Meta earnings call discloses Muse agent progress: over 10 million users in the US and Canada, plans to improve compute efficiency 10x within a year

Article body EM_StockImg_Start

EM_StockImg_End

After the market closed on Wednesday, US Eastern Time, Meta released its third-quarter 2026 financial results and held an earnings call. The company's overall revenue and fourth-quarter guidance both exceeded market expectations.

The biggest highlight of this earnings report is clearly Meta's personal AI agent Muse. After delivering impressive early user and monetization data, CEO Mark Zuckerberg said plainly on the earnings call that the company plans to increase Muse's compute efficiency tenfold within a year and is expected to achieve substantial profitability within a few years.

However, while looking ahead to AI's prospects, compute spending, Muse's path to profitability, and the pace of narrowing Reality Labs' losses also became the core topics that Wall Street institutional investors pressed on.

Key takeaways from the earnings call

I. Core financials

1. Q3 total revenue was $64.6 billion, up 26% year over year, with earnings per share of $3.83. Expenses rose sharply due to a one-time $9.9 billion litigation settlement charge.

2. Advertising remains the foundation of Meta's revenue, with ad revenue of $63.1 billion, rising in both volume and price. Reality Labs revenue was $443 million, and its loss is expected to narrow to about $15 billion next year.

3. Guidance: Q4 revenue of $73-76 billion; full-year capital expenditures narrowed to $135-145 billion; free cash flow was negative this quarter.

1. User numbers: In the US and Canada, more than 10 million users have set it up within a month and a half of launch, 25% of users use it daily, and retention is better than similar products; shopping is the top use case, with average order value exceeding $200, including travel.

2. Current status: Compute costs are high, and it has not yet rolled out globally; the plan is to increase compute efficiency tenfold within a year, with substantial profitability expected within a few years.

3. Model: Basic features are currently free, monetized through transaction commissions and premium/enterprise subscriptions; it is advancing enterprise APIs and adaptation for small and medium businesses, with the long-term goal of building an interconnected personal-enterprise agent network.

1. Daily active users across the company's family of products reached 3.64 billion, and Threads' monthly active users have surpassed X.

2. On the advertising side, AI tools continue to improve conversion, and annualized revenue from creator partnership ads has exceeded $15 billion.

3. Hardware: Smart glasses will integrate Muse next month; VR is slowing its iteration pace, with resources prioritized toward AI.

IV. Capital and Q&A highlights

1. A new non-voting Class C common stock was introduced, equivalent to a 2-for-1 split, effective in January 2027, to be used for financing, employee incentives, and M&A, safeguarding the founder's control; there are no specific financing plans at present.

2. Analysts focused on Muse's compute bottlenecks and commercial friction in agent-based e-commerce; Meta and industry partners launched an open protocol for agent interaction; the AI business may begin contributing incremental revenue in 2027, but large-scale commercialization will still take time.

Full transcript of Meta's Q3 2026 earnings call (AI-assisted translation, some content abridged)

Rob Bhavani (Director of Meta Investor Relations): Good afternoon, and welcome to Meta Platforms' third-quarter 2026 earnings call. Joining me today to share results are Chief Executive Officer Mark Zuckerberg and Chief Financial Officer Susan Li. I'll now turn it over to Mark Zuckerberg.

Mark Zuckerberg (Meta CEO): All right, hello everyone, and thanks for joining. The past few months have been busy and exciting. Daily active users across our family of apps have reached 3.64 billion; thanks to rapid growth in video, engagement on both Facebook and Instagram hit all-time highs. Threads' monthly active users have surpassed X and it is still expanding quickly. WhatsApp messaging volume hit a record, and it is gradually becoming an important platform for running AI agents.

Over the past few months, we have hit several important milestones toward our goal of "personal superintelligence for everyone." The Muse personal agent became one of the most successful product launches in our history; we launched our first frontier model, Muse Wonder; and we released several smart glasses and hardware devices with built-in Muse capabilities, with all-new audio glasses coming soon.

