The stock market is a machine for pricing things before anyone knows what they cost, and it has been running hot on Meta. Since the launch of its AI agent, Muse, the company has added nearly $200 billion in market value, on the assumption that the product will scale. As a Forbes analysis this week points out, investors have priced in automation before finding out how much of the service is actually automated, or what it costs Meta each time the thing successfully completes a task. Those are, arguably, the two questions.
The exuberance is not mysterious. Muse launched on Sept. 8 and shot to the top of the U.S. app charts. It sends emails, arranges travel, shops online, fills out forms and calls businesses on a user’s behalf. More than 2.5 million downloads arrived in roughly two weeks, and Meta shares jumped 11% in a single session. After years of watching Meta spend on AI with the returns dissolving invisibly into ad targeting and engagement, shareholders finally had a consumer product they could point at.
The valuation math is easy, which is exactly why everyone did it. Meta has around 3.6 billion daily users across Facebook, Instagram, WhatsApp and Messenger. Subscriptions are believed to be under consideration in the $20- to $100-a-month range. Multiply even a small conversion rate by 3.6 billion and you get billions in recurring revenue, and the market got there first. What the market cannot yet calculate is the other side of the ledger: what each paying customer costs to serve.
Context helps. Second-quarter revenue rose 28% to $60.8 billion, though operating margin slipped to 31%. Meta spent $31.1 billion on capital investment in the quarter and expects 2026 capital expenditures of $130 billion to $145 billion, largely for servers, data centers, networks and the AI buildout. The advertising machine throws off enough cash to fund all of this experimentation, which no startup competitor can say. It also explains why investors were so eager to assign a revenue line to the spending line.
Now for the detail Wall Street has treated as routine. Reuters reported that Meta is testing a “human concierge” feature in which contractors handle some Muse calls. The trial is being offered to half of Meta’s own employees, on an opt-out basis, and the company describes it as a way to learn about safety and privacy before a public release. Which is all reasonable! There is nothing inherently sinister about humans inside an unfinished product. The bullish view — human involvement declines as the system learns — may well be right. Investors simply do not have the information to assume it, and they have paid as if they do.
Think about which tasks are easy and which are not. Writing an email or finding a restaurant should automate beautifully. Disputing a hotel charge, rebooking a missed trip or canceling a subscription whose provider does not wish you to cancel involves negotiation, missing information and counterparties with incentives to be unhelpful. Someone also has to be accountable when the agent books the wrong date or shares something the user thought was private. The Forbes writer notes having examined the same problem with Google’s Gemini: an agent needs no malice to generate real costs, only permission to act and an incorrect understanding of the assignment.
Here is where the income statement diverges from the product demo. A single task might require repeated model calls, several outside services and, sometimes, a person to clear the exception. And the customers paying $100 a month will probably be the heaviest users, meaning the most valuable subscribers could also be the most expensive to serve. That is a strange shape for a software business and a familiar shape for a concierge business.
There is also the small matter of the commercial internet’s cooperation. Shopify has welcomed Muse, added Shop Pay and treats the agent as another channel to its merchants. Amazon has banned Muse from buying on its site, citing concerns about unlawful activity, privacy and user experience. Meta cannot force merchants to admit its agent; the eventual economics will be set by agreements, integrations and possibly revenue-sharing, none of which appear on an app-store download chart.
The bull case remains genuinely strong. No company in consumer AI has distribution like Meta; WhatsApp alone gives Muse a familiar home, and the core business can fund a long learning period in which contractors’ interventions teach the system where it fails. The strategic prize may exceed subscription revenue: if Muse becomes the place a consumer starts a purchase or contacts a company, Meta captures commercial intent that currently begins at Google, Amazon or a retailer’s own app, and can feed that back into the advertising machine.
The Forbes piece proposes one number above all others: the cost of a successfully completed task. That single figure bundles the compute, the share of jobs finished without human help and the expense of fixing mistakes. If it falls, Muse is learning and subscription revenue scales. If it stays stubbornly high, Meta has built a concierge service wearing a software company’s valuation.
None of this makes Muse a failure; two-point-five million people wanted it before anyone knew what it cost, which is its own kind of evidence. It just means the market has awarded roughly $200 billion in advance, for an answer only the unit economics can provide. Downloads decide very little from here. The verdict arrives each time Muse finishes a job, and someone at Meta adds up what that just cost.