Meta conference call: Capital expenditures mainly directed towards AI infrastructure, Meta AI will launch a paid premium version and advertising features

Meta significantly raises capital expenditure guidance, and Mark Zuckerberg promises that the AI strategy will yield high returns.

After the U.S. stock market closed on Wednesday, Meta announced its first-quarter financial report, showing revenue of $42.2 billion, a year-on-year increase of 16%, exceeding Wall Street's expectation of $41.3 billion. Among them, advertising revenue grew by 10% year-on-year, more than double the expectation, becoming a key factor driving performance.

Moreover, Meta further raised its annual capital expenditure expectations, a bold strategy that excites and worries investors. In the earnings call, Zuckerberg stated that the company will continue to bet heavily on AI, not only increasing capital expenditure expectations but also planning to add paid subscription and advertising features to Meta AI.

Here are the highlights from the earnings call:

1. The company raised its annual capital expenditure expectation from $60-65 billion to $64-72 billion, primarily directed towards AI development. The reason for increasing investment in capital expenditure is to build the capability for world-class infrastructure, to truly establish capacity faster in 2025 and 2026, providing an advantage for the company to develop leading AI technologies and services in the coming years.
2. Both the supply and demand for AI are very unstable. There is no fixed answer to when there will be enough supply to meet all demands, which is part of the reason we are accelerating the increase of data center space this year; it is still too early to discuss plans beyond 2025.
3. An important trend is the personalization of cross-agent experiences. Once AI begins to understand you and what you care about in context, and can accumulate memories from your past conversations with it, I believe this will start to become different.
4. Meta AI will launch a paid premium version and advertising features, exploring new revenue sources. The Meta AI application as a standalone app will be particularly important in the U.S., helping to establish our leadership in being the primary personal AI used by people.
5. There is clearly uncertainty in the earnings guidance for the second quarter, considering how the macro environment will evolve over time and how this may affect different areas of our business, which will impact our revenue outlook for the second quarter.
6. It is expected that by mid to late next year, AI coding agents will complete a significant portion of AI research and development work.
7. Increased regulatory pressure in Europe, violations of the Digital Markets Act could result in fines of up to 20% of annual revenue.

Here is the full transcript of the call:

Conference Operator (Krista):

Good afternoon, I am Krista, today's conference operator. Welcome to the Meta first-quarter earnings call. All lines have been muted to avoid background noise. After the main speaker finishes, we will enter the Q&A session (operating instructions follow). This conference will be recorded. Meta's Director of Investor Relations, Ken Dorell, please begin

Ken Dorell, Director of Investor Relations:

Thank you all for attending. Joining me today are CEO Mark Zuckerberg and CFO Susan Li. Our remarks will include forward-looking statements, and actual results may differ due to risk factors. The financial data includes GAAP and non-GAAP metrics, with specific comparisons available in the earnings press release. (Technical difficulties) Now, I will turn it over to Mark.

Mark Zuckerberg, Founder and CEO:

Alright. Thank you, Ken, and thank you all for being here today. We had a strong start this year. Our community continues to grow, with over 3.4 billion people using at least one of our apps daily. I believe we are well-positioned to navigate macroeconomic uncertainties. The main theme currently is how AI is changing everything we do. As we continue to increase our investments, I want to outline our focus areas through five key opportunities.

Advertising Optimization: AI has enabled precise targeting, with new tools in testing increasing click-through rates by 5%, and 30% of advertisers using AI creative tools.

Content Interaction Upgrade: Improvements in recommendation systems have led to user time increases of 7%, 6%, and 35% for Facebook/Instagram/Threads respectively. Threads has surpassed 300 million monthly active users.

Business Communication: WhatsApp has over 3 billion monthly active users and is lowering usage barriers in developed countries through AI customer service agents.

Meta AI: Nearly 1 billion monthly active users, with a newly launched standalone app supporting personalized voice conversations, planning to expand user base before commercialization.

AI Devices: Sales of smart glasses have tripled annually, and Quest 3S is driving VR adoption.

On the technology foundation side, the Llama 4 model leads in multimodal capabilities and energy efficiency, with a larger "Behemoth" model set to launch soon. To support AI development, we are accelerating data center investments, raising capital expenditures for 2025 to $64-72 billion.

All of this is built on our long-term investments in developing general intelligence and leading AI models and infrastructure. Even with significant investments, we do not need (technical difficulties)... If we can realize these visions, I believe these AI investments will yield unimaginable returns.

