The ugly economics of consumer AI

After this week, you could say that consumer AI is making a comeback.

Meta’s personal AI assistant, Muse, and its plush mascot, Jolly, have become unexpectedly popular. OpenAI’s Dots, released just yesterday, appears to be pursuing the same cartoonish personal assistant idea. And the up-and-coming Instinct assistant has reached a $10 billion valuation on the strength of agent errands focused on travel bookings, restaurant reservations, and subscription cancellations.

Making a blue case is easy. Agent AI is finally reliable enough to handle everyday tasks. Companies are increasingly marketing their services to the public, and people are getting real value from them. If you are an investor, this is very similar to the ChatGPT announcement in 2022. This is where the raw power of AI opens up product categories that were previously impossible. Who doesn’t want a piece of the action?

But there’s a reason Frontier Labs is cautious about consumer AI. It’s not because the technology isn’t good enough. Even surprisingly popular high-tech products are starting to hit a ceiling on what consumers are willing to pay, and it’s unclear whether better models are actually leading to more profitable consumer businesses. The result has been an industry-wide shift to a human model focused on corporate contracts and expansion by vertical.

If products like Muse and Instinct buck that trend, it’s because they don’t care as much about monetization. But the underlying economics of consumer AI haven’t improved, and those entering this business will eventually have to deal with them.

We were reminded of these economic conditions in Andreessen Horowitz’s semi-annual Market Situation Report. This report pulls numbers from this summer’s PNC research report. Two charts track a slow increase in the proportion of consumers paying for AI services and a slow increase in the amount consumers pay. As of May, 2.2% of consumers were paying for AI, with an average spend of $31 per month.

Andreessen is upbeat about this, saying, “It’s still too early to adopt and leverage mature AI.” There’s a lot of room to grow! However, in both graphs, the pace of growth appears to be very linear. Despite significant model improvements, there has been no significant change in the number of customers willing to pay for AI or the amount of money they are willing to pay for AI. For example, the significant performance improvement from GPT-5.2 to Astra is barely visible in the charts.

While the per-consumer numbers are less impressive, they are still well below the standard break-even point. If we were to adopt Netflix as the standard for a market-saturated online service (325 million subscribers), its annual revenue of $34 per customer would be only $11 billion, less than a third of OpenAI’s operating costs.

If you think PNC is underestimating adoption, similar numbers are available from Bank of America. In March, the company revealed that about 3% of U.S. consumers are paying for AI, an increase of 40% from a year ago. Menlo’s September survey provides a slightly more positive outlook, finding that a quarter of adults use AI every day, and half of those users pay for it.

The problem with the consumer approach is related to cost rather than revenue. AI is unusually expensive to operate, especially when compared to lighter weight previous technologies such as social networking and cloud computing. Even if you have hundreds of millions of paying customers, you can’t guarantee break-even.

To its credit, OpenAI seems to have adapted well to these facts. The company’s widely reported enterprise transformation has been largely successful, with enterprise bookings reportedly doubling since July. The launch of Dots also had a strong enterprise focus, showing how the new personal agent could benefit software engineers and agency creatives. One of the long-standing ways to make money from popular but inexpensive consumer services is to mark them up and sell them to businesses, and OpenAI appears to be following that strategy.

It’s hard to say what this means for Muse and Instinct. Muse has the powerful capabilities of Meta’s personalized ad targeting behind it, giving you more monetization options and more time to issue. Notably, Meta is already exploring the enterprise space.

Instinct has a separate plan that includes receiving a percentage of purchases made through agents, which may result in higher limits. It would probably also avoid the cost of training frontier models, which would help a lot.

But the ugly economics of consumer AI severely limit how big the company can grow without leveraging corporate revenue. This is a lesson the big labs have already learned, and one of the few things in the industry that doesn’t seem to have changed.