Is the Jevons paradox the new “money-good”?
It may seem strange to relate AI-model competition to the direction of Treasury yields but increasingly, the former is having an impact on the latter. In my yen carry trade article, I discussed how international balance-of-payments dynamics explain foreign-capital financing of US technology companies. Today, it is worth looking at the dramatic impact the AI spending by these same companies is having on the US economy.
Estimates vary, but many economists think AI infrastructure spending will contribute 0.75% – 1.5% to US GDP this year.1 Until recently, there was little reason to believe that would not materialize. Now, however, the Chinese company DeepSeek claims it was able to create a top-tier AI model with significantly less resources and training than its US rivals. The company also announced plans to provide its model via open-source to developers. If both statements are true, then today’s prevailing narrative that massive capex for AI-model development — which has been fueling US growth —may face scrutiny, particularly as models become commoditized and free.
Following the DeepSeek news, US tech companies have not announced plans to reduce capex, nor has the market drastically penalized them for that. In my mind, there are two reasons for this. First, over the past few years, the stock market has rewarded companies with the highest earnings momentum. In an elevated capex cycle, earnings momentum should be strong, as a company’s return on investment can show up immediately in earnings while capex can depreciate over many years. That explains substantially higher earnings growth, compared to free-cash-flow growth.
The second reason has to do with the Jevons paradox, which holds that as technology becomes more efficient, consumption counterintuitively increases, often with unintended consequences. A couple of months ago, no one was thinking about the Jevons paradox in the context of AI capex. Tech companies were spending massive sums on AI to develop new cutting-edge models. Post-DeepSeek, they are now arguing that major capex is necessary because of the widening model usage and the need to fund further investment in inference and data-intensive agentic AI. So far, they have generated an attractive return on these investments, so perhaps they are correct. But this is a significant shift in the narrative, similar to that with mortgage bonds throughout 2007 and 2008.