We have been discussing the vision of personal superintelligence for more than a year. We want everyone to have a capable personal agent: one that understands your goals and helps you achieve them and improve your life—whether it's maintaining relationships, saving time and money, running a business, creating creatively, or bringing all kinds of ideas to life. Muse is the product embodiment of that vision. The underlying foundational technology is now ready, there is still a lot of work ahead, but we are off to a very strong start.

Just a month and a half after launch, more than 10 million users in the US and Canada have completed setup of the Muse personal agent, and user growth momentum is strong, with the app holding the top spot on the App Store almost every day.

Usage activity and retention have started off impressively: about 25% of users who complete Muse setup use the product on any given day. The longer early users use it, the better retention gets. For reference, that retention level is more than double that of Threads in its early days and higher than the starting level of other AI apps on the market.

The most mainstream use case for Muse is shopping. More than a quarter of Muse users perform shopping-related actions, with an average order value of over $200. Payment service provider Plaid reported that it has never seen a connector generate business traffic on such a large scale on any other service. In the first week after launch, we also had to work with Stripe to raise our credit limit; the actual business scale exceeded everyone's expectations.

Beyond shopping, other high-frequency use cases include personal productivity tools, as well as connecting Muse to third-party services such as email and calendars. This also highlights the value of our self-developed Muse secure virtual machine, which provides industry-leading privacy and security protection. Soon we will also launch the Muse confidential virtual machine; once enabled, even Meta itself will not be able to read user content.

Muse Wonder further improved answer quality, user activity, and retention. This signals that, as model intelligence continues to iterate, Muse's product performance will keep improving. Muse has now validated product-market fit, and we are entering a new, more operations-focused phase of scaling. This new phase has three major goals: expanding the product's user base, increasing revenue, and optimizing compute costs.

Our goal is to serve billions of users with Muse. This is both a mission and a strategy. From the mission perspective, we believe personal superintelligence should be placed in everyone's hands; individual access to technology is the source of social prosperity, and checks on power are the cornerstone of safety—this is the path to an inclusive and better future. In terms of values, we believe promoting this technology is our most important undertaking.

On the business side: the average annual ARPU (average revenue per user) of existing users of our full suite of apps in the US reaches several hundred dollars. Based on the scale of e-commerce transactions already emerging, we expect Muse's long-term ARPU level to be even higher. Achieving this goal will take time, but we have already begun laying the groundwork.

The business model consists of multiple parts: the vast majority of users can use the product for free; we build e-commerce tools and take a commission from the transactions we facilitate, with merchants bearing this cost. Users who need access to top-tier models and large amounts of token compute can pay for a subscription directly. Business operators who want to use Muse to expand their businesses, or companies that need to use Muse in work scenarios, can also pay for a subscription. Users can also connect Muse to the Meta ecosystem to optimize the personalization and quality of content and ads.

A supplementary note: the ARPU figure mentioned above refers to revenue generated by Muse users across Meta's full suite of apps. On compute, over the next few years we are building a compute foundation capable of serving more than 1 billion users.

Running Muse consumes enormous compute in the short term, but we expect compute efficiency to improve more than 10x over the next year. Once revenue scales up, the product can become profitable. If compute efficiency can continue to improve rapidly over the next few years—which we judge to be highly likely—Muse can achieve high profitability within a few years.

In the short term, relying on the strong cash flow from our core business, we can complete the expansion of the Muse network first, and we also believe Meta is the company best suited to execute this strategy.

One more point on compute economics: model training is a large fixed cost. Given that we hold top-tier models, we should take on as much high-margin inference business as possible. That is why we launched the Meta enterprise platform led by CJ Desai.

Muse Wonder and stronger models trained later can, in addition to serving the Muse network, also be offered externally through API interfaces, Muse Code, and other high-margin businesses such as enterprise services. One of our major advantages is that compute can serve a wide variety of business needs.

It is difficult to precisely predict the growth rate of the Muse network and the enterprise platform, or to quantify how much models such as Muse Wonder will drive our core business. But by releasing advanced model capabilities through multiple channels, at least some of these businesses are highly likely to exceed expectations.