Susan Li, CFO:

Financial Performance:

• Q1 revenue of $42.3 billion (up 16% year-over-year, up 19% adjusted for currency)
• Advertising revenue of $41.4 billion (up 16%), with e-commerce being the largest contributor to growth
• Reality Labs revenue of $412 million (down 6%), primarily due to a decline in Quest sales
• Operating profit margin of 41%, free cash flow of $10.3 billion

Business Outlook:

• Q2 revenue guidance of $42.5-45.5 billion (with a currency tailwind of about 1%)
• Full-year expense guidance lowered to $113-118 billion
• EU regulatory risks: Adjustments to subscription models could impact 16% of revenue in the European region

Brian Nowak (Morgan Stanley):

That's great. Thank you for taking my questions. I have two questions. The first is about Llama. Mark, the landscape of LLMs is still evolving and highly competitive. Can you share with us some key areas of progress that you are most focused on and excited about as we look towards Behemoth and future versions of Llama?

The second question is about Meta AI, which has nearly 1 billion users globally. Could you talk about your views on the usage of this among U.S. users, as well as the types of repeat user behaviors you are seeing in the early applications of Meta AI? Thank you.

Zuckerberg:

Sure. I can talk about LLMs. Regarding the usage of Meta AI, I'm not sure if we have more data to share at this point. Yes, I'll ask Susan to respond and see if we are ready. Regarding LLMs, yes, there has been significant progress in many different areas, and the reason we want to build it is that, first, we believe it is crucial for our business, and we need to take control of our own destiny rather than relying on other companies to do such critical things. Secondly, we want to ensure that we can optimize development based on our infrastructure and the application cases we want.

To this end, Llama 4, which is a model with 17 billion parameters per expert, is specifically designed for our infrastructure and aims to provide a low-latency experience for voice optimization. One of the key things when you are having a voice conversation with AI is that it needs to have low latency. This way, when you stop speaking, there isn't a long pause before the AI starts to respond. So, from the shape of the model to the research we are conducting, to the technologies being integrated, everything aligns with this.

Similarly, another thing we are focused on is the length of the context window. In some of our models, we are indeed leading the industry in terms of context window length, and we believe this is important partly because we are very focused on providing personalized experiences. You can incorporate personalized content into LLMs in different ways, but one way is to include some content in the context window. Having a long context window that can incorporate a lot of background information shared by users in our applications is one way to achieve this. This gives you insight into the products we are trying to build and some specific technical architecture decisions and research priorities we are making to provide the specific experiences we want. I could go on a lot more. The reason is that I think delivering large models like Behemoth is also important, not just because we will ultimately deploy them in production, but because the technologies distilled from larger models, right?

The Llama 4 model we have released so far, as well as the models we use internally and some models we will build in the future, are essentially distilled from the Behemoth model to achieve 90% to 95% of the intelligence of large models in a lower latency and more efficient manner So all of this is very important. Clearly, we cannot extract this from other closed models. So this gives you an understanding of our development thinking in this area. Of course, the models and infrastructure we are building support all the opportunities I mentioned earlier.

Susan Li:

Brian, I’m happy to answer your second question about Meta AI. The primary application case for Meta AI from a query perspective is indeed information gathering. People are using it to search, understand, and analyze information, followed by social interaction, from casual chatting to deeper discussions or debates. We also see people using it for writing assistance, interacting with visual content, and seeking help. We observe that people interact with Meta AI through several different entry points. WhatsApp continues to see the strongest usage of Meta AI within our family of apps. Most interactions on WhatsApp occur in one-on-one Threads,

followed by Facebook, which is the second-largest driver of Meta AI interactions. We have seen strong engagement from deep dives in the Feed, allowing people to ask Meta AI questions about the content recommended to them. We are clearly excited about the launch of the standalone Meta AI application.

Eric Sheridan (Goldman Sachs):

I’ll follow up on Brian’s question from a different angle and thank you for providing details about the Meta AI use cases you see today. How do you think these use cases will evolve with the development of the standalone application? Can you walk us through the process of deciding to launch a standalone application? What changes might this bring in terms of utility, frequency, or scale compared to what you see today within the family of apps? How do you view positioning Meta AI as a standalone application to compete against other independent consumer-grade AI applications today? Thank you.

Zuckerberg:

Yes, I can talk about that. We will focus on integrating it into our family of apps in more ways and building a standalone experience. I think some people want faster access to it or want a more comprehensive feature set than what you get in apps like WhatsApp, so a standalone application will be valuable.

