today versus developing future assets for the future. I think this is always a mix, right? You don't want to only do long-term things, and you don't want to not do any short-term things and not validate the near-term market. But I also think simply selling all your computing power for short-term profit is foolish. When you have the opportunity to build intelligence on top of it, which will be a multiple, and compound the value of the computing power on top of it.
So, I think the answer is exactly what we are doing, which is essentially using a lot of our capital to build computing power, being confident in our ability to directly monetize computing power when it makes sense to do so, while also knowing that we have many different use cases to monetize the intelligence on top of the computing power. This includes the enterprise use cases we talked about, the consumer use cases we talked about, and also the core business, which isn't necessarily a new product we didn't talk about, but using it to further increase intelligence, improving ranking, recommendations, and advertising in our core services.
So, I think all of these are real. This creates a dynamic where we are now investing to build these data centers, and they will come online at some point in the future. Obviously, you won't get value from them until they come online. But basically, we see the demand for all of these as very large, and we want to maximize the opportunity to build all these different businesses.
Operator
Your next question comes from Justin Post of Bank of America. Please go ahead.
Justin Post, Analyst
Great. Thank you. Mark, you hired senior leadership for the AI lab about a year ago. Can you talk about how the lab is performing? Do you think Wall Street will see a real uptick in product velocity in terms of models, chips, or something else? And then, what sustainable competitive advantage do you think the lab is building? Thank you.
Mark Zuckerberg, Founder, Chairman and CEO
Yes. So, I mean, I'm fairly satisfied with our trajectory so far. We have released some impressive models on our early scaling ladder. As I said, we are scaling up to larger, more advanced models, and we are excited about that as well. I think there is the intelligence side and the data side. I think in any product category you want to enter, there will be some kind of flywheel effect where you can learn from the behavior of people using the product in that community. So, the better you are at that, the more feedback you get, and then this inevitably makes the product better.
So, I think this is an area where people can be understandably obsessed with intelligence, but the knowledge of data and serving users is very important. If you are talking about personal superintelligence, then obviously having a very clear mental model of everything that happens in a person's life and their goals will be a very important aspect of it. Certain companies have different advantages in different markets and different parts of this field.
But I think part of the reason for working on these different things in advance is that, first, the technology is general. So, when you build intelligence, it can be applied to many different uses. Then you want to invest in building flywheels on those uses, because I think that is how you build a sustainable advantage over time.
But I think we at Meta have clearly shown that when we have a product in a valid form, I would say we are probably the best company in the world at scaling those experiences to billions of people. So, I feel good about the personal agent work. I feel very good about the business agent work, with our existing advertiser base and the small businesses we serve, and our ability to scale and promote them. I think these are also durable advantages, along with the data flywheels we build on top of them.
Then, obviously, on the research side, you want to build a good culture. I mean, it's just basic things, but I feel you just have to make sure you are building the team in a well-managed, consistent, stable, low-drama, compounding way, which is what you aim for. I think if you can do that well, then maybe it's not a clear mathematical explanation, but I think that's how business works. So, yes, that's what we do.
Operator
Your next question comes from Ross Sandler of Barclays. Please go ahead.
Ross Sandler, Analyst
Yes. Hey, Mark. Continuing on the comments about the AI lab. Muse Spark 1.1 is very close to the frontier, but it's on the low end of the intelligence spectrum, I should say the low-cost end. From your previous answer, it sounds like you think it's important to compete on both the low-cost end and the more expensive, high-performance end. Can you talk a bit about that? And then, your lab leadership has also talked about returning to open source, returning to where you were a few years ago. So, how does that fit into the strategy, and all this discussion around monetizing AI products and models? Thank you very much.
Mark Zuckerberg, Founder, Chairman and CEO
Yes. So, let me go a little deeper into the scientific process of this. Basically, at each stage of training a large model, you see new behaviors. So, basically, what you want to do is build up from training smaller models to building larger models. The models we have released so far are based on—at a certain scale on the scaling ladder, we are continuing to scale up to larger models.
So, I think for Muse Spark 1 and Muse Spark 1.1, I think they are very impressive models for the scale of the model, the stage of lab development, and I feel good about them. You also want to have more advanced models, which is why we are scaling up to larger models, which is why we are building a lot of research infrastructure around this.
But at the same time, we also want to have good, more efficient models, which will be many of the services we provide to consumers at scale. But if you serve billions of people, you want to have the ability to use more advanced models for very difficult problems, and you also want to have the ability to use simpler, more efficient models for the vast majority of prompts. So, I think both are very important. Efficiency is really important, but I also think we want to be able to solve the most difficult problems for businesses and customers around the world. So, we care about both.
Regarding open source, we have always believed that open source is an important part of the ecosystem, good for the world, and it creates a feedback loop that benefits us, allowing the community to invest in our infrastructure stack and our work and contribute improvements. But we have always said we will take a mix of open source and closed source. That remains true.
