Meta (Minutes): Investing in AI without hesitation, planning to invest 100 billion next year

The following is organized by Dolphin Research regarding$Meta Platforms(META.US) the Q2 2025 earnings call minutes, for financial report interpretation please visit《Meta's Unprecedented Surge!》

I. Review of Core Financial Information:

II. Detailed Content of the Earnings Call

2.1. Key Information from Executive Statements:

1. Business Progress

1) Overall Performance

- Over 3.4 billion users globally use Meta applications daily, maintaining high overall engagement.

- AI systems have begun to show signs of self-improvement, with a slow but clear trend of enhancement.

- Meta's vision is to bring personal superintelligence to everyone, developing "superintelligent" AI that surpasses human intelligence, believing it has the potential to usher in an exciting new era of personal empowerment.

- Meta Superintelligence Labs has been established, including foundational, product, fairness teams, and a new lab focused on developing next-generation models.

- Progress on Llama 4.1 and 4.2 models is going well, with next-generation models under development, expected to push the frontier in about a year.

- An elite, talent-intensive AI team is being formed, led by Alexandr Wang, with Nat Friedman responsible for AI product and application research, and Shengjia Zhao as chief scientist.

- Meta possesses all the elements needed to build leading models and deliver them to billions of users, including unparalleled computing power.

- The Prometheus cluster will go online next year, expected to be the world's first gigawatt-level cluster; Hyperion will expand to five gigawatts in the coming years; multiple Titan clusters are under development.

- Firm belief that super AI will improve every aspect of the company's operations.

2) Application Product Side

a. WhatsApp: As the largest driver of Meta AI query volume, users send messages directly to Meta AI for tasks such as information gathering, homework help, and image generation. Business message revenue growth on the WhatsApp business platform contributed to other revenue growth in the Family of Apps.

b. Facebook: User activity continues to grow, especially in the U.S. market where video watch time increased by over 20% year-on-year.Meta AI usage on Facebook is expanding, with people using it to inquire about posts seen in their news feed and to search for content on the platform and in search.Advancements in recommendation systems have increased Facebook usage time by 5%.

c. Instagram: Global video watch time increased by over 20% year-on-year. Recommendation system advancements have increased Instagram usage time by 6%. Over two-thirds of recommended content in the U.S. now comes from original posts.

d. Threads: Community and interaction depth continue to grow, with the integration of LLM (large language models) significantly driving growth in viewing time related to rankings on Threads. As of May, global advertisers can now place video and image ads to Threads users in most countries, including the U.S.

3) Meta AI

a. Meta AI Improvement: AI-driven content recommendation systems have significantly improved content quality. Meta AI is now available in over 200 countries and regions.

b. Llama Models: Llama and several other backend processes are being utilized, including handling error reports, which have led to a roughly 30% decrease in top-line error reports in Facebook news feeds and notifications in the U.S. and Canada over the past 10 months.

c. Investment Planning: The company will heavily invest in AI, believing superintelligence will improve every aspect of the company's operations.

4) Reality Labs and Wearable Devices

a. Ray-Ban Meta Glasses: Continued strong performance, with accelerating sales growth. Sales growth accelerated in the second quarter, with the most popular SKU still in short supply, and the company is working to increase supply to meet consumer demand.

b. AI Glasses: Launched new high-performance AI glasses, Oakley Meta Houston's, featuring longer battery life, higher resolution cameras, and designed specifically for sports.

c. Quest Ecosystem: People continue to spend more time in the Quest ecosystem, with community growth steadily increasing.Last month, the Meta Quest 3S Xbox edition was launched, and record interest in cloud gaming was observed.In addition to gaming, media and web browsing also contributed significantly to engagement.

5) Business Outlook

a. User Experience and Interaction: Committed to continuously increasing daily active users on Facebook, Instagram, and WhatsApp through further improvements in recommendation systems and product experiences.

b. Short-term Strategy: Further improvements are expected throughout the year as model scale expands and recommendation content becomes more aligned with user interests during sessions. The focus in the second half of the year will be on further enhancing the freshness of original posts.

c. Long-term Strategy: Long-term ranking innovation is progressing well, expected to bring the next wave of improvements in the coming years. Research work on cross-platform foundational recommendation models continues to advance.

d. Meta AI: As models continue to improve, user engagement continues to grow. Meta AI is expected to help content discovery by automatically translating and dubbing foreign language content into the audience's local language.

e. Threads: User and interaction depth continue to grow.