I also think the standalone application will be particularly important in the U.S. because, as Susan mentioned, WhatsApp is the largest platform for people using Meta AI. If you want to text AI, it makes sense to have it tightly integrated and have a good experience in the messaging app you use. However, while there are over 100 million people using WhatsApp in the U.S., we are clearly not the leading messaging app in the U.S., which is iMessage. We hope to become the leader over time, but our situation in the U.S. is different from what we see with WhatsApp in the rest of the world

So I believe that the Meta AI application as an independent application will be particularly important in the United States, helping to establish our leadership in being the primary personal AI used by people. But we will continue to comprehensively advance experiences in all these different areas.

Justin Post (Bank of America):

Regarding the guidance for the second quarter, there have been reports of potential supply issues related to e-commerce. How are you considering this in your guidance, and what is your outlook for the second half of the year? Then, from a broader perspective, your capital expenditures are now approaching some very large-scale companies with substantial customer bases. Please help us conceptually understand the ecosystem you are building with your capital expenditures. I know you provided a lot of help in your introduction, but perhaps the ROI does not require direct corporate expenditures to drive revenue. How are you considering this issue? Thank you.

Susan Li:

Thank you, Justin. Regarding the guidance for the second quarter, there is clearly uncertainty about how the macro environment will evolve over time and how this may affect different areas of our business. Our revenue outlook for the second quarter aims to take these factors into account, partly because the range of $3 billion reflects the possibility of broader outcomes.

Specifically, we have seen a decrease in spending from e-commerce exporters in Asia in the United States, which we believe occurred before the removal of the minimum exemption on May 2. This portion of spending has been shifted to other markets, but the total spending of these advertisers is below the levels seen before April. However, our second-quarter outlook reflects the trends we are currently seeing in April, which are generally healthy. So it is still too early to know how things will develop throughout the quarter, and of course, it is even harder to know for the remainder of the year.

Your second question is about why we are investing more in capital expenditures. We truly believe that our ability to build world-class infrastructure provides us with a meaningful advantage in developing leading AI technologies and services in the coming years. We also have many opportunities to improve our core business by investing more computing power in our advertising and recommendation efforts. Therefore, even if we invest in capacity in 2025, it will be difficult to meet the demand for computing resources from various teams across the company.

Thus, we will continue to make meaningful investments in our infrastructure footprint here, but we are also really committed to building this capacity in a way that allows us to deploy it flexibly in the coming years. This way, we can respond to market and technological developments.

Douglas Anmuth (JP Morgan):

Thank you for taking my question. I just want to follow up on capital expenditures and infrastructure spending. Regarding the cap on capital expenditures, can you help us understand how much of it is related to additional data center investments, how much is related to increased hardware costs, and what are the actual factors driving these higher hardware costs?

Then, additionally, some articles indicate that you have been seeking partnerships to share some of the costs of building AI infrastructure. Can you help us understand your thought process and some of the pros and cons of going solo versus collaborating? Thank you.

Susan Li:

Thank you, Doug. The outlook for our capital expenditures reflects these two updates, with an increase in data center spending this year due to some adjustments we made to our build strategy, which will enable us to build capacity much faster in 2025 and 2026. We haven't broken down the exact drivers.

This year, we expect to pay higher costs for infrastructure hardware, which actually comes from suppliers around the world, and there is a lot of uncertainty in this regard given the ongoing trade discussions. So, this is reflected both in the broader range we provided, and we are also working to address our mitigation measures in the supply chain. Our outlook does attempt to reflect our best understanding of the potential impacts this year, which involves all of these uncertainties.

Regarding the second part of your question, we are pleased to have partners investing alongside us and bringing Llama to market, such as AWS and Azure, who help us host Llama. We have been looking for opportunities to continue deepening or expanding these partnerships, but we are funding the infrastructure used to train Llama, and we expect that this situation will not change for now.

Mark Shmulik (Bernstein):

Yes, thank you for taking the question. Mark, in your conversation with Satya last night, I believe you both discussed the proportion of code written internally for AI. Going back to your earlier comments about how we might see AI taking on the role of mid-level engineers this year. As the world evolves rapidly, can you share where you see significant progress in this area? Do you think we are making progress toward this milestone faster or slower than you expected?

And then, Susan, given the slight decrease in spending guidance, how should we view the expected overall pace of spending as it relates to core business performance and the reality we are in? Thank you.

Zuckerberg:

I can talk about the work of coding agents. I think our predictions about the timing of this have not changed materially. So, I would say it is basically still on track, and around sometime this year, mid-level engineers will start to become viable and will expand next year. So I expect that by mid to late next year, AI coding agents will handle a significant portion of AI R&D work, so we are focused on that.

