AI economics: when only 2.2% of users pay, which solutions AI companies choose to make money

This week the two most mentioned consumer AI products are Meta's Muse and OpenAI's Dots. While Muse exceeded 5 million downloads thanks to a large advertising campaign, Dots is an "Always-On Agent" product that users can assign tasks anytime, anywhere. Both are real, working products but the question is: who will pay for these two products.

The model got better very quickly, but the number of people paying only grew slowly and steadily

According to analysis by Andreessen Horowitz based on PNC research, as of May 2026 only 2.2% of consumers paid for AI services, with an average spend of $31 per month. The growth rate of paying users is almost a straight line, not accelerating. While the model has improved a lot, the number of paying customers has not increased as expected, if a better model attracts more people to pay, the growth curve must curve at an exponential rate. The harsh reality is that it still goes straight.

Only 2.2% of users are willing to pay for AI products, with a payment of about $31/month This is important because AI has a different cost structure than traditional software. With a regular software application, once built, serving an additional user costs almost no extra money, or the marginal cost is close to 0. Meanwhile, with AI, each query will force the server to run the model and spend money on computation. And so, using it more will cost a lot. Therefore, a free user is not a "free" user because then there will be an amount that the person paying, revenue from advertising, and the business compensates for that free part. And if only 2.2% of consumers pay, the above compensation is probably not the ones who pay.

Four directions companies are trying to offset costs

When individual subscriptions weren't enough, companies tried different directions, including four. The first direction is enterprise, and OpenAI is moving in this direction and according to reports, their order volume has doubled since July. This was also shown at the Dots launch when they emphasized the applicability for engineers and service companies, not individual users. So, perhaps if AI companies want to be successful in the long term, businesses are the ones they need to target to be profitable, and leading laboratories have been understanding this lesson most clearly.

OpenAI launches Dots but is also aimed at business users The second direction is advertising, with Meta following this path. Muse relies on Meta's existing advertising infrastructure, so users do not need to pay directly. This is an advantage that pure AI modeling firms do not have. But everything has a trade-off, here it is user trust. Another hot news this week is that Muse was able to read private messages on the user's Mac even though he had turned off Full Disk Access. Meta countered, saying that Muse's messaging integration is entirely optional and that users must go through multiple authorization steps. Although the two sides have not agreed, but honestly with Meta's history, I don't have too much faith in how they treat user data. Obviously, with a product that lives on advertising, reading user data is a must and users have every right to doubt this.

Meta just got caught in a scandal of secretly reading users' messages, but now every time he touches Meta, he feels worried The third direction is to collect commissions with each transaction. Instinct is basically an agent-like AI assistant founded by Noah Shinn, with the ability to text or call to do work on your behalf such as booking trips, restaurant reservations, canceling subscriptions, and has just been valued at 10 billion USD even though it only has 14 employees. They plan to collect commissions from bookings and orders made through agents. In other words, the agent makes money when it completes transactions on behalf of the user The fourth direction is to charge according to usage, and Google follows this direction in parallel with subscriptions. The newly launched Gemini 4 Argon is open to Google AI Ultra subscribers and paid API customers. For a fixed customer group like Google AI Ultra, the fee is like that. If you use it more or less, Google will have to shoulder the burden for the users. The less users will partly offset the cost, but in this package, it is easy for people to use it less. The point I find noteworthy is the price list for the API group: the introductory price is 2 USD per million input tokens and 10 USD per million output tokens, then increased to 4 USD and 20 USD. With this calculation method of Google, customers who use more pay more, if they use less they pay less.

Each way to make money has its own weaknesses

Of course, each of the above methods has certain disadvantages. While individual subscriptions provide clear and predictable revenue, the conversion rate is currently only 2.2% and the number of subscribers is not growing as quickly as expected. As for Meta's advertising solution, it allows them to serve the masses for free. For Meta, the risk of data, personal security and trust is a huge problem, which is no different from pushing that difficulty towards dfung users. The commission solution through revenue brings real value from the agent's activities, but it is only applicable when the user actually assigns the agent tasks such as making a reservation or placing an order, and I have not seen any data to show that this method has worked on a large scale. The final solution, cost-based and business-friendly billing, is facing stiff competition with many competing names here: Google Gemini 4 Argon, Claude Opus 5.5 and GPT-6 Astra But in general, after all, the product is not non-existent, but what AI companies lack is probably a breakthrough way to make users willing to spend money, thereby collecting their money.