For six years, I taught my Trinity College students that switching costs were the moat.
The textbook example was the iPhone. Once one person in a family had one, the rest tended to follow, because iMessage worked best when everyone was inside it. The green bubble was a tax on the outsider. Group chats split along device lines. Switching meant degrading the experience for everyone you loved.
The example I would give them now is the platform you are reading this on. LinkedIn does not lock you in with photos or messages. It locks you in with the network you spent a decade building. Every connection you accepted, every comment that sparked a conversation, every recommendation that took ten minutes to write. You could move to a competitor tomorrow. Your network would not come with you. The cost is not a subscription fee. It is a decade of social capital.
I taught these examples for six years. A few months ago, I dissolved one of the deepest versions of the same idea on my screen in 60 seconds.
I imported my ChatGPT memory into Claude. The thing my students were taught was the most defensible asset in modern technology, the personal context that took a year to build and supposedly could not be transferred, moved between providers with a single prompt. A year of preferences, conventions, working style, copied from one foundation model to another in less than a minute. I sat there staring at my laptop trying to reconcile what I was watching with what I had taught years ago.
The journalism is calling this a switching cost story. Built In frames it as portability eroding lock-in. a16z walks through Hamilton Helmer’s Seven Powers framework and concludes that switching costs are the one moat AI is genuinely changing, but argues this is good for software because companies will have to compete on quality. TechBuzz calls it the AI equivalent of mobile phone number portability. All of this is true. None of it is the whole picture.
What the journalism is missing is a distinction my students used to write on their exams.
There are two kinds of network effects, and most commentary conflates them. The first is the direct effect, the one that operates between users. iMessage is more valuable because my family is on it. LinkedIn is more valuable because my professional network is. The second is the indirect effect, the one that operates between users and the producer. The more people use a product, the more data the producer collects, the better the product gets, the more users it attracts. The first is visible. The second is invisible.
AI assistants never had strong direct network effects. I do not need other people to use Claude for Claude to be useful to me. The lock-in was always something else: the personal context I had accumulated inside one provider’s memory. Memory was the moat, dressed up in network-effect language. Memory Import did not weaken a network. It dissolved a switching cost that was never really a network in the first place.
The indirect effect, the one underneath the surface, is intensifying. Every conversation a user has with a foundation model becomes part of the training distribution that improves the next version. The more users a model has, the better it gets, which attracts more users. This is not a metaphor. It is the actual production function. The INET working paper from October 2024 modeled this carefully. Compute, data, and talent all exhibit massive economies of scale. Frontier model training now costs hundreds of millions of dollars. The barriers to entry at the foundation layer are getting higher, not lower.
So the consumer story and the production story point in opposite directions at the same time. At the surface, switching is becoming radically easier. Underneath, the producers are consolidating. The market is becoming more contestable for users and more concentrated for builders. Both things are true. Most commentary picks one and ignores the other.
For the people whose careers were built between these two layers, this matters in a concrete way.
For the last two years, an entire category of companies has lived in the space between the foundation models and the end user. Wrappers. Connectors. Vertical AI tools that took a base model and bolted it to a specific data source or workflow. Their pitch was that switching was hard because they owned the connection to your data. The teams inside those companies built careers on the assumption that the connector layer was where the durable value would accrue.
Then MCP and native connectors arrived. Foundation models can now plug directly into your email, your calendar, your documents, your codebase. The integration layer that made its money being the connection is being compressed by an open protocol the foundation labs themselves are publishing. The companies in the middle are being squeezed from both directions. The consumer side is becoming more contestable. The production side is becoming more concentrated. The middle is being thinned.
This is the part of the story that does not fit cleanly into the journalism’s frame. It is also the part that affects the most people. The workers inside connector companies, the founders who raised on integration moats, the engineers whose specialization was the seam between models and data. These are the careers being restructured by a shift the headlines are still describing as a consumer switching story.
For the people building inside this market, including me with my own small AI products, the implication is the same one that has always been true and is now louder. The defensibility of the previous era does not transfer. Switching costs from accumulated context are being engineered away by the foundation labs themselves. Connector economics are being compressed by open protocols. What remains is what was always going to remain in the long run: the quality of the product, the trust of the user, the depth of the workflow, the specificity of the problem.
I am not writing this as a lament. The textbook I taught for six years described an equilibrium that was always more fragile than it looked. What is happening now is not the death of network effects. It is their migration.
The direct effects that powered the social era never really came to AI assistants. The indirect effects are concentrating in the foundation layer. The connector layer in the middle is being thinned. The map of where defensibility lives in technology is being redrawn in real time, and most of the people writing about it are still using the old map.
The moat I taught is dissolving on my screen. The one I will teach next will be a different shape entirely.
— Irena