6) Monetization Efficiency

a. Ad Placement Optimization: Continued optimization of ad supply on each platform to better deliver ads at the most relevant times and places. Ads have begun to be introduced in Threads' news feed and WhatsApp's "Updates" tab in the second quarter. Initial ad supply on WhatsApp is expected to be low, with low monetization capability, and is not expected to make a significant contribution to total impressions or revenue growth in the short term.

b. Ad Performance Improvement:

- AI has unleashed higher efficiency and revenue in the ad system, driving strong performance this quarter.

- The new AI-driven ad recommendation model has expanded to new display surfaces and improved performance by using more signals and longer context,driving approximately 5% ad conversion rate on Instagram and 3% on Facebook.

- Enhancements and expansion of the Andromeda model architecture to Facebook Reels have increased conversion rates for Facebook Mobile Feed and Reels by nearly 4%.

- Improved performance of the new generative ad recommendation system GEM increased ad conversion rates on Instagram by about 5% in the second quarter, and Facebook Feed and Reels by 3%.

- Expansion of the Lattice model architecture coverage increased ad conversion rates for Facebook Feed and Reels by nearly 4% in the second quarter.

c. Advantage+ Solutions: Generative AI-driven ad creative tools have achieved strong advertiser retention rates.

- Promotion of Advantage+ sales and app campaigns to simplify the ad campaign creation process has been completed.

- Nearly 2 million advertisers are using video generation, image animation, and video extension features.

- Testing of AI-driven translation features began in the second quarter, allowing advertisers to automatically translate ad copy into 10 different languages.

- Incremental attribution features have been globally promoted.

- Omnichannel ads have been globally launched and are in testing, with advertisers using omnichannel ads seeing a median total cost per purchase decrease of 15% compared to optimizing solely for websites.

Financial Performance

1) Overall Overview

a. Revenue: Total revenue for Q2 2025 was $47.5 billion, up 22% year-on-year both in reported and constant currency terms.

b. Expenses: Total expenses were $27.1 billion, up 12% year-on-year. Cost of revenue increased by 16%, R&D expenses increased by 23%, marketing and sales expenses increased by 9%, and general and administrative expenses (G&A) decreased by 27%.

c. Net Income: Net profit was $18.3 billion, with earnings per share of $7.14.

d. Operating Profit Margin: Operating profit was $20.4 billion, with an operating profit margin of 43%.

e. Free Cash Flow and Capital Expenditure: Free cash flow reached $8.5 billion. Capital expenditure (including principal payments on finance leases) was $17 billion,primarily driven by investments in servers, data centers, and network infrastructure.

f. Stock Repurchase and Dividends: $9.8 billion of Class A common stock was repurchased this quarter, and $1.3 billion in dividends was paid to shareholders.

g. Cash and Marketable Securities: At the end of the quarter, $47.1 billion in cash and marketable securities was held, with liabilities of $28.8 billion.

2) Family of Apps

a. Revenue: Q2 revenue was $47.1 billion, up 22% year-on-year.

- Advertising Segment: Advertising revenue was $46.6 billion, up 21% year-on-year (22% in constant currency terms).

The online commerce verticalwas the largest contributor to year-on-year revenue growth.

ARPU growth was highest in Europe and the rest of the world, increasing by 24% and 23% respectively; North America and Asia-Pacific grew by 21% and 18% respectively.

Ad impressions increased by 11%, primarily driven by the Asia-Pacific region.

Average price per impression increased by 9%.

- Other Segment: Family of Apps other revenue was $583 million, up 50%, primarily driven by growth in WhatsApp paid message revenue and Meta-verified subscriptions.

b. Expenses: Family of Apps expenses were $22.2 billion, up 14% year-on-year, accounting for 82% of total expenses.

c. Operating Profit: Family of Apps operating income was $25 billion, with an operating profit margin of 53%.

3) Reality Labs Business

a. Revenue: Q2 revenue was $370 million, up 5% year-on-year,primarily driven by growth in AI glasses sales, partially offset by a decline in Quest sales.

b. Expenses: Reality Labs expenses were $4.9 billion, up 1% year-on-year, primarily driven by increased non-personnel-related technology development costs.

c. Operating Loss: Reality Labs operating loss was $4.5 billion.

4) Capital Allocation Strategy

The main focus is on reinvesting capital into the business, with infrastructure and talent development as top priorities.

a. Personnel Adjustment Plan: As of the end of the second quarter, the company had over 75,900 employees, a 1% decrease quarter-on-quarter. However, the company continues to recruit in priority areas such asmonetization, infrastructure, Reality Labs, AI, and regulatory and compliance.

