Meta has just released its latest earnings report, with revenue beating expectations, but free cash flow turned negative. What the earnings report doesn't reflect is the data on Meta's personal agent Muse.
From the earnings call, it's clear that virtually all questions were about Muse. As the number one downloaded free app in the U.S. and the latest AI application surpassing ChatGPT, Muse is indeed Meta's most closely watched product in the market right now. Let's take a look at Zuckerberg's latest views on Muse and his plans for the future.
What is the most critical constraint on Muse's subsequent user growth—capability, model efficiency, or compute? Also, what do you think of the Meta enterprise platform, and what is its core value to enterprise customers?
Mark Zuckerberg: Let me start with Muse. I mentioned that more than 10 million people in the US and Canada have already created a Muse, but one of the biggest constraints right now is that this service is very compute-intensive, so we haven't launched globally yet. Part of the reason is that we want to optimize for each market, and for that we have some workaround connector solutions; another part is that we want to make sure we do this efficiently. There are trade-offs here.
The good thing is that 10 million people signed up within six weeks, and that number comes only from a few markets in the US and Canada, which makes it even more remarkable. If we really roll it out to more markets, that number should be significantly larger than it is today. Conversely, we must truly improve efficiency before globalization makes sense, and that's a big chunk of work for us.
Compute efficiency is also critical here: we need to build enough compute to serve the more than 1 billion users we expect—provided we hit certain compute efficiency targets. We have a lot of work to do there as well, and it will be one of the constraints next year. At the same time, we need to build the business around it—improving compute efficiency while building a business model—so that this works commercially, and I'm optimistic about that.
Then there's continuously improving product value, which mainly comes from constantly developing new models. So I mentioned that Muse launched on Muse Spark, which is also the default model on the free tier; Muse Wonder is already live and available today on the paid tier, and we know it performs much better, so we're working on a more efficient version to bring it to more people.
You can imagine that after two or three more generations of models, with significantly improved intelligence and capability that can help people solve more problems, while being delivered at a much lower price, that would unlock much greater value.
On the enterprise side, it's connected to the consumer side. The main enterprise products in the near term will be the model API and Muse Code. We've been working on business agents for a while, and all three are growing fast.
Longer term, we want to form a Muse agent business network. One is the business agent, which I talked about in previous quarters, helping merchants with sales and customer service. We see many people already using Muse today to help run their small businesses, especially small merchants. The opportunity here is to build an agent that automates the tedious, time-consuming parts of starting a small business that people don't really want to do themselves—I think that's a big deal. Also, building a version of Muse that can be brought into different enterprise environments to work will be a big deal too.
Further ahead is what I call the Muse network: making Muse no longer a "single-player game"—not just having your own Muse, but having your Muse able to collaborate with the Muses of friends and family; when you want to buy something, you don't open that merchant's website, but instead talk directly to their agent. Connecting all of this will unlock enormous value at scale. Enterprise use will be a big part of that, and it will also bring significant utility gains to consumers, so I'm excited about both sides.
Susan Li: Let me add a bit to Mark's view on the enterprise side. AI is creating massive disruption in the enterprise market today. From what we see, we don't believe traditional enterprise software companies have locked up this opportunity; in fact, some of the leading companies in this space today didn't exist a few years ago. Those companies have invested heavily in model development, and so have we, and our models are becoming increasingly competitive, able to serve all our users, including small businesses and large enterprises. Today, 500 million merchants are already using our products. We do believe AI will reshape how enterprises operate, and businesses of all sizes will use AI agents in software development, productivity, sales, marketing, customer service, and other areas.
We have differentiated assets in this market: models, infrastructure, global distribution capability, and long-term relationships with merchants. We hope this will become a new, diversified recurring revenue stream that can support multiple business models, including subscriptions, usage-based APIs and messaging, and other revenue lines. We expect this to be a multi-year opportunity, and we will scale our investment based on validated customer demand and product-market fit.