Now, in scaling up the Meta superintelligence lab, it's actually somewhat counterintuitive that doing open-source models requires more work, because if you build a closed system for your own use cases, it can be a bit rougher. Whereas if you release it as open source, it will be used for many things, and you want it to be more comprehensive. I just want to make sure the MSL team can build the smartest models we can without restrictions, and we expect to resume releasing some open-source models sometime in the near future. But as we have always said, we are not dogmatic about this. We think open source is important. We want to contribute to that ecosystem. We plan to combine open-source and closed-source models.
Operator
We have time for one final question, and that question comes from Ken Gawrelski of Wells Fargo. Please go ahead.
Ken Gawrelski, Analyst
Thank you. If I may, I have two questions. First, Mark, I want to touch again on your point about open-weight models. There's a lot of discussion on this topic, and you've been involved in commentary. Why or why not? Would this change Meta's view on developing closed proprietary frontier models? If open-weight models proliferate, is there an opportunity for Meta not to have to develop its own frontier models? That's the first question.
Second question, if I may, for Susan. You mentioned in your prepared remarks that the plan is to maximize capacity for '26 and '27. Is this a demand or supply comment? Meaning, are you implying that Meta plans to use all its built capacity internally through '27? Or will you evaluate expansion for '28 and beyond based on Meta product and service demand? Thank you.
Mark Zuckerberg, Founder, Chairman and CEO
I can take the open-source question. Let me see, is the essence of the question whether we can rely on open-weight models because we have some? I mean, currently, open-source models are not as powerful as frontier models. So, the basic answer is no.
Then, there is also the perennial long-term policy debate and question about the actions of other companies, and whether this is something a company like Meta can rely on. I think that is very tricky. So, I think in both respects, we believe we can do better, and we think relying on it is risky. I don't think that's the right approach.
I think we are a company—if you look at Meta as a whole, many people see the surface layer: we built some social media apps, we have an ad business, but we are actually a full-stack technology company. We build our own data centers, infrastructure, chips, and underlying software. When we started, because of my engineering background, I wrote a lot of system code. A big part of Facebook's success was because it actually worked, right? When other social networks weren't fast and efficient, it worked.
I think we are fully capable of building more personalized, optimized, and efficient things. Certain qualitative experiences, others can't even build, because we have been deep in the entire technology stack. In my view, having the ability to build your own models will obviously be a very important part of the future technology stack, which is why this is important for Meta, and why others care about open source, and why open source overall is important, because other companies, even if they don't have the ability to build these models, don't want to rely on just a few closed labs.
I know open-source models have a very important place in the world. To be clear, this doesn't affect the API opportunity or anything else I talked about, because someone still needs to run the models, do the inference, and run them efficiently. Having the computing power to do so will continue to be a source of commercial advantage. So, I don't think these are necessarily contradictory.
But when I look at what many discerning customers and companies around the world want, they want to know they are in control of their own destiny, that they can trust the models they use, and that their data is secure and not sent to competitors, etc. So, I think open source will be important, but I think for us, building models will also be a key part of it.
By the way, I know all models are evaluated on a common set of benchmarks and then distilled down to, well, this model is within a few points of another model or something, but they do have different skill sets, just like people, right? Just like people have strengths in certain areas, different personalities, and perform well or poorly on different tasks.
If you are trying to build personal superintelligence for people, or business agents for small businesses, that might require different skills than what other labs are tuning for. Being able to do full-stack work on our Instagram recommendations, ad system, is how we have achieved these results over time. I believe building full-stack models will largely be an advantage in how we build personal superintelligence agents, business agents, and all these different use cases for the different customers we want to serve.
So, in addition to the distribution advantage we have and the ability to reach all these people and scale effective products, I think this is a lot of durable advantage: you build something specific to that use case and excel at it. I mean, I know this is a huge investment, a big bet. We see the technology working. We are satisfied with the lab's trajectory. I am excited about the upcoming products, and we believe this will be a big deal.
So, I mean, I know this whole industry is a big bet. My personal bet is that those who invest in this will be rewarded over time and feel very satisfied.
Susan Li, Chief Financial Officer
Ken, I'll briefly answer your second question. When we mention focusing on capacity for '26 and '27, there are actually two factors. One is that we are in a demand-constrained state now and for the foreseeable future, which also includes our core business. If we had computing power, we still have many places with positive ROI to deploy it.
Second, of course, just the uncertainty about long-term building capacity constraints, we talked about some needs in our previous comments to further build the supply chain. So, I think after '27, when we look at '28 and beyond, the world will be very different. Our own internal demand will evolve. A lot of things will be settled by then.
Therefore, when we think about planning for '28 today, we are really focused on flexibility. This just gives us the ability to have land and power, but the real decisions about buying chips and other commodities will be made much later. So, for now, I think we know there are many good use cases requiring capacity in '26 and '27, and that is indeed what we are building towards.
Ken Gawrelski, Analyst
Great.
Mark Zuckerberg, Founder, Chairman and CEO
Thank you all for joining us today, and we look forward to speaking with you again soon.
This article is originally from "Wall Street Sights," edited by Feng Qiuyi of Zhitong Finance.
Comments