Meta Muse: 10 Million Setups, Monetization Plans and the Compute-Cost Problem

Meta disclosed new Muse usage and business plans on October 21, 2026. More than 10 million people in the US and Canada have set up its personal agent. The company now expects a greater than tenfold improvement in compute cost efficiency over the next year.

Those are two different kinds of evidence. The usage is a company-reported result. The efficiency improvement is a forecast, and the distance between them explains the next challenge for consumer agents.

What Meta disclosed about Muse usage

In Meta's third-quarter earnings call, Mark Zuckerberg said the setup milestone took about a month and a half. Roughly 25% of people who have set up Muse use it on a given day.

Shopping is particularly important. CFO Susan Li specified that more than 25% of Muse's weekly active users use it for shopping. Zuckerberg put average purchase size above $200. These denominators matter: weekly shoppers, daily users and total setups are not interchangeable metrics.

A $200 purchase is not $200 of revenue for Meta. It measures transaction size, not the fee an agent provider receives or the profit left after serving the user.

For an agent business, the useful question comes after adoption. Does a completed task generate enough value for someone to pay for the reasoning, searches, retries and other work behind it?

How Meta plans to make money from Muse

Zuckerberg described a free experience for most consumers, eventually supported by transaction fees paid by businesses. He also outlined subscriptions for advanced models or extra usage, and paid business and workplace capabilities.

But that is a business model being developed, not a newly available merchant-fee product. In the follow-up analyst call, Li said Muse was not being monetized outside subscriptions at present. More users and more connectors or business partners were needed before monetizing transactions.

This distinction matters for founders building around agent commerce. A product can demonstrate purchase intent before it has settled the commercial relationship with merchants. Technical integration, payment completion and revenue capture are separate problems.

There is also a product-design question to watch. If a business eventually pays when an agent helps complete a transaction, the user still needs a clear explanation of how recommendations are chosen. Commercial distribution becomes part of the product experience, rather than something that happens only before someone opens the app.

Why tenfold efficiency does not mean a smaller total bill

Meta expects total Muse serving costs to rise as adoption grows. Its efficiency plans include smaller models distilled from Muse Wonder, specialized models, inference improvements and better use of off-peak capacity. In the follow-up call, Li cautioned that reaching tenfold efficiency still requires further technological advances.

Lower cost per task and higher total spending can happen together. If the cost of an equivalent workload falls tenfold while the number of workloads grows twentyfold, the total bill doubles. That is an illustration, not Meta's usage forecast.

The engineering challenge is also about what counts as an equivalent workload. A cheaper model that needs more retries or more human review may not produce a cheaper completed task. Teams need to compare successful outcomes at the required quality, rather than only token prices.

The infrastructure spending is broader than Muse

Meta's quarterly results put capital expenditure, including finance-lease principal payments, at $34.91 billion. Full-year guidance narrowed to $135 billion to $145 billion from $130 billion to $145 billion. Quarterly operating cash flow was $32.79 billion and free cash flow was negative $2.12 billion.

These are company-wide numbers, not a Muse cost breakdown. Meta describes capacity serving model training, its existing business and new AI products. Dividing the entire capital budget by Muse users would produce a dramatic number, but not a meaningful cost per user.

For AI operators, the next useful disclosure is therefore less about another adoption milestone and more about the relationship between task quality, recurring use and contribution after serving costs. A personal agent becomes a durable business when the work people keep asking it to do is also work it can afford to keep doing.

I'm Paras. I work at the intersection of AI, engineering and distribution.

I help AI founders turn product economics and technical updates into clear stories that developers and business buyers can understand and share.

Book a 30-minute call.