Beyond serving enterprise customers externally, we have not yet reached the upper limit of how compute can empower our own core business. We have already begun testing the use of Muse Wonder to autonomously iterate and optimize our recommendation and ads systems. As the capabilities of the next generation of models continue to strengthen, the iteration speed of such systems will accelerate further.

Next, let me talk about the models. Muse Wonder reaches top-tier levels across many key dimensions required for personal superintelligence. This confirms that self-developed training models are the only way to give users the best experience. In addition to Muse, since the new models rolled out in April, average time spent per user on Meta AI has doubled.

Muse Wonder is the most important version iteration, and I am also satisfied with the overall execution pace. Over the past few months we have released Muse Spark 1.2 and 1.3, Muse Code, two open-weight models Muse Glimmer and Muse Spark Open, as well as a Muse voice version.

The industry is developing rapidly. Going forward, on the one hand we need to push toward the frontier of intelligence capabilities, and on the other hand distill frontier capabilities into smaller, more efficient product models for personal agents. My judgment: continued scaling is enough to achieve superintelligence, and we are advancing step by step in an orderly manner.

A larger-scale model beyond Muse Wonder is already in development and will be extremely powerful; at the same time, we are also developing a new lightweight model that will be released soon, achieving Muse Wonder-level intelligence with higher efficiency.

Now, personal superintelligence has also become a core part of the Reality Labs division's work. Muse will launch on Meta smart glasses next month, which is an important step.

Users already love the glasses' existing features: music, calls, image capture, hearing assistance, and more. But in the long run, the personal superintelligence Muse that you carry with you will be the most central part of the hardware experience. Muse is likewise key to the glasses business model.

Our mission and strategy remain about serving the masses. We do not price hardware to earn high margins on the hardware itself; the goal is to build a large software and superintelligence business on top of the hardware. With glasses sales growing rapidly, combined with Muse, the paid subscription scale is expected to rise significantly, and the segment's operating condition will improve substantially as a result.

I mentioned the VR business strategy adjustment at Connect: we will only release new hardware when there is a major technological breakthrough, and this year's Meta VR glasses are that kind of product—the first ordinary glasses with an integrated VR experience. I still believe that the holographic experience blending virtual and real will start from VR technology. But we will adjust our investment pace and prioritize delivering personal superintelligence.

Taking all these factors together, Reality Labs' operating condition will improve. We expect operating losses to narrow to about $15 billion next year and to keep moving toward profitability.

Finally, a word on overall iteration speed. Fast iteration, fast learning, and serving users at scale have always been our methodology, and recently iteration has accelerated further. Over the past year, Edits launched and monthly actives have already surpassed 50 million; new products such as Forum and Seller were built by small teams and quickly serve deeply engaged users. Muse went from an idea on a whiteboard to a mature product in only about 8 months. The upcoming Muse Charm, from concept to mass-produced hardware in the field, took less than a year.

I am proud of what the team has achieved. Views on the future differ, but Meta believes technology should serve everyone, and putting capabilities in the hands of ordinary people is the way forward. This is crucial—the world needs companies that hold this belief. It is not just about ideas, but about execution. Thank you to all my colleagues for working to bring products to billions of people around the world. As always, thank you to everyone on this journey with us. I will now hand it over to Susan Li.

Susan Li (Chief Financial Officer): Thank you, Mark, and good afternoon, everyone. First, let's look at segment results; unless otherwise noted, all comparisons are year over year.

Q3 Family of Apps revenue was $64.2 billion, up 26% year over year; Family of Apps advertising revenue was $63.1 billion, up 26% year over year on both a reported and constant-currency basis. Total ad impressions across our platforms rose 10% year over year this quarter. Benefiting from growth in user scale and engagement, as well as optimized ad load strategy, ad impressions grew solidly in all regions. Driven by improved ad performance and a macro environment that improved from a year earlier, the global average price per ad rose 14% year over year; however, sharp impression growth in some lower-monetizing placements and regions partly offset the increase in average price.

Family of Apps other revenue was $1.1 billion, up 59% year over year, mainly from WhatsApp paid messaging and subscription businesses.

Reality Labs segment: Q3 revenue was $443 million, down 6% year over year, as Quest headset sales declined, but revenue from AI smart glasses continued to grow strongly, partly offsetting the pressure.

