Meta Muse AI: 10M Users and the Business Model
Meta says its Muse AI app has passed 10 million users. See what shopping, productivity, monetization and compute costs reveal about its business potential.
Meta Muse Has 10 Million Users. Can Its AI Business Scale?
Meta says its AI app Muse has passed 10 million users. But the more interesting part isn't the install count - it's what people actually do with it. CEO Mark Zuckerberg said roughly one in four people who set up Muse use it every day. Shopping is a standout use case: more than a quarter of engaged users who shop through Muse have an average purchase size over $200. Meta also says users connect Muse to personal tools like email and calendars.
Taken together, those details suggest Muse is meant to be more than a chatbot. Meta's goal is for the agent to help people make decisions, take actions, and eventually interact with businesses and other AI agents. That's a bigger opportunity than selling simple access to a model, but it also raises the stakes: an agent that can shop or touch a calendar has to earn trust, produce reliable outcomes, and keep its compute costs in check.
This article breaks down what Meta's numbers do - and don't - tell us, how its revenue model could work, and why compute efficiency and privacy may decide whether Muse turns into a lasting business or just another expensive feature.
What Meta's 10 million user figure tells us
Meta said on its Q3 earnings call that Muse has more than 10 million users, confirming an estimate released earlier by Sensor Tower. The number is a decent signal for distribution and early adoption. It doesn't, by itself, prove the product can become a sustainable business.
For an AI assistant, the more revealing question is what happens after installation. Zuckerberg said about 25% of people who set up Muse use it daily. That's a meaningful usage signal - daily behavior is harder to earn than downloads. Still, the figure needs context before you can compare it to other products. Meta didn't spell out how it defines a "daily user," what the measurement window was, or how retention looks by cohort in the remarks reported by TechCrunch.
Those missing details matter. Someone who opens the app once a day for a quick question is using Muse differently from someone who delegates multi-step work. Both are "active," but they likely drive very different outcomes, costs, and willingness to pay.
Adoption is not the same as product-market fit
Zuckerberg said Muse has reached product-market fit and that Meta plans to scale the agent, increase revenue, and reduce compute costs. That's Meta's claim, not something the reported figures alone prove. The user and activity numbers point to traction, but they don't yet show whether usage sticks, whether the service is profitable, or how resilient it is against competitors.
A practical way to evaluate an AI product is to separate four stages:
- Reach: Can the company get the product in front of people who might want it?
- Activation: Do users complete a useful first task?
- Retention: Do they come back when there's a real need?
- Value capture: Does repeated use generate revenue that covers the cost to serve?
Meta has an advantage on reach because it already runs large consumer platforms. That can reduce friction for trying Muse. But reach doesn't solve the later stages. An assistant has to be good enough to earn repeat use, and the business model has to cover inference, infrastructure, support, security, and ongoing product work.
You can see the same pattern across the broader AI market: a big user count pulls attention, but the real test is whether the product becomes part of a recurring workflow. For Muse, shopping and productivity are encouraging because they lead to actions with tangible results. The next question is whether users keep delegating those tasks over time.
Shopping and productivity are different tests of trust
Shopping is one of the most commercially interesting uses Meta disclosed. Zuckerberg said that more than a quarter of Muse's engaged users who were shopping through the agent have an average purchase size above $200. He also pointed to Plaid, whose connector Muse uses for agentic shopping, saying it hadn't seen similar activity on another service.
That's a useful data point, but it's worth reading it carefully. The "average purchase size" applies to a subset of engaged users shopping through the agent, not necessarily to all Muse users. And a higher order value doesn't automatically tell you how often people buy, whether the purchase was actually completed through Muse, or how much of the revenue Meta keeps. In other words, it suggests commercial intent - but it doesn't yet add up to a settled unit economics story.
Why an AI agent can change the shopping journey
In a typical setup, search or social ads help a user discover a product. An agent can take on more of the work: interpreting the request, comparing options, and helping complete the transaction. That moves the assistant from influencing a purchase to actively participating in the purchase journey.