As agent commerce advances, there will be friction with some aggregator platforms or merchants unwilling to be integrated by agents. How do you plan to solve that? Also, the new class of stock announced with the earnings report—does it give you any new thoughts on funding sources and capital structure?
Mark Zuckerberg: What we're seeing so far is that the vast majority of merchants are very enthusiastic about enabling this, because they recognize it's a new opportunity: it can both improve sales and experience quality with existing customers and potentially bring in new customers. Across the commerce ecosystem, adoption pace will vary, and our ecosystem strategy in this area will evolve as things progress. But overall, we've been pleasantly surprised—a large number of merchants not only want to figure out how to let Muse better use their services, but are actively coming to us for integration. That's a great signal. Not everyone wants to be an early adopter, and that's fine; we'll work with partners on terms they're comfortable with. Long term, doing this well will be valuable for everyone, and our trajectory is good so far.
Susan Li: I think agents will be an important part of the reality of the future, and everyone will have to adapt to what the new era of agent commerce means. We obviously have many merchants who are very willing to work with us, and some merchants have considerations around protecting their own customer relationships, which is understandable. We believe the market will ultimately not stay at "agents operating every digital experience like humans do," and in the long term will move toward more permissioned, structured interactions, including authenticated APIs, approved agent partnerships, or consumers' personal agents communicating directly with merchants' own agents. We are helping to establish the standards needed to support this model.
Earlier this month, we previewed the Personal Agent Protocol, an open standard spanning Muse and Meta Business Agents, which we are working on together with industry partners such as Sierra, Walmart, Stripe, and Shopify. This protocol is intended to define how consumers' personal agents can interact securely and reliably with merchants' agents, while preserving appropriate control for both merchants and consumers. Again, adoption will likely be uneven and take time, and in the near term we expect many experiences will still be AI-assisted rather than fully autonomous. As trusted integrations are gradually established, agents prove they can bring incremental value and incremental demand, and can operate safely, we expect more and more merchants will be willing to participate.
Your second question is about the announcement of a new class of stock. We want to clarify that the purpose of establishing this new class is to create a publicly tradable non-voting stock that can be used for financing, employee compensation, M&A, and other general corporate purposes, while preserving the voting rights of existing Class A and Class B shareholders. We believe this class will maintain our founder-led structure, expand our financing tools, and give us the flexibility to invest in the unprecedented momentum we see across all aspects of our business brought by AI. We currently have no specific financing plans to announce, but we believe it is prudent to first put the appropriate capital structure framework in place to ensure we can access diverse sources of funding.
What will truly be unlocked as we move toward more permissioned, agent-to-agent transactions? Also, what clues can help us judge 2027 expenses and capital expenditures?
Susan Li: Let me start with how the agent ecosystem will evolve. I wish I could predict it more accurately, but this will take some time, and I likewise think it will be uneven. But what is certain is that I felt it very clearly when I used Muse today: if Muse could interact more seamlessly with a merchant's agent on the other side, many parts of the experience would be better. For example, if Muse needs to call a merchant on my behalf, if that call is answered by a business agent representing the merchant, it would be easier for that merchant.
One more thing I think is very unique about Muse and one of the reasons the product experience is so good is the proactiveness it shows for me. I can give it a task and it goes off and executes it, and even if it hits obstacles along the way or temporarily can't find the right thing, it will find another path to help me look.
This kind of proactiveness shown by consumer agents or merchant agents will cause more transactions and more activity than happens today. Today, actions mainly happen when latent demand is explicitly expressed; but with agents, Muse can continuously help me look for things, regularly update me on what I'm looking for, and push ten new books to me every month. Agents should do similar things for merchants, and I think this will bring an opportunity for a new, more proactive agent-to-agent interaction framework.
