June AI Reality Check: AI Tokens, Subsidies, Lock-in Risk

It seems there is another AI-related concern generating a lot of buzz. The talk of a bubble has gone away, and talk of token costs has taken its place. Specifically, there is widespread concern that the costs of using AI, which shows up as costs for use of tokens (a measure of how much data is consumed by an AI model doing something) have risen and become a real budget line item for some companies. This, in turn, has raised concerns that these costs will rise even further once companies are locked into using AI. 

Tokens are just a measure of how much of the AI infrastructure in data centers you, or an AI agent, are using. We’ve known for 2 years that there isn’t enough capacity in data centers to meet the skyrocketing demand for AI tokens, and that has been made much, much worse as the industry has shifted from chatbot-based usage to agent-based usage. Agents use many times more tokens to do something than a simple chatbot response does, so in just the past few months, as companies have started deploying agents at scale, it has suddenly become an issue, causing the fear of lock-in that so many are familiar with from SaaS contracts.

The difference, and why token costs are not going to be the cause of lock-in, is that data centers, and the tokens they produce, are infrastructure, not a user-facing application. In every infrastructure buildout in history, from railroads to chip fabs, overinvestment leads to price collapse and the commodification of infrastructure. But early in the infrastructure buildout, when capacity hasn’t caught up, infrastructure has pricing power, so prices go up. This was true for broadband internet, but it wasn’t when broadband became a commodity. It was true with mobile phone minutes, which were metered and expensive, until mobile coverage became a commodity. And it will be true when the current data center buildout reaches saturation, likely in a few years. 

This does not mean there won’t be some form of lock-in risk associated with AI, but history has shown that lock-in and pricing power happen closer to the user - in applications, or in the case of cell phones, the phone itself.