Are LLM Providers the Next Utility Companies? Why the Analogy Falls Short

With Anthropic filing for IPO, the debate on the future of LLM providers is heating up. Some have compared these companies to electric utilities. Over time, these services will become a commodity. It would become easier to stand up AI data centers. Just as we design appliances to be energy sensitive, we will design applications that consume fewer tokens. The commodity nature will drive down prices and governments may be tempted to regulate price and access.

Like all analogies, the LLM comparison with electricity is limited. The analogy is useful to understand the ubiquity of demand but fails in other aspects. Tokens are not a commodity. They have character. LLMs are sensitive to training and hence quality of outputs vary. Anthropic's surge in popularity for businesses and OpenAI retaining hold on consumer questions is a case in point. So we will not have a PG&E and Con Edison for AI.

If you want to imagine where Anthropic and OpenAI will be in a few years, a more relevant comparison is with investment-intensive services like cloud and airlines. The airline example works because airlines require heavy investments. The capital barriers to entry limit the number of players. Yet each airline has a character, switching costs are low and there is choice for consumers. Project that to LLM providers and you can imagine the emergence of three to four global players, a few national players and some low-cost or niche players.

All the above debate is about public AI. Only 5% of internet is public data. LLMs train on this limited dataset of roughly 11 TB. The battle to package and deploy proprietary data has not yet started. The market for private AI will outstrip the demand for public AI by a significant factor. Think of them as captive power plants or charter planes. That is the next wave and we have seen nothing yet.