On a consolidated basis: Q3 total revenue was $64.6 billion, up 26% year over year on both a reported and constant-currency basis. Total expenses were $51.6 billion, up 68% year over year, including a $9.9 billion charge corresponding to the multistate attorneys general settlement reached in August 2026. The year-over-year increase in expenses mainly came from legal-related expenses, infrastructure investment, and third-party AI token costs. Higher infrastructure costs came from third-party cloud purchases, depreciation, and data center operating expenses.

Headcount was 66,000 at the end of Q3, down 12% from Q2 following the layoffs in May 2026. GAAP operating income was $13 billion, down 37% year over year, with an operating margin of 20%. Excluding the one-time $9.9 billion legal settlement charge, Q3 operating income rose 11% year over year. The tax rate this quarter was 24%, above the expected range, mainly due to the non-deductible portion of the legal expenses. Net income was $9.9 billion, or $3.83 per share.

Capital expenditures (including principal payments on finance leases) were $34.9 billion, concentrated in servers, data centers, and network infrastructure. Free cash flow was -$2.1 billion. Cash and marketable securities were $82.4 billion at quarter end, and debt was $83.7 billion.

Business performance is driven by two core things: first, building highly engaging product experiences for the community; second, continuing to monetize user engagement.

On the first point, content recommendation programs continue to deliver significant gains. Instagram time spent per user reached an all-time high, up double digits year over year in Q3, with both Reels short video and feed strengthening. Facebook Reels time spent hit new highs globally and in the US, up double digits year over year. Threads time spent per user rose more than 20% year over year.

Engagement rose across products, driven by continuous improvements in content relevance, personalization, and timeliness. The two major recommendation engines—retrieval and ranking—kept iterating, and both Instagram and Facebook saw positive results. For example: Instagram Reels launched relevance optimization in Q3, lifting Reels watch time by more than 4%; we upgraded the model architecture so that instead of scoring each short video individually, it scores a sequence of short videos together, while also optimizing for watch-time signals.

Facebook Feed and Reels completed the global rollout of a generative retrieval architecture, drawing on large language model approaches to tag content with semantic labels and proactively generate candidate recommendations. This makes it possible to surface high-quality new posts before they have accumulated much engagement data. The architecture now covers about a quarter of non-friend Feed and half of Reels traffic, driving a 1% increase in Reels watch time. The recommendation algorithms also make content fresher and more personalized.

In the Instagram Reels recommendation feed, nearly 45% of content was published less than a day earlier, up more than 11 percentage points from Q2. The Facebook Reels ranking model can now read twice as much historical content, lifting Reels watch time in the US by 1%.

Large language models also deepen content understanding. We use large models daily to process public Instagram posts, and by the end of the quarter this signal had lifted engagement with Instagram search results by 1%. This quarter we began switching our content understanding models to our in-house Muse model family. Offline evaluations show the new models outperform the old production models; at the same time, per-unit content processing costs fell by more than 80% across Instagram, Threads, Facebook, and the ads business. Owning the full model stack also lets us flexibly iterate models around the signals that matter most for personalization.

The second major driver: improving monetization efficiency. We continue to help businesses on our platforms improve their results. The ads system deployed more prediction models to deeply understand user preferences and match more relevant ads, while also optimizing infrastructure efficiency. Last quarter we launched the Meta generative recommender, which uses large models to interpret both ad creative and user preferences and predict the best ad. In Q3, Facebook Feed launched a generative retrieval module, lifting conversion rates in targeted conversion scenarios by nearly 2%.

The GEM ads ranking model expanded to more scenarios, lifting conversion rates by 1.7% for off-site conversion ads and 9.6% for app download ads. The Andromeda retrieval system was iterated to increase model complexity and throughput, lifting Facebook ad conversions by 1.4%.

AI also lowers the barrier for advertisers, helping businesses plan campaigns, generate creative, and capture opportunities. Meta AI business assistant has become a trusted tool for advertisers. Advertisers who adopt the assistant's suggestions saw conversions rise nearly 5% with cost per conversion unchanged, and their ad spend increased 9% compared with similar non-users. We will later add capabilities such as creative insights, business analytics, and intelligent task execution. Marketers can complete planning, asset creation, and optimization through conversation, with approvals and previews, all under their control.