The increase in talent in all priority areas is expected to continue driving overall employee growth in 2025 and 2026, while employee growth in other functional areas will remain limited.In the AI field, there is a particular emphasis on recruiting top talent in the industry to build Meta Superintelligence Labs, accelerating AI model development and product planning.

b. Capital Expenditure Plan: Having sufficient computing power is expected to be central to realizing many of the company's biggest opportunities in the coming years. AI capacity investments in core advertising and organic interaction plans bring attractive returns, with continued significant investment expected in 2026. Developing leading AI infrastructure is a core advantage for developing the best AI models and product experiences,thus similar significant dollar growth in AI field investment is expected in 2026.

5) Q3 2025 and Full-Year Guidance

a. Revenue Guidance: Total revenue for Q3 2025 is expected to be between $47.5 billion and $50.5 billion. The guidance assumes a favorable impact of approximately 1% from foreign exchange on year-on-year total revenue growth. Year-on-year growth in Q4 2025 is expected to be slower than Q3, as it compares to a stronger growth period in Q4 2024.

b. Expense Guidance: Total expenses for 2025 are expected to be between $114 billion and $118 billion, narrowed from the previous forecast of $113 billion to $118 billion, reflecting a year-on-year growth rate of 20% to 24%. The total expense growth rate for 2026 is expected to be higher than that of 2025, mainly driven by infrastructure costs (depreciation expenses and operating costs will accelerate sharply) and employee compensation (increase in technical talent in priority areas and confirmation of employee compensation expenses for full-year recruitment in 2025).

c. Capital Expenditure: Capital expenditure for 2025 (including principal payments on finance leases) is expected to be between $66 billion and $72 billion, narrowed from the previous forecast of $64 billion to $72 billion, with a median year-on-year increase of approximately $30 billion. Similar significant dollar growth in capital expenditure is expected in 2026 to increase online capacity and meet the needs of AI work and business operations.

d. Tax Rate: U.S. federal cash taxes are expected to decrease for the remainder of this year and in the coming years. The tax rate for 2025 is expected to be higher than the second quarter's rate.

e. Regulatory Risk: Closely monitoring legal and regulatory challenges, including those in the EU and the U.S., which could have a significant impact on business and finances. For example, communication with the European Commission regarding "Less Personalized Ads" (LPA) is ongoing, and the Commission may seek further modifications, which could substantially worsen user and advertiser experiences and potentially have a significant negative impact on revenue in Europe as early as later this quarter.

2.2. Q&A Analyst Questions and Answers

Q: Regarding the evolution of the Meta AI business unit over the past three to six months, what are the key learnings after delving into this strategy? How have these learnings influenced shifts in talent recruitment and computing resources, and how has the strategy evolved based on these key learnings as mentioned in your recent blog post?

A: From a macro perspective,we have observed that the more aggressive or fastest assumptions in AI progress have often most accurately predicted what actually happened, and this year is no exception.The internal team has made significant progress in adjusting Llama 4 to build autonomous AI agents that can help improve Facebook algorithms, enhance content quality, and increase user engagement, although the scale is currently small, the development trajectory is very optimistic.

We believe the world will be very different in the coming years, with super AI soon shaping all our systems, changing many of the fundamental assumptions of work within the company. The continued acceleration of AI progress supports our emphasis on having the best and most elite talent and teams, and ensuring the necessity of having leading computing clusters. We are committed to providing researchers with more computing resources per capita so they can lead research and extend results to billions of users across our product lines.

Q: Based on Mark's comments about expanding talent and computing resources, can you explain more deeply how we should view these two components driving operating expenses and capital expenditures over the next 12 to 18 months?

A: Given this is a very dynamic operating environment, we have not yet initiated the budgeting process for 2026, but we have a rough understanding of the infrastructure plans for 2026 and the growth in AI talent compensation expenses.Infrastructure will be the largest single contributor to expense growth in 2026.

This is mainly driven by the sharp acceleration of depreciation expenses in 2026, including incremental depreciation of assets purchased and put into use in 2026, and full-year depreciation confirmation of infrastructure deployed in 2025.

Additionally, the proportion of short-term assets in capital expenditures for 2025 and 2026 is expected to be higher than in previous years. Infrastructure cost growth also includes higher operating expenses such as energy, leasing, and maintenance costs, as well as increased operational expenses related to cluster maintenance and cloud service spending.

Employee compensation is the second largest driver of expense growth in 2026, mainly driven by investment in technical talent, including full-year compensation expenses for AI talent recruited this year. In terms of capital expenditure, the main driver of increased capital expenditure in 2026 will be the expansion of generative AI capacity for building training capacity, leading to increased spending on servers, networks, and data centers.