The second question is about 2027. We are still at a very early stage in budgeting and capacity planning for next year. It's fair to say we are evaluating many investment proposals—which have good return characteristics across different time horizons—while thinking about how to fund as many of these proposals as possible within the framework of our overall budget and financial targets. But the situation is very dynamic and it's still early, so we don't have anything more specific or quantitative to share before the Q4 earnings call—which, as usual, is when we give guidance for the next year.
How much room is there to improve by applying new models to the ads system? Will next year's strong growth come from ads, or does it already include some incremental revenue brought by AI?
Susan Li: Our ads business has benefited over the years from two generations of investment: one is continued investment in what we used to call traditional AI machine learning models, and the other is the more recent generation of models. There are three themes worth mentioning here.
First, there is the more traditional performance improvement roadmap we already have in the pipeline, closer to "routine machine learning work." This pipeline remains solid and healthy. Looking at 2027, we likewise don't provide an outlook, but our investments and product improvements on both the engagement and ads sides give us confidence that there is still considerable room for growth in the core business throughout 2027. This also answers your second question by the way: our ranking and recommendation pipeline is healthy.
Second, as we migrate to more LLM-oriented AI infrastructure, we have some new research bets aimed at reasoning more deeply about people's preferences and the things they are interested in, including both the organic content side and the recommendation side. We also expect to gain improvements in our ranking and recommendation engines from this research.
Third, we expect generative AI tools to continue expanding the capabilities available to advertisers. These tools are already very valuable to advertisers, and we will continue adding features mainly through the Advantage+ suite: enabling advertisers to generate creative faster and more smoothly, or turn an image into video and generate voice, all aimed at making ads increasingly personalized for the people who see them. So across these three things, we feel there is considerable room for growth.
Next year's revenue growth, as I said, we feel good about our budget investments in core ads and engagement for next year. At the same time, we're building several new AI businesses that we expect to start contributing incremental revenue in 2027, but more importantly, they can significantly scale up in the long term. These include monetizing our AI products through subscriptions, transactions, and ads, as well as the enterprise products we mentioned earlier. So I'd say many of these opportunities are still early, and our near-term focus is really on building solid products, achieving product-market fit, and scaling product-market fit; however, our track record of reaching billions of people and scaling businesses gives us confidence that AI will become an important source of revenue diversification in the long term.
After Connect, the company seems to have entered a new phase, with everything being restructured around Muse—is that judgment correct? Also, what else needs to be done to scale up products serving small and medium-sized businesses?
Mark Zuckerberg: More than a year ago, I wrote that letter making "personal superintelligence" our North Star. This is actually in line with what the company has been doing for over twenty years: empowering individuals, giving people tools for expression, and letting people connect in the ways they want. Extending this to today's AI work, we believe the most important thing we can do for the next generation of technology is to put the power of this technology in everyone's hands—that is the vision of personal superintelligence.
We've been working on this for a while, including restarting around MSL and building up the lab. To a large extent, this stems from the idea that we need to specifically make models suitable for being personal agents, and only by owning, building, and designing end-to-end technology ourselves can we create the best experience. This includes both models and agents. This direction has been brewing for some time. The vision for glasses is the same: we want to build the next-generation hardware platform to deliver the best end-to-end experience to people. These are now starting to come together and will become more evident in the coming years. These are long-term bets.
I think this is what we exist for: making long-term technology bets and designing the best end-to-end experiences for billions of people around the world. As long as we keep doing this, the business will take care of itself in the long term, provided we continue to execute well—and I'm proud of how the team has performed in this regard.
Susan Li: The version of Muse for small businesses launched in late September. You can think of it as an extension of Muse rather than a standalone product: it's a set of new skills and connectors within the existing Muse app, aimed at helping small businesses with operations and finding new customers. With a few clicks, they can connect their Instagram account insights, Facebook Page, or Meta ad account, plus dozens of third-party connectors that businesses already use, such as Shopify, QuickBooks, Canva, and more. It can already understand what the business sells, what the brand tone is, and what customers are asking.