One example: Simply Organic Maple, a family-owned maple syrup business in Vermont, used Meta AI business assistant to optimize its ad copy; after following the suggestion to test a longer, more descriptive copy, purchase conversions rose 57% and cost per purchase fell 36%; the best-performing new ad delivered a 52% higher return on ad spend, three times the original.

Generative AI ad creative tools continued to spread, with nearly 10 million advertisers using at least one of them in Q3. Adoption of the video generation feature among advertisers rose 25% quarter over quarter; upload one image and a finished video is generated in seconds.

Creator ad partnership business is growing rapidly, allowing advertisers to run ads together with creators and brands; the business grew 97% year over year, with annualized revenue surpassing $15 billion. In Q3, Creator Marketing Hub launched globally for all advertisers, bringing together creator search, asset management, and partnership ad delivery, with more than 10 million creators searchable on the platform. Meta AI business assistant will also be embedded in the hub, letting advertisers search for creators, find assets, and get end-to-end guidance through conversation.

Beyond ads, there are more monetization paths across the ecosystem. In September we launched Meta One, a unified subscription service rolling out gradually around the world for everyday users, creators, and businesses, covering more than 50 features across Instagram, Facebook, and WhatsApp, including Meta business agents and Meta AI; subscriptions and trials have now surpassed 15 million. Core experiences across the family of apps and Meta AI remain free, while subscriptions unlock advanced capabilities and more AI compute.

Beyond monetizing existing products, we are building new experiences and revenue streams around consumer AI agents. Mark described Muse's strong early performance. Usage scenarios are diverse, with the top three being shopping, personal productivity, and creative work; shopping is used by more than a quarter of Muse weekly active users.

For now, the primary goal is to expand Muse's user coverage, while also having multiple monetization paths, with subscriptions being the starting point. The opportunity for Muse among small and medium-sized businesses is equally huge. According to research, an estimated at least 500 million SMBs use Meta's family of apps to reach customers and generate revenue. Muse can help businesses acquire customers, improve operational efficiency, and facilitate user transactions. Over the long term, consumer agents and business agents can transact directly with each other, bringing multiple opportunities across paid usage, e-commerce, and enterprise services.

Beyond growth, we are pursuing a healthy unit economic model. Muse consumes a large amount of compute, and user expansion will push up total serving costs; however, through idle-time compute scheduling, model lightweighting, and inference optimization, we expect comprehensive compute efficiency to improve more than 10x over the next year. As revenue scale grows, efficiency gains give us confidence in building a highly profitable business over the long term.

Now let's look at the guidance. For Q3 2026, total revenue guidance is $73-76 billion. This guidance assumes currency will create roughly a 1% headwind to year-over-year revenue.

Expense and capex outlook: expense guidance is raised, incorporating a $9.9 billion one-time legal charge in Q3. Full-year 2026 total expense range is $175-179 billion. Excluding this legal expense, full-year operating income is expected to be higher than in 2025. By segment, Reality Labs' full-year loss narrows versus 2025, landing in the $18-19 billion range. Full-year capital expenditures (including finance lease principal) narrow to $135-145 billion, versus the prior range of $130-145 billion. Assuming no change in the tax environment, the Q4 tax rate is expected to be 15%-17%.

We continue to monitor various legal and regulatory risks. Youth-related regulatory reviews continue in multiple markets, and several youth-related court hearings in the U.S. will take place in the coming months.

Beyond 2026, next year's budget and capacity plans are still being formulated, but we will continue to invest heavily to support the core business and emerging AI projects. Existing and upcoming compute allows us to allocate resources flexibly: model training, core business growth, and new product development. AI investments are already generating incremental revenue for the advertising business by increasing engagement, improving recommendations, and enhancing monetization. Investment in MSL labs will unlock a large number of new revenue opportunities and strengthen product capabilities across the company. The AI investment payback period is long, but multiple monetization paths and internal efficiency gains support our confidence in the long-term potential.