Q: Regarding the vision of intelligent labs and superintelligence, how do you view it now compared to 12 months ago? Can you introduce what technical limitations or bottlenecks you are most focused on overcoming in the next 24 months, which may differ from the past, to ensure you truly lead the concept of superintelligence in the next 10 years?

A: In the research agenda and many areas we are highly focused on, we do believe that focusing on "self-improvement" is a very important research direction. For developing superintelligence, it cannot simply learn from humans, as it aims to build entities smarter than humans, thus needing to learn how to self-improve.

We believe this is a very foundational part of building superintelligence, which will have profound impacts on how we build products, operate companies, invent new things, make new discoveries, and broader society. In terms of the overall effort, I am more convinced that small, talent-intensive teams are the best configuration for driving frontier research. This differs from the model where we efficiently use hundreds or even thousands of people to improve systems in Instagram, Facebook, or ad systems. For frontier research in superintelligence, what is truly needed is the smallest team capable of comprehensively mastering the entire system.

Q: You have made many improvements to core products to enhance engagement, recommendation effectiveness, etc. Can you introduce what factors you most anticipate will further enhance engagement on core platforms in the next 18 months?

A: Regarding the forward-looking roadmap of the core recommendation engine, our recent short-term focus includes:focusing on making recommended content more aligned with user interactions during sessions, making it more relevant to what users are most interested in at the time.We are optimizing to help the best content from small creators stand out by matching it to the right audience more quickly.

At the same time, we are also working to improve the system's ability to discover more diverse and niche interests for everyone through interest exploration and learning clear user preferences. We also plan to further expand model scale and incorporate more advanced technologies to improve the overall quality of recommendations. In terms of long-term investment, we are developing foundational recommendation models that support recommendations across multiple services and integrating large language models (LLM) more deeply into our recommendation systems. A major focus of this work will be optimizing the system to make it more efficient, allowing us to continue expanding recommendation system capacity without affecting return on investment (ROI).

Q: Meta has been a strong advocate of open-source AI. As you pursue superintelligence and seek greater returns on your massive infrastructure investments, has your thinking changed in this regard?

A: Regarding open source, we believe the company's thinking has not particularly changed; we have always open-sourced some models, but not everything we do is open source.We expect to continue producing and sharing leading open-source models.There are also some trends, such as models becoming so large that they are impractical for many others, so we consider whether sharing their productivity or primarily helping competitors.

Additionally, as we approach true superintelligence, there are a series of completely different security issues that need to be taken very seriously. We expect to continue open-source work, maintaining a leading position, while not all work is open source, which is consistent with our long-standing approach.

Q: Your comments on 2026 capital expenditures suggest spending may exceed $100 billion next year. Do you still expect to finance it all yourself, or are there opportunities for collaboration?

A: Regarding the financing of next year's growth in capital expenditures, we certainly expect to bear a large part of it ourselves. But at the same time, we are exploring ways to collaborate with financial partners to jointly develop data centers.Currently, we have no finalized deals to announce, but we generally believe there will be some models that can attract substantial external financing to support large data center projects. These projects can leverage our ability to build world-class infrastructure for development while providing us with flexibility as our infrastructure needs change over time. Therefore, we are exploring many different avenues.

Q: Your spending is now approaching some of the largest hyperscale providers. Do you think all this capacity is mainly for internal use, or do you see any way to share, or even come up with a business model to utilize this capacity for external use?

A:Currently, we are focused on ensuring sufficient capacity to meet our internal use cases.This includes all core AI work we do to support organic content recommendations, all ad ranking and recommendation work. We are also committed to building the necessary advanced AI model training capabilities and ensuring readiness for future potential inference use cases.

Inference use cases not only focus on developing advanced models but also consider how to expand to consumer use cases that are expected to be widely useful and attractive to users.Currently, we are not really considering external use cases for infrastructure.

Q: When considering the return on investment (ROI) of these capital expenditures, I believe you have internal models. How do you view ROI, and are you optimistic about long-term returns?

A: Regarding the ROI of capital expenditures, in core AI, we continue to see strong ROI, and our ability to measure this is quite good, and we are very satisfied with the rigorous measurement and the returns obtained.In generative AI, we are clearly still in the early stages of the return curve and do not expect generative AI work to be a significant driver of revenue this year or next.But we are generally optimistic about the monetization opportunities that are about to open up.

