Meta Platforms' Muse personal AI agent is an early hit, holding the iOS App Store top spot since soon after its Sept. 8 release. According to research firm Human, it was already the No. 1 agent last month, responsible for 40% of agent-based traffic on the internet. Largely because of its release, agent traffic tripled in a month.

Meta is probably losing money on Muse now, and it will be very hard for it to defray the costs down the line.

As it demonstrated with its Threads microblogging platform, Meta's strategy is to scale a new product and then monetize it with ads. But Muse may be different, because it's promising to bring a level of privacy and security to it that Meta isn't known for. Soon, interactions with Muse will be opaque to everyone aside from the user, including Meta. The kind of highly effective ad targeting that Meta's data collection enables in Facebook and Instagram may not work as well in Muse. If user behavior in Muse is anonymized, it won't provide Meta with data to further refine its targeting.

In any case, Meta isn't yet talking about ads in Muse. For now, it is looking to e-commerce and premium subscriptions to subsidize the costs.

And those costs may be very high. At least to begin with, Meta is being very generous with the Muse free tier, allocating up to 100 million tokens a week for each user. Tokens are the basic unit of artificial-intelligence text; for example, this sentence is made up of about 30 tokens when measured by common methods. Though we don't know exactly what the computing costs of 100 million tokens are to Meta, we have a decent idea of what it would charge if someone were using the underlying model, Muse Spark, to do other things like writing software code using the company's application programming interface, or API.

On the API, there are different prices for tokens depending on how they are being used, with the cheapest being "cache" tokens. Even if 97% of a Muse user's tokens are for cache, free-tier users are getting the equivalent of an $1,100 annual API credit. In the past 12 months, Meta has generated only $63.54 per each of its 3.6 billion daily app users, almost entirely through ads. The tough job ahead of Meta is bridging that gap for Muse.

It's unclear what Meta's profit margin is on the Muse Spark API, but we can deduce that it isn't as high as the top models from OpenAI and Anthropic. In the "cost per task" metric of research firm Artificial Analysis, Muse Spark is closer to inexpensive Chinese models than the ones from the U.S. frontier AI labs. The Muse Spark API is a great deal for users, and Meta may well be underpricing it to gain market share.

Revenue per user from the Muse agent is likely going to have to be much higher than what Meta earns from its apps to recoup the costs. If Meta were able to capture all of Visa's revenue with its Muse e-commerce fees -- $45 billion in the past 12 months -- that could be enough, but scaling to the size of Visa won't be easy.

This year, Meta intends to spend up to $145 billion on capital expenditures for new AI data centers.

The Muse math gets even tougher for Meta in the paid tiers. The $240-a-year plan offers the equivalent of $5,500 in annual API credit, and the $1,200-a-year tier comes with a staggering $33,000 of annual credit. On the free tier, where most of the casual users of Muse will be, Meta can count on many of them not coming close to their free token allocation, bringing average token use down. But people on the paid plans will likely be power users who will be using Muse to run their lives like a full-time human personal assistant would. They may come close to using up their entire token budget every week.

Meta didn't respond to a request for comment on the wide gulf between its average revenue per user and Muse token allocations.

The company is hoping that everyone becomes a power user, and that they make Muse a daily habit like its social-media apps. But the more successful Muse is, and the more people use it, the more difficult profitability will become.

This dovetails with a recent report from research firm SemiAnalysis, which showed that the API costs of the token allocations for OpenAI and Anthropic subscription plans are well in excess of their subscription prices. In effect, at all three companies, API customers are subsidizing free and subscription users.

The situation is indicative of where we are right now in AI. The costs are evident in the hundreds of billions of dollars being spent for new AI data centers. Cloud providers like Amazon.com, Microsoft, and Alphabet are earning revenue from renting out these servers, but companies with no cloud service like Meta are fishing around for income, and it isn't clear exactly where the profits come from when the expenses are so high.

Meanwhile, Meta shareholders have to endure reduced operating profit margins, down to 37% in the latest quarter (after excluding some one-time charges) from 43% last year. Making Muse a mass-market service is going to be no mean feat, but making it a profitable one may prove the bigger hill to climb.