Although we are not disclosing full-year 2026 revenue guidance, the gains AI brings to the core business, combined with model iterations and new products, are sufficient to support strong overall revenue growth in 2027. In addition, we will continue to drive improved profitability at Reality Labs; as Mark said, next year's loss will narrow to about $15 billion, with resources prioritized toward AI.

Today we are also announcing a new non-voting common stock proposal: distributed to all existing shareholders as a stock dividend, with each share of existing stock entitled to one new non-voting share, equivalent to a 2-for-1 stock split, expected to begin trading in late January 2027.

In summary, Q3 saw strong momentum in core advertising and product execution, while we are steadily advancing on the path to personal superintelligence for everyone. Krista, let's open it up for questions.

Operator: Thank you. We will now open the line for questions. The first question goes to Brian Nowak of Morgan Stanley.

Brian Nowak (Morgan Stanley): Thanks. Congratulations on Muse's success. From a user behavior and product perspective, what do you see as the bottlenecks constraining Muse's user growth slope that need to be addressed first? For example, capabilities, model efficiency, compute capacity. Second question, Mark, could you describe the vision for Meta's enterprise platform and its core value for enterprise customers?

Mark Zuckerberg: I'll take both questions first, and Susan can add. Currently 10 million users in the U.S. and Canada have completed Muse setup, but the biggest constraint comes from the enormous compute consumption, and we have not yet launched globally. On one hand, we need to adapt the product for each region and connect with local third-party connectors; on the other hand, we need to first solve operating efficiency to ensure commercial viability.

And precisely because we have only launched in a few markets, this 10 million result is even more impressive. If we opened globally, the user scale would be far higher than now. But the efficiency optimization work must be done. Compute efficiency is the top priority; we need enough compute, and only after achieving efficiency targets will we be able to serve more than 1 billion users. This is a major constraint for next year.

Commercialization buildout must advance in sync with compute efficiency improvements, and I remain optimistic about that. The product's own capabilities must also continue to iterate, achieved through new models. The free version uses Muse Spark by default, while the paid version can experience Muse Wonder, which is much more capable. We are developing an efficient version of Muse Wonder to roll out broadly later.

After two to three generations of iteration, model intelligence will improve significantly while costs decline, unlocking enormous user value.

Turning to the enterprise business. The two are connected: near-term enterprise products are mainly model APIs and Muse Code. Business agents for enterprises are already scaling, and all three are growing well. Over the long term, we want to build a complete Muse business agent network.

Business agents can handle sales and customer service; today a large number of SMBs already use Muse to run their operations. We want to build dedicated business agents that automate many tedious, non-core business processes for SMBs. Launching a version of Muse adapted to enterprises' internal environments is equally significant.

Going further, building the Muse network: no longer single-person use, personal agents can collaborate with one another. When shopping, the user's agent connects directly to the merchant's agent instead of visiting the merchant's webpage. Once this system is proven out, it will unlock enormous value at scale. The enterprise scenario is an important part, and it will likewise enhance the value of consumer products; I'm excited about both sides.

Susan Li: Let me add on the enterprise business. AI is profoundly reshaping the enterprise market. Traditional enterprises do not have a monopoly on this track; many of today's leading players did not exist a few years ago, and they are investing heavily in model R&D, as is Meta, and our models are now competitive for both SMBs and large enterprises. 500 million businesses worldwide already use our platform. AI will completely change how enterprises do development, productivity, and marketing and customer service.

We have multiple differentiated assets: models, compute infrastructure, global distribution, and long-term merchant relationships, and we want to build diversified, durable revenue streams spanning subscriptions, usage-based API pricing, messaging services, and more. This opportunity will take years to cultivate, and we will scale investment gradually based on real customer needs and product-market fit.

Conference operator: Next question, Eric Sheridan of Goldman Sachs.

Eric Sheridan (Goldman Sachs): Thanks for taking my question. First question, there is friction in agentic commerce; some aggregator platforms and suppliers are unwilling to have AI agents call and access their services. How does Meta view this kind of tension? How do you reduce friction, balance existing consumer services, and the industry changes agents will bring in the medium and long term? Second question, with the new share class announced today, how do you view future capital sources and capital structure optimization to ensure you can seize investment opportunities in the AI era?

