AI Has The Strongest Network Effects Ever, And That Is a Problem

I kept running into the same mechanism in different papers this semester. Katz, Shapiro, Farrell, and Saloner built the theory of network externalities across the 1980s, and the core claim is deceptively simple: a technology becomes more valuable as more people use it. A telephone is useless if you are the only owner. A fax machine becomes valuable only when your supplier and your customer and your regulator all have one. Same-side network effects mean that every additional user makes the product more attractive to the next potential user. The value lives in the collective, not in the device.

I thought I understood this until I tried to apply it to AI and realized the scale is not comparable to anything we have studied before. A fax network or a social media platform has strong network effects, but the value increase is roughly linear with new users. More users on Facebook means more content, more connections, more ad inventory. It is a strong effect, but it has a ceiling. There are only so many people on earth and only so much time they can spend on a platform. AI is different. The relationship between users and value is not just additive. It is recursive.

The data flywheel is what makes AI network effects the strongest the economy has ever seen. Every interaction with a frontier model generates data. That data trains the next version of the model. A better model attracts more users. More users generate more data. The loop does not stop. It compounds. OpenAI is the clearest example. Every ChatGPT conversation, every thumbs up or thumbs down, every time a user edits a response or asks a follow-up, that signal goes back into the training pipeline. The chat data is not just usage telemetry. It is the raw material for GPT-5. So the user who joined ChatGPT yesterday is not just a customer. She is also a contributor to the product's improvement, whether she knows it or not.

Katz and Shapiro (1985) distinguished between the value of the installed base and the value created by future adopters. In AI, the installed base generates the training signal that determines how good the future model will be. The user today is not just adding value to the network. She is building the network's next version. That is a fundamentally different dynamic from anything the telephone or the operating system or the social network produced.

The difference between same-side effects and cross-side effects becomes important here. Same-side effects in AI are visible in consumer chat products. More users attract more users because the model gets better. Cross-side effects show up in the API ecosystem. When developers build applications on the OpenAI API, they create value for end users who never open ChatGPT. A startup that embeds GPT into its customer service tool is a complementor. The more complementors join, the more use cases the model covers, and the more valuable the platform becomes for everyone. This is the platform ecosystem logic Parker, Van Alstyne, and Jiang (2017) described. The core improves the periphery, and the periphery feeds data and signals back to the core.

What makes this a problem is that the flywheel creates winner-take-most dynamics faster than any industry I can track. Once one model pulls ahead in quality, it attracts the most users, which gives it the most training data, which extends its lead. The leading model does not just win the current generation. It locks in the data advantage for the next generation. Competitors face a compound disadvantage. They have to build a better model with less feedback data, while the leader improves faster precisely because it is already ahead.

The compute cost amplifies this concentration. Recent reporting suggests training a frontier model now costs upward of a hundred million dollars, and inference costs for a large user base run billions annually. A new entrant needs not only comparable compute but a user base large enough to generate comparable feedback data. The two requirements reinforce each other in a way that makes entry nearly impossible. You cannot bootstrap the flywheel without users, and you cannot attract users without a model that is already competitive with the leader. This is the same logic Tushman and Anderson (1986) applied to technological discontinuities, but the compounding data dynamic makes the incumbency advantage stickier than any capital equipment barrier I have seen.

I think this explains why we have three serious foundation model providers instead of thirty. In 2023, there were dozens of teams with plausible claims to frontier capability. In 2026, the serious conversation is about three or four. The rest either exited, pivoted to applications, or are running on fumes. Network effects concentration, amplified by compute cost, filtered the field faster than any antitrust intervention could have.

And that is the problem I keep coming back to. The market has not tipped yet, but the direction is visible. If one provider locks in the dominant position through the data flywheel, switching costs for developers and enterprises will be enormous. A company that has fine-tuned on one API, integrated one model into its workflows, and trained its staff on one interface cannot switch providers without retraining models, rewriting integrations, and absorbing a performance hit during the transition. The DOJ should be studying this through network effect theory right now, before the market tips, not after. I wrote about why platform governance determines who wins in multi-sided markets, and the same logic applies here. The rules that govern access to the model, the pricing of the API, and the terms of data use are not neutral. They determine who can compete and who cannot.

I do not know whether the right answer is interoperability mandates, training data disclosure requirements, or something we have not designed yet. What I know is that the IS field has a framework for this. Network externalities theory gives us the language to explain why concentration is happening and why it is accelerating. We should be using it before the market closes in a way that no regulation can reverse.