Internally, we are also very focused on building AI agents or systems that can help run different experiments to enhance our recommendations in other AI products (such as those that recommend across our Feed). So I think if it works, it should accelerate our progress in these areas. This is the fundamental bet we are making

Susan Li:

Regarding your second question about our outlook for reduced expenses, in fact, we are already four months into this year, and the reduced outlook reflects a more refined forecast, including updates on employee compensation and some other non-personnel-related operating expenses for this year. This is partially offset by higher expected infrastructure costs, which are related to our increased capital expenditure outlook and the higher expected cost of goods sold for Reality Labs. Given the more dynamic operating environment we are in, we maintain a range of $5 billion.

What I want to say is that our current investment stance reflects the significant opportunities we see internally within the company, as well as the priorities we are investing in this year. Clearly, we will continue to assess this, depending on how macro conditions evolve more broadly. But we do believe these are our major strategic focuses, which are critical for us and must continue to be invested in.

In fact, I think one of our goals in improving efficiency over the past two years has been to put us in a stronger financial position so that we can continue to invest in key priorities during tougher financial cycles.

Ross Sandler (Barclays):

Mark, in your many podcasts or keynote speeches yesterday, you mentioned a bunch of projects your team hopes or desires to complete are constrained by AI capacity bottlenecks, which Susan has talked about before, and even some tests the ad ranking team wants to run are delayed. So, looking ahead to this year, next year, or whenever, when do you expect to see this constraint alleviated? More broadly, we have passed three years since IDFA impacted your business. So, where do you think we are in terms of overall improvements in the ad ranking system and the ROI you can provide, and at what stage do you think we are? Thank you very much.

Susan Li:

I can try to answer both of those questions, Mark, and you can certainly jump in. Regarding the first question, our current capacity landscape is very active, reflected in many changes in our capacity expansion as well as the demand from different product departments within our company, whether they are in the Gen AI team or doing more core AI work related to ranking and recommendations. Therefore, both supply and demand are very volatile, so we do not have a fixed answer as to when we expect to have enough supply to meet all the demand, but we are working very hard to alleviate this issue, which is also part of the reason we are accelerating the increase in data center space coming online this year. We are also very focused on improving the efficiency of our workloads this year.

Regarding your second question about ad performance and ad ranking, we have been investing for years and continue to invest in driving improvements in ad performance. Year-over-year conversion rates remain strong, and in fact, we continue to see the growth rate of conversion rates outpacing the ad impressions in the first quarter, reflecting the improvement in conversion rates. Improvements in ad ranking and modeling are the main drivers of overall performance enhancement. We have made many innovations in model architecture during the ad retrieval and ranking stages of the ad delivery process to provide people with more relevant ads

In the first quarter, we introduced a new GEM advertising recommendation model, and we discussed some previous model architecture improvements in the past quarters, such as Lattice and Andromeda. For us, we really believe that advertising is a relative performance game, which is especially important for us because most of our business is driven by direct response advertising. So we are pleased that our previous investments are paying off, and we continue to invest in many different initiatives to continuously improve our advertising rankings and recommendation performance.

Ken Gawrelski (Wells Fargo):

Thank you very much. First, perhaps Mark. How should we think about the timing of driving the adoption of AI capabilities for WhatsApp for Business in a labor market with high labor costs? What measures is Meta taking to accelerate this adoption? Do you see this primarily as incremental to the small and medium-sized business advertising spend you have already captured?

Then, Susan, please comment on this. What does the revised capital expenditure outlook for 2025 mean this year? What does it mean, or you mentioned this as an acceleration in your revised outlook statement. Should we view this as a new starting point—considering 2026 and beyond? Or should we restart in 2026 and consider the needs and capacities at that time? Thank you.

Susan Li:

No, I’m happy to answer, and I will continue to address both questions, Mark, feel free to jump in at any time. Mark touched on a bit about our overall vision, which is that every business will soon have an AI that is an expert about its business to talk to its customers, just like they have email, websites, social media presence, and so on today.

We are currently testing business AI with a limited number of businesses in the U.S. and several other countries on WhatsApp, Messenger, as well as Facebook and Instagram ads. We started with small businesses, focusing first on helping them sell their goods and services through business AI. But ultimately, we are developing tools to support businesses at every stage of the customer journey (from lead generation to order management and customer service), and one core area we are currently addressing is the ability for businesses to customize and control agents to achieve the results they want. Therefore, we launched a new agent management experience and dashboard that makes it easier for businesses to train their AI-based tools based on existing information from their website, WhatsApp profile, or their Instagram and Facebook pages, and we first allow businesses to activate AI in their chats with customers.

