How my time at an ISP shaped how I view AI businesses

“Spectrum licenses don’t depreciate.” “Yes, the operating margins are huge, but the capex is large. Look at Comcast’s debt-to-equity ratio.” “Build-to-demand is a viral marketing campaign that makes the installations cost-effective.” “Content acquisition eats at the margins of providing television service, but people don’t want OTT.” “The operating expenses for business customers are significantly higher.”

The job doubled as a business class. When I joined Google Fiber in 2012, it had a hundred or so employees, an order of magnitude smaller than the thousands at YouTube. That meant visibility into every aspect of the business. Understanding those mechanics benefited me significantly in the roles I have taken since. Here are two examples that serve as instructive use cases:

1. Connecting a home to the Internet requires an up front investment, but once established, is relatively inexpensive to operate, resulting in a high operating margin. Furthermore, customers tend to be sticky. Thus, a company that is able to handle the capital expenditures to connect homes, could, given a large enough user base, benefit from the healthy operating margins to offset that cost. It’s also worth noting that certain investments to provide connectivity, like spectrum licenses, are assets that don’t depreciate, so they effectively function as cash and could be sold at a later date.

2. Internet providers charge a lot for TV service because customers want it. Unlike the Internet connection itself, the cost of providing that content on a monthly basis is high, so the margins are very low. However, because other ISPs also offer TV service, this becomes table stakes in order to have parity with one’s competition. There are multiple ways to respond to this. One is to make up for the low margins with high volumes, which is what the largest providers are able to do. Another is to bundle the television service with higher margin offerings (triple play). A third could simply be to encourage customers to use over-the-top services instead of their own TV offering, which may reduce revenue but depending on the price point could actually increase profits.

I’ve thought back to these examples when thinking about how the AI landscape might evolve. There are currently multiple companies spending a lot of money to build large models. The bet is that there will be a way to recoup these costs by charging customers for the privilege of using these models (even so-called open models have language around this). If costs for updating and serving these models could eventually be reduced, then one could image a a business model similar to an ISP’s connection to the home.¹

What about downstream business customers of these AI services? If the costs of serving these models cut significantly into their margins, then they are in a similar position to the TV service. Options available to them may include making up for the lower margins through higher volumes, bundling their AI product with a higher margin offering, or developing a strategy in which the cost associated with using these models does not scale with the size of their business.

It’s early days for AI companies, who themselves have an interest in creating a healthy ecosystem to nurture business investment in this space. As these AI models continue to amaze me in the art of the possible, I am equally excited to see how the business models evolve.

1. There is also a world in which either some of those costs do not come down and/or some data to train these models needs to be purchased/licensed, in which case the business model may resemble something more like TV service. ↩