Mark Zuckerberg: Let me answer the first question.

Susan Li: It's about the friction points in agentic commerce.

Mark Zuckerberg: Got it. Overall, the vast majority of businesses are enthusiastic about this new model; they see that it can not only improve the existing customer experience and expand sales, but also bring in new customers. Different players in the industry will adopt at different paces, and our ecosystem strategy will adjust accordingly. A large number of merchants have reached out proactively, hoping to integrate with Muse, which exceeded my expectations. Of course, there will also be businesses unwilling to try early on, and we will negotiate reasonable partnership terms. In the long run, this system is valuable to all parties, and overall progress is positive.

Susan Li: To add. Agents are part of the future reality, and everyone needs to adapt to the era of agentic commerce. Many merchants are actively seeking partnerships, but some businesses want to protect their own customer relationships. We believe that in the future agents will no longer browse webpages like real people; it will move toward authorized, structured interaction: certified APIs, official agent partnerships, and consumer agents communicating directly with business agents. We are working with partners such as Sierra, Walmart, Stripe, and Shopify to launch an open standard, the "Personal Agent Protocol," which defines rules for secure interaction between consumer agents and merchant agents, balancing merchant and user permissions.

Back to your second question, the original intent behind the new share class: to create a publicly traded non-voting share class for fundraising, employee compensation, acquisitions, and other corporate purposes, while protecting the voting rights of existing Class A and B shareholders. This both maintains the founder-led structure and broadens our financing tools, giving us ample flexibility to seize AI-driven opportunities across the entire business. There are no specific financing plans at present, but building a complete capital framework in advance makes it easier to draw on diverse funding in the future.

Conference operator: Next question, Doug Anmuth of JPMorgan.

Doug Anmuth (JPMorgan): Great, thank you. Following up on authorized agent transactions. When a large number of businesses connect to agent-to-agent interaction, what changes will that unlock over the long term? Second, on expenses and capex, although the 2027 plan is not yet finalized, looking back at historical statements: 2026-2027 full-scale compute expansion, 2028 greater freedom; a year ago you mentioned that 2026 expense growth would be significantly greater than 2025, while also pursuing annual operating profit growth. Is there any clue to help understand 2027?

Susan Li: The agent ecosystem is hard to predict precisely, and the pace of development will be uneven. But using Muse today, you can already feel that if you could connect directly to business-side agents, the experience would improve by a large margin. For example, when Muse calls a merchant on the user's behalf, if the other side is answered by the merchant's agent, the whole process would be much smoother.

One of Muse's defining traits is proactivity: after receiving a task, when it encounters obstacles it autonomously finds a path to complete the goal. The proactive capabilities of consumer and business agents will generate more transactions and business activity than today. It is no longer the case that agents act only when users explicitly state a need; agents can continuously monitor user needs and periodically push results; the same goes for business-side agents. Over the long term, this will form an entirely new proactive interaction paradigm.

Second, on 2027. Next year's budget and capacity planning are still in early evaluation, and a large number of investment plans with different return cycles are being assessed. We want to execute as many high-quality projects as possible within the framework of our budget and financial targets. There are many variables, and we will not disclose quantitative guidance for next year until the Q4 earnings call.

Conference operator: Next question, Justin Post of Bank of America.

Justin Post (Bank of America): One question for Mark and one for Susan. Mark, the new large model has great potential internally as well as for external products. The ads system has already been rebuilt on AI; after deploying the new generation of models, how much more room is there for improvement internally? Susan, you mentioned strong revenue next year; does that include only advertising, or does it already factor in some of the new AI revenue?

Susan Li: I'll answer both. The advertising business has benefited for years from both traditional machine learning and the new generation of large models.

There are three main threads: First, the traditional machine learning iteration pipeline remains ample. Even if we don't disclose guidance for 2027, product iteration gives us confidence in growth of our core business, and we have ample reserves in ranking and recommendation systems.