Mark also mentioned these opportunities in his speech, and we believe that in the medium to long term, these opportunities are very close and intuitive to the current state of our business, so they will be our huge opportunities, and each opportunity will have a huge market. When building infrastructure, we consider interchangeability. While many aspects such as data center rooms and network infrastructure need to be built in advance, we will order servers as needed,servers will ultimately be the largest expenditure in capital expenditures, and we will make the best decisions as needed to determine where the capacity goes.

Q: When pursuing the vision of superintelligence, what markers or KPIs are you tracking to determine if you are on track and making progress? Is it really related to the five pillars you outlined above, or should we think more broadly?

A: Internally, the indicators I focus on include: the quality of team members, the quality of models we are producing, the speed of improvement of our company's other AI systems, and the extent to which leading foundational models facilitate improvements in all other AI systems and everything we do. I believe this will naturally translate into our standard product and business strategy, turning the technology into new products that will first expand to billions of people and then be monetized over time.

I believe there will be a lag period, which is also our consistent way of working, whether developing new social products or products like Meta AI. We will strive to reach leading scale, build the highest quality products, and focus on this for several years, and then once we are truly confident in our position, we will focus on developing business around it. We believe that if you are building superintelligence, you should use all your GPUs to ensure you can serve your customers well with it, and by generating directly, you can achieve higher returns.

Q: AI now brings huge ROI, and all these investments are also working towards long-term goals. Do you have any changes or adjustments regarding the relationship between revenue or core business performance and the pace of investment?

A: We have previously stated that from a profitability perspective, our main focus is on driving long-term growth in consolidated operating profit, which will not be linear. In some years, we will achieve above-average profit growth, while in years when we make significant investments, I believe this will affect the amount of operating profit growth we can achieve.

Currently, we see many attractive investment opportunities that we believe will allow us to bring substantial profit growth to all investors in the coming years. Therefore, when we pursue these investments, we focus on limiting investments in other areas. We do believe that now is the time to truly invest in the future of AI, as this will not only open new opportunities for us but also strengthen our core business.

Q: Overall engagement with Meta AI is growing, especially on WhatsApp, where we have a billion users on our platform. The current focus is on driving personalization. I would like to know more about how these next-generation models help drive adoption here, especially as the AI flagship model Behemoth goes live at some point. What are your thoughts on search and queries and potential monetization when people use AI on WhatsApp?

A: I won't delve into the specific roadmap in this area, but the basic situation is,we do see that as we continue to improve the models behind Meta AI and conduct post-training, user engagement increases.When we switch to updated models, such as from Llama 4 to Llama 4.1, we expect these models themselves to be highly versatile.

Therefore, while it can focus on specific areas, overall, it will become better at many different things people want to do. I believe each version, whether it is the continuous training we conduct weekly or when we release new generations or major version updates, will continuously improve user engagement. We are focused on this and are very excited about it.

Q: The Ray-Ban project has been a flagship achievement for you so far. How are we progressing in the development of glasses and the new computing platform you have talked about in the past? Is it faster or slower than you imagined? When you leverage Meta AI, do you think glasses will eventually replace smartphones, or do you need a new AI-first form?

A: Ray-Ban Meta and Oakley Meta, as well as other products we plan, are performing well. This product is not only a stylish pair of glasses with interesting features, but the use of Meta AI in them continues to grow, with the percentage of daily users increasing, all of which is delightful.

I think glasses will essentially be the ideal form for AI because you can have AI see what you see, hear what you hear, and talk to you all day.Once a display is added to them, whether it's a wide holographic view or a smaller display, it will unlock a lot of value, enabling all-day multimodal interaction with AI systems. I personally believe that in the future, if you don't have glasses with AI capabilities or interact with AI in some way, you may be at a cognitive disadvantage compared to others you work or compete with. This is also the area we have been investing heavily in research at Reality Labs for the past five to ten years. Additionally, glasses will be the ideal way to merge the physical and digital worlds, and the entire metaverse vision will eventually become extremely important, and AI will accelerate this process.

Q: How do you view the progress of SBC (stock-based compensation) in the coming years? Can it be assumed that it will grow faster than revenue and operating expenses? And how do you minimize shareholder dilution?

A: The impact of increased compensation costs, including this year's stock-based compensation (SBC), has been reflected in the revised 2025 expense outlook. These are clearly also a major driver of expense growth in 2026, as we confirm full-year compensation for newly recruited talent.

Nevertheless, we are very focused on closely monitoring the dilution effect.We generally believe that our strong financial position will allow us to support these investments while continuing to repurchase shares as part of the buyback program to offset stock-based compensation, and provide quarterly cash dividend distributions to our investors.

Risk Disclosure and Statement of this Article:Dolphin Research Disclaimer and General Disclosure