Making sense of AI in the economy

In 2025, cap-ex spending related to data centers powering LLMs accounted for much of the strength in the US economy. Data center construction by the major tech companies represents around 1/3 of all cap ex by the S&P 500 companies. One big question for the economy in 2026 is whether these spending levels are supported by real demand for AI by end users.

The case for optimism

Token demand in 2026 is clearly going to be much higher than 2025. Two giant use cases for LLMs have emerged and will continue to expand massively.

Search

I don’t know if 100% of search will move to LLMs, but LLM powered search is going to expand substantially. The results from LLM searches are undeniably better than 10 blue links. ChatGPT is topping the charts for mobile downloads in the App Store and users are predominantly using this app for search queries. This trend will continue to play out in 2026 and for the next few years.

Software

Using LLMs in software development could be the biggest change in the profession since the introduction of compilers. You don’t need to believe that LLMs are taking all the jobs to recognize that they will have a giant impact on the profession.

Looming uncertainty

There’s no doubt AI demand is bomming, but how much of it is economic. Datacenter investment must be supported by revenue from end user demand. At present, the companies representing most of the LLM demand are burning piles of cash.

Of the four major model publishers, three are losing billions of dollars. OpenAi, Anthropic, and xAi are all known to be deep in the red. Only Google is able to support their AI expansion with revenues from other lucrative products.