The Second Quarter Chapter IV · Technology

Chapter IV · Technology

Why AI Subscriptions May Not Be Enough

9 min read Miracle Roch

I have been thinking about the business model for AI companies, and I have become increasingly convinced that the subscriptions model, which underpinned the modern day internet, will not work in the AI era. Or, perhaps, subscriptions will work, but they may not be enough. Today, the path to monetization for most AI companies is fairly straightforward; give users the product for free, limit what they can do, and then charge extra to the users who want more. $20 a month. $30 a month. $200 if you are a power user. There is nothing particularly wrong with this model. The problem is that it creates what I have previously called the Netflix problem.

I have written about the Netflix problem before. Netflix made an extraordinary bet on subscriptions – pay us a relatively small amount every month and we will give you an endless amount of entertainment. The problem, of course, is the word endless. If I finish a good show today, Netflix needs another good show waiting for me tomorrow, and another one after that. So Netflix is stuck in this strange cycle where it has to keep spending enormous amounts of money creating and buying content because the moment viewers stop finding things they want to watch, they switch to other streaming platforms and soon enough, start wondering why they are still paying for Netflix. But there is a limit to how much Netflix can charge me every month. Netflix’s cost of keeping me happy can continue to increase, but my willingness to pay them does not necessarily increase with it. Recent reporting about Netflix’s content challenges is another confirmation that I’ve been right.

I think AI companies are going to face a version of this problem, perhaps an even worse one because AI is way more expensive that content. The models have to get better, the chips have to get faster, more data centers have to be built and those data centers need enormous amounts of power. After spending all that money training the models, the companies still have to spend money when we actually use them. The usual response to this argument is that compute will become cheaper, and I am sure it will (maybe in 50 years). But I don’t think that solves the problem people think it solves because, as AI becomes cheaper and better, we are going to ask it to do considerably more, so the demand for even more compute will be exponentially higher than the user forecast. A few years ago, we were impressed that ChatGPT could write an email. Now I can ask an AI agent to research a holiday, compare flights, check hotels, look at reviews, build an itinerary and potentially make the booking for me. Tomorrow I will expect it to do all of that while also knowing my calendar, my budget, which airlines I like, where I prefer to sit on a plane and that I absolutely do not want a six-hour layover in Chicago to save $26. So yes, the cost of one unit of intelligence may fall dramatically, but if we consume 100 times more intelligence when it does, the economics don’t automatically disappear.

There is another problem with subscriptions that I think is even more important; subscriptions don’t price individual actions. When I pay $20 for an AI subscription, the AI company gets the same $20 whether I ask ten questions that month or use the product relentlessly every day. The economics therefore rely on a kind of Russian roulette in which the company needs enough low-consumption subscribers to subsidize the people hammering the product. This is not unique to AI. Gyms have done some version of this forever, Insurance companies too. The person who pays for a membership and shows up twice a month is a wonderful customer; the person who shows up every morning, showers there and uses every machine is considerably more expensive. The economics work because there are enough of the first set to subsidize the second. The difference is that with AI, particularly agentic AI, the gap between the cost of serving those two customers could become enormous.

There is a meaningful difference between asking an AI to rewrite an email and asking an agent to spend twenty minutes searching the internet, comparing dozens of flights, querying several services, reasoning through the options and eventually executing a transaction. Yet under a pure subscription model, those two actions may generate exactly the same incremental revenue for the AI company – zero. I don’t think that makes sense long term. The better model, I suspect, is one where the subscription pays for access to the intelligence while certain actions generate additional revenue. You might still pay $20 every month for your AI, but if that AI books a $1,200 flight for you, somebody should probably pay the AI company for successfully facilitating a $1,200 transaction. Maybe that is you. Maybe it is the airline. Maybe it is both. The exact structure matters less to me than the principle – revenue should increase when the AI performs economically valuable work.

This is why I find what Meta is doing with Muse so interesting. Muse itself is impressive, but I actually find the proposed business model more interesting than the product. Mark Zuckerberg has said Meta intends to keep Muse free for most users while making money from some of the economic activity Muse facilitates. When I first heard that, my reaction was that this simply makes much more sense. Instead of trying to squeeze progressively larger subscriptions out of users as AI becomes more expensive and capable, Meta can potentially make money from the economic activity the AI creates.