For businesses, that could be valuable if the agent consistently matches users with relevant products. It could also change competitive dynamics. If customers talk to an agent instead of bouncing between retailer sites, businesses may need to make their products, pricing, and policies easier for the agent to interpret - not just persuasive to a human browser.
But there's a trade-off. Once the assistant is responsible for more than a recommendation, the bar rises. It needs accurate details, must handle constraints, and should clearly show what it's doing (and give the user a way to review or stop a transaction). If a recommendation is wrong - or if a purchase happens without clear consent - the trust issue doesn't stay with the seller. It lands on the agent, too.
Productivity raises a higher privacy bar
Meta also said users are connecting Muse to personal productivity tools like email and calendars. Those integrations can make the assistant more useful because it can work with a person's real schedule and information, instead of relying only on what they type into a prompt.
They also make mistakes more consequential. A poor recommendation might cost time or money. An assistant with access to a calendar or inbox could expose sensitive information, misunderstand an instruction, or take action at the wrong time. The more powerful the integration, the more carefully the system needs to define permissions, data handling, and user control.
Meta pointed to its Muse Secure VM as a privacy protection step, and Zuckerberg said a forthcoming Muse Confidential VM would prevent even Meta from seeing users' content. Those are strong design claims, but the real question is whether the implementation holds up and whether outsiders can verify it. Users also need clarity on what data is accessed, what's stored, which actions require confirmation, and how they can revoke access.
In practice, trust isn't one feature you flip on. It's a combination of permissions that make sense, predictable behavior, good failure handling, and a clear way to undo or review what happened. If an assistant asks for broad access before it proves value, adoption friction is likely - even if the underlying model is impressive.
The business model: commerce, subscriptions and costs
Meta says it wants to keep Muse free for most users and monetize through commerce by taking a percentage of transactions that happen via the platform. It also expects power users to pay for access to more advanced models and for more tokens. That gives Meta two possible revenue streams: transaction-linked income and paid access for higher usage or capability.
The mix is plausible because different users value different things. Some people may be fine with a free assistant supported by commerce. Others - especially heavy or professional users - may pay for more capacity. Still, each revenue path depends on different fundamentals. Commerce depends on users actually completing purchases and merchants agreeing to participate. Subscriptions depend on a recurring benefit users can't easily recreate elsewhere.
Why revenue per user is an ambitious benchmark
Zuckerberg said Meta expects Muse, eventually, to generate more revenue per user than users of its family of apps - Facebook, Instagram, WhatsApp, and Messenger - generate today. He described current annual revenue per user for that group as several hundred dollars.
It's an ambitious comparison. It isn't a forecast with a timeline or published assumptions. Meta's apps already have a mature advertising business and massive reach. For Muse to match that kind of revenue, it would need to create real per-user value through commerce, paid features, or another stream - while also paying the cost of running an AI service.
This comparison also highlights a strategic difference. Advertising monetizes attention and audience scale. An agent could monetize completed tasks or transactions instead. That may feel more direct per interaction, but it also comes with heavier operational responsibilities: payment plumbing, transaction disputes, merchant quality, customer support, and transparency around commercial recommendations.
The key isn't total shopping volume. It's contribution after costs and obligations. A workable model needs to account for:
- What Meta actually retains from transactions or subscriptions.
- Compute costs for both successful and unsuccessful interactions.
- Payment, support, security, and compliance costs.
- User return and conversion rates.
- Refunds, cancellations, and overall experience quality.
A higher average purchase size helps only if enough purchases happen and Meta retains a reasonable share of the transaction. Subscriptions make sense only if the paid tier's value exceeds both the cost to serve and the cost of acquiring and keeping that user.
Compute is a product constraint, not just an infrastructure issue
Zuckerberg said Muse is currently very compute-intensive to run. Meta expects to improve cost efficiency by more than 10x over the next year and said the service could become profitable once revenue lines scale. The company also said it's building capacity to serve more than a billion people in the coming few years.