We are also testing business AI on Facebook and Instagram ads, where you can ask questions about products and return policies or help you make purchases in our in-app browser. Again, the ultimate vision is to build an experience that can serve customers across all these different services and applications No matter where you interact with enterprise AI, it should be an agent that can access your history and preferences, and we are hearing encouraging feedback, especially regarding the adoption of these AIs, which are saving a significant amount of time for the enterprises we are testing and helping to identify which conversations are meaningful to them so they can spend more time on those.

Then, your second question is about capital expenditures in 2026. As I mentioned earlier, due to the ongoing development of AI, and for us, we continue to discover many good use cases to allocate capacity for our core AI ranking and recommendation work, infrastructure is just a very dynamic planning area. Therefore, I would say it is still too early to discuss plans beyond 2025.

Youssef H. Squali (Truist Securities):

Great. Thank you for taking the questions. So Mark, in a world where we might now have 5 to 10 chatbots on our smartphones, including Meta AI, all doing almost the same thing, do you think this is a market very similar to search, where the winner takes all, or is it more likely to become more decentralized? But in either case, what do you think are the top two to three competitive advantages of Meta AI?

Then, Susan, regarding the EU's decision related to the DMA, what modifications do you need to make to the applications? Can you help us assess the potential financial impact, although it may still be too early? Thank you.

Zuckerberg:

Yes, regarding Meta AI, I mean, I think people will use many different agents, just as they use different applications to do different things. I'm not sure if people will use multiple agents to do exactly the same thing, but I think that things more focused on enterprise productivity may differ from those optimized for personal productivity, and that may be slightly different from those optimized for entertainment and social connection. So I know there will be—there will be different experiences. One trend that I think we are starting to see now is personalization across these experiences. Now, if the experience is not personalized, then you can go to different applications and get reasonably similar answers to different questions, but once AI starts to understand you and what you care about in context, and can accumulate memory from your past conversations with it, I think that will start to become different. So that’s a point we think is important.

Then, of course, there are all the different modalities, not only being able to answer questions about text but also being able to engage in voice and multimodal interactions, and being able to generate images and videos and understand all of that, and have good conversations about it, I think that will be important overall. So yes, I mean, I think Meta AI is in a favorable position, but we have a lot of work to do to make it a leading personal AI

Susan Li:

Regarding your second question, it may be too early to talk about what these changes might be, as we are in communication with the European Commission. I think the most useful indicator I can give you is that our advertising revenue in the affected geographic areas, namely the European Economic Area and Switzerland, will account for 16% of our global total revenue in 2024. Again, we are continuing to actively communicate further with the European Commission. So we hope to have more clarity in the next conference call.

Mark Mahaney (Evercore ISI):

Thank you. I just want to add two questions. What are the advertising trends in vertical industries? And regarding Reality Labs and the losses associated with Reality Labs, they have been very consistent, anyway, at $4 billion each quarter, for a long time. Is there light at the end of the tunnel? Is there reason to think? Are there factors that could lead to a decrease in these losses, when might that happen, but more importantly, what would cause these losses to decrease? Thank you very much.

Susan Li:

Mark, let me answer your first question about other verticals. Generally, we have seen healthy growth in most verticals in the first quarter. We have indeed seen some weakness in gaming and politics. Overall, this is still just a very small vertical.

Zuckerberg:

I mean, we are basically focused on getting things done more efficiently, but as AI glasses really take off, as I have talked about many times, I think it makes sense to invest more to ensure we can distribute it and grow it very quickly. I mean, if you look at some leading consumer electronics in other categories, by the time they develop to the third generation, they usually sell 10 million units and start to expand from there.

I'm not sure if we will do it exactly like that, but I think that's roughly the scope of the opportunity we have, and I think we are focusing on expanding to that point and then expanding for the subsequent generations.

So I think some of the efforts we are making will— we will become more efficient in some of the work we are doing, but when a bunch of those products start to come out and begin to grow, even larger than the numbers I just mentioned, that will be like a recent milestone, and then I think we will continue to expand in distribution, and at some point, just like with other products we have built, we will feel that we have reached sufficient scale, and we will primarily focus on ensuring we are monetizing and building an effective business around it.

But that's roughly where we are at the moment. We are definitely focused on getting things done more efficiently, but we are also very optimistic about the results we are seeing, especially with AI glasses