Second, the migration to large-model infrastructure: by using large models to deeply understand user interests, organic content and the recommendation engine can continue to unlock value.

Third, generative AI advertising tools are already delivering value to advertisers, and we will continue to iterate the Advantage+ suite to rapidly generate creative, image-to-video, and audio, enabling highly personalized ads. All three of these still have substantial room for improvement.

On the second question, next year's growth. We will continue to invest in core advertising and user engagement initiatives; at the same time, several new AI businesses are being built and will begin contributing incremental revenue in 2027, but the larger-scale release will come later. Revenue sources include AI subscriptions, transaction fees, advertising, and enterprise services. We are still in the early stage of refining products and validating product-market fit. But we have the experience of serving billions of users, and over the long term AI will become an important source of revenue diversification.

Conference operator: Next question, Mark Shmulik of Bernstein.

Mark Shmulik (Bernstein): Thank you. The feeling from Connect is that the company has entered a new era, with Muse at the center of all businesses. Is that really how it is positioned internally? If so, what changes have been made to business assessment and goals? Susan, the enterprise product was released before the SMB version; talk about what needs to be done to bring Muse to small and medium businesses.

Mark Zuckerberg: More than a year ago I put forward personal superintelligence as the company's north star. It continues Meta's core over more than twenty years: empowering individuals and giving people tools for expression and connection. Extending into the AI era, the most important undertaking of the next generation is putting superintelligence into the hands of ordinary people.

To that end, we restarted the MSL lab, with the goal of building a full end-to-end in-house technology stack for personal agents, including models and agent products. The smart glasses hardware vision follows the same logic: building the next-generation hardware platform and delivering a complete end-to-end experience.

These are all long-term bets. Meta's mission is to make long-term technology bets and refine the best products for billions of people worldwide. As long as we execute well, commercial results will follow, and I'm proud of the team's progress.

Susan Li: The SMB version of Muse launched at the end of September. It is not a standalone app but a new set of capabilities and connectors added within the existing Muse. It helps small and medium business operators run their businesses and acquire customers; with one click it connects to Instagram analytics, Facebook Pages, and Meta ad accounts, and also integrates third-party tools such as Shopify, QuickBooks, and Canva. The agent understands the merchant's products, brand tone, and customer needs; it can automatically analyze ad and content performance, write marketing copy, reply to customer messages, and identify business risks.

We continue to collect feedback from small and medium businesses to identify the most frequent needs: ad optimization, message organization, expanding sales, syncing inventory and product updates to websites, and staffing. We closely track usage and iterate on features. There are 500 million small and medium businesses worldwide, and much of the technology we have built for consumers can be reused in SMB scenarios. We have mature channels for reaching small and medium businesses, and a strong foundation platform. AI also opens up entirely new enterprise market opportunities for us.

Conference operator: Next question, Ross Sandler of Barclays.

Susan Li: There are no specific figures we can disclose publicly. As Mark said, compute costs are high. Compute efficiency is the core lever for achieving scale and meeting business goals. Near-term measures: using idle compute during off-peak times to run some tasks, multi-dimensional inference optimization, improving GPU utilization, along with various product optimizations. Longer term, after large-model iterations, completing the same tasks will consume fewer tokens; at the same time, the AI models themselves will optimize the underlying runtime software. There are still many projects to land on the efficiency roadmap, and technological evolution will continue to bring costs down.

Conference operator: Next question, Mark Mahaney of Evercore ISI.

Mark Mahaney (Evercore ISI): Meta's historical approach to new products: first build scale and engagement, then layer on monetization. Muse's compute costs are far higher than past products; does that old playbook still apply, or do you need to move faster on monetization? Second, does the e-commerce data on 25% of users and average order values above $200 include travel spending, or is it retail only?

Mark Zuckerberg: Given the scale of costs and the market opportunity, Muse will monetize earlier than our traditional approach. Second, that statistic includes travel; travel itself is also a high-frequency use case for Muse, with users using it for all kinds of trip planning.

Conference operator: That concludes the Q&A session. This concludes the entire Meta earnings call. Thank you for participating; you may now disconnect.

(Source: Cailian Press) Mid-article embedded ad