I book almost all my flights using AI now. I tell my personal agents where I am going, my dates, what airports I am willing to use, whether moving the trip by a day saves money, what kind of layovers I will tolerate and so on. Now imagine using Muse for this. I tell Muse that I need to be in New York next Friday, find me the best flight under $700, I don’t want a layover and don’t put me on a 6 a.m. flight. Muse searches the airlines, checks my calendar, knows my preferences, finds a 10:15 a.m. nonstop for $612 and asks if I want it booked. I say yes. Now comes the interesting question: who should pay Meta for that? The current AI business model says I should. I pay Meta $20 or $30 a month because Muse helped me. But there is another answer, the airline should pay Meta too. The airline just acquired a $612 customer. Airlines already pay Google, travel websites, credit-card companies, affiliates and all sorts of intermediaries for access to customers. Hotels do the same thing. So do insurance companies, restaurants, retailers, banks and almost everyone selling something on the internet. If Muse delivers a customer who is ready to buy, why shouldn’t Meta participate in that transaction? Perhaps I pay something too. Maybe booking the flight through my agent costs $2 or $5. Perhaps I pay nothing because the airline’s commission covers it. Or airlines or aggregators pay Meta to maintain connectors that allow Muse to search inventory and execute transactions directly. There are many ways to structure it, but they all produce the same important result: Meta makes money not just because I have Muse, but because of what Muse does for me.

This creates a very different economic relationship. Imagine I use Muse to book flights, reserve hotels, order groceries, find a plumber, buy concert tickets, renew my insurance and find a restaurant for Saturday night. Under a pure subscription model, the more I use Muse, the more compute I consume and the more expensive I become as a customer, while my subscription remains largely fixed. Under a subscription-plus-transaction model, the relationship changes. The subscription gets me through the door, but my activity creates another revenue stream. If the AI becomes twice as useful and begins doing twice as much economically valuable work for me, the company has some mechanism for participating in that additional value. That is a much better loop because revenue can begin to scale with usefulness rather than merely with the number of people willing to maintain a subscription.

There is another part of this that I think people are underestimating. For most of the internet era, the most valuable thing technology companies competed for was our attention. Facebook wanted your attention, Instagram and TikTok too. Google was slightly different because Google could often capture something even more valuable: your intent. If I search for “best shoes for running,” I may be researching. If I search for “Nike Pegasus size 12,” I am getting closer to buying. Google Ads built an enormous advertising business around understanding this difference. AI agents could go one step further because when I tell Muse, buy me the Nike Pegasus in size 12 if you can find it below $130, there is almost no ambiguity left. I am no longer merely browsing; I have given a machine permission to transact. That could make AI agents one of the most valuable positions in the commercial internet: the layer sitting between intent and transaction.

And that changes something else. Today businesses obsess over Google search rankings and spend billions on SEO because appearing first when someone searches for something can mean money. But what happens when people stop searching? Instead of Googling “best hotel in Miami,” I may simply tell my AI to find me a nice hotel in Miami for three nights, somewhere quiet, close to the beach, under $400 a night, and remind it that it already knows what kind of hotels I like. Now the hotel doesn’t need to convince me first; it needs to convince my AI. I suspect an entire industry will eventually emerge around this, just like one did around SEO. Companies will optimize their products, prices, APIs and information so that AI agents can discover them, understand them and recommend them. We spent the last 25 years optimizing websites for search engines. We may spend the next 25 optimizing businesses for machines that make decisions on our behalf. This is not good news if you work for Google search.

There are obvious problems with all of this, and perhaps the biggest is trust. If businesses are paying the agent, can I still trust the recommendation? If Delta pays Meta more than United, does Muse suddenly think Delta has the “best” flight? What happens when every business is bidding to become the answer my AI gives me? These are not small questions. If AI agents become as important as I think they will, they will eventually become enormous questions because an agent that controls transactions could become far more powerful than an advertising platform that merely influences them. But these problems don’t invalidate the business model. If anything, they demonstrate how valuable the position could become.

Many years ago, I wrote about Twitter, and argued that the company needed to rethink how it made money. Twitter had become a running revenue joke despite becoming something enormously important. News broke there, politicians communicated there, companies managed their reputations there and entire cultural movements happened there. There was enormous economic value moving through the platform but Twitter simply wasn’t capturing and retaining enough of it, their stock price did not reflect this neither did their revenues which was mainly based on ads. I think there is a lesson there for AI companies because creating value and capturing value are two different things. Right now AI companies are creating an absurd amount of value. I use AI every day for things I would happily have paid another human considerably more money to do for me, yet the industry’s dominant answer to monetization is still basically would you like the $20 plan or the $200 plan? That cannot be the end of the story.

Meta’s idea is interesting because it asks a different question. Instead of asking how much more the user can pay every month for AI, ask how much economic activity AI can create and then find a sensible way to participate in that activity. If my AI saves me three hours, finds me a cheaper flight, books my hotel, orders my groceries and eventually manages half the transactions in my life, there is an enormous amount of money flowing through that relationship. The company that owns that relationship does not necessarily need to charge me $100 a month. It needs to figure out how to stand close enough to the money when it moves. That may turn out to be a much bigger business.

Someone at the frontier labs should give me a call!