These are big goals, not results. The reason the gap matters is simple: AI can eat resources even when the user experience looks straightforward. An agent may need to interpret a request, search or retrieve information, call tools, check results, and then continue the conversation. More capable behavior often means more compute per task.
Meta says it can improve the economics in multiple ways: more efficient models, better routing between models, caching repeated work, shorter or more focused interactions, and infrastructure tuned for the workload. Each approach involves trade-offs. A smaller model might be cheaper but less reliable on harder tasks. More confirmation steps may reduce risk but add latency. Aggressive cost cutting can degrade quality if it ignores which tasks truly need extra capability.
The system-level goal shouldn't be only "spend less compute everywhere." It should be: spend compute where it improves the user outcome and avoid spending on low-value work. That requires measuring performance at the task level - completion rate, latency, error rate, cost per successful task, and repeat usage. Without those metrics, a "10x" efficiency target is hard to evaluate from the outside.
The Muse network and Meta's hardware ambitions
Zuckerberg also talked about a future "Muse network" that could move the assistant beyond a single-user experience. In his framing, a person's Muse could work alongside the Muses of friends or family. He also suggested shopping might involve chatting with a business's agent instead of navigating directly to its website.
This points toward a more agent-driven internet, where software agents exchange requests and information for people and organizations. If it works, it could reduce friction in tasks like coordinating plans or asking a retailer about products. It could also open up new discovery and transaction channels for businesses.
But interoperability isn't a minor detail. Agents need a shared way to identify themselves, express permissions, interpret requests, and confirm outcomes. A business agent has to be accurate about inventory, pricing, and terms. A personal agent shouldn't disclose more information than the user intended. And when two systems disagree - or fail - users need to know who's responsible and how to resolve the issue.
The network effect has to be earned
Calling something a "network" doesn't automatically produce network effects. For the Muse network to be genuinely useful, enough people and businesses have to participate - and the interactions have to be better than the alternatives. If users still have to verify answers manually or repeat the same information across services, the network adds an extra layer without removing much work.
There's also a governance question. If Meta controls the assistant and the user relationship and the discovery channel, businesses may worry about how visibility and commercial recommendations are determined. Users will want to know whether suggestions reflect genuine fit, paid placement, or both. Clear disclosure and consistent rules matter here if Muse is going to retain trust.
Muse also fits into Meta's smart glasses strategy. Zuckerberg said the assistant will play a central role in the company's hardware plans. Reality Labs, which includes smart glasses, virtual reality, and metaverse work, has continued to report large losses. Meta is reported to expect those losses to fall from $19.1 billion in 2025 to $15 billion this year.
The logic is understandable: an assistant that can interpret context and respond hands-free could make wearables more useful. Still, hardware and AI together make the execution challenge bigger. The product has to work reliably in real settings, protect sensitive data, and deliver enough daily value to justify both the device and the service. A great demo won't be enough to prove a durable business.
For Meta, the upside is leverage across software, commerce, and devices. The risk is that each layer adds costs and new failure points. The real test isn't how many product surfaces can include Muse. It's whether the combined experience makes user tasks meaningfully easier without worsening privacy, reliability, or economics.
Conclusion
Meta's early numbers suggest Muse is starting to do more than replace another chatbot. Some people are using it to shop, and others are connecting it to personal productivity tools. That's commercially interesting, but it also raises the bar for reliability, privacy, and unit economics.
Meta has distribution, potential commerce revenue, and a hardware strategy built around the assistant. What it still has to prove is that users come back, that transactions create revenue Meta can keep, and that inference costs fall without degrading quality. Those are the real operating questions - not branding ones.
If you're evaluating Muse as a business signal, focus on three things next: cohort retention, revenue actually captured per active user, and cost per successful task. They'll tell you more than a headline user count. For the original earnings-call reporting and figures, read TechCrunch's coverage of Meta Muse.