Muse can take actions, including analyzing ad and content performance, drafting marketing campaigns and customer replies, and flagging issues. We are actively listening to feedback from small business users to understand their primary use cases; we see them using it for ad optimization, organizing direct messages, expanding sales outreach, keeping websites updated based on inventory and new products, and balancing employee workloads.
We are closely monitoring how businesses use Muse to ensure that the new features we roll out in Muse are thoughtful, most useful for small and medium-sized businesses, and most productive. I also want to say that we have unique reach—I mentioned at the outset 500 million users who self-identify as small businesses; and much of our investment in consumer products will spill over to help these businesses. We have multiple great channels to present these opportunities to small businesses globally. This is an area where we already have an extremely strong platform, and we are very excited about it. We also believe that the shift to AI overall opens up new opportunities for us in the enterprise market.
What does Muse's per-capita cost structure look like today, and how will this cost curve come down going forward? What are you doing to bring it down?
Susan Li: Regarding Muse's cost structure today, we don't have specific numbers to share, but it is indeed, as Mark said, very compute-intensive. Getting compute efficiency right is one of the key paths for us to both scale Muse and make it the business we want it to be.
I think there are several layers to efficiency. The nearer-term parts include: using off-peak or idle compute to run some tasks, using smaller models for certain tasks; and there is a lot of inference efficiency work to do, essentially all aimed at higher GPU utilization, as well as other efficiency work we can do on the product side.
Longer term, as LLM scale continues to expand, we expect to accomplish more tasks with fewer tokens, and to rely on AI models themselves to help us improve the efficiency of the software we use to run models. I think there is a lot of work on our efficiency roadmap that can help us bend this curve down; and we also expect the broader trend of AI technology progress to help us achieve greater savings on this front.
Will Muse's monetization happen faster than Meta's usual "build users first, then monetize" approach because costs are higher? Also, does the e-commerce data you mentioned—25% of users, average order value above $200—include travel?
Mark Zuckerberg: My guess is that we will start monetization earlier than the usual playbook, for two reasons. First, the magnitude of these numbers we're talking about—whether costs or opportunity—makes it important to make progress on monetization, otherwise we can't scale. Second, e-commerce accounts for a large part of what people do with Muse.
To answer your question, that data includes travel. We might view travel as a distinct category, but travel is actually one of Muse's primary use cases: people use it to plan trips and plan various things they want to buy, things they are less likely to do within our family of apps.
I think this matters for delivering the consumer value users want from us. By doing so, we can also establish a model: taking a certain percentage of the value we create for people. Getting that math to work is, I think, very critical so that we can continue advancing the compute buildout we want and ultimately bring this service to everyone.
Beyond what Susan just answered, I'd add one more point: one trend we're seeing is that a year from now, we should be able to deliver the same level of intelligence as today's Muse Wonder or any current frontier model with a much smaller model at very low cost. If that constitutes some kind of baseline that people consider sufficient for Muse to reach product-market fit, then we'll keep moving forward along that efficiency curve through continued model development.
What's less certain: I don't think most people a year from now will still be willing to use today's level of intelligence, because models will have advanced by then. So we need to do all of these things—the revenue side, scaling across various dimensions, plus model development and efficiency gains—to both have the funding and support the compute buildout, and ultimately reach billions of people; and that, we hope to do as quickly as possible.
Part of the reason we don't charge everyone from the start is that we want to build up the Muse network. We believe Muse becomes more valuable to everyone when many people and many merchants are using it, because they can collaborate and do things together. I think that's a very important state, and one we're very much looking forward to. Again, I don't think there are many companies in the world positioned to pull this off the way we are. We're well positioned, and we intend to push very aggressively.
Welcome to join the official BlockBeats community:
Telegram Subscription Group: https://t.me/theblockbeats
Telegram Discussion Group: https://t.me/BlockBeats_App
Official Twitter Account: https://twitter.com/BlockBeatsAsia
