A counter-intuitive theory is emerging to explain the skepticism around the trillion-dollar AI infrastructure boom: what if AI gets too efficient? Sebastian Siemiatkowski of Klarna has floated this idea, noting that new models from companies like DeepSeek are drastically reducing the computing power needed for AI tasks. If this trend continues, the massive data centers currently being built could turn out to be redundant.
This “efficiency paradox” poses a massive financial risk. The current market valuation of companies like Nvidia and the data center REITs is based on the assumption of exponential demand for chips and power. If software improvements reduce the need for hardware, the demand curve collapses. Siemiatkowski calls the current spending “nervous” and lacking “thoughtful thinking.”
The market is starting to price in this risk. The sell-off in tech stocks and the broader market malaise—visible in the FTSE 100 and Stoxx 600 declines—reflects a growing doubt about the return on investment for these massive capital projects. If the hardware becomes a commodity or is less needed, the $4.5 trillion valuation of Nvidia looks perilous.
This theory also impacts the energy narrative. Sundar Pichai has warned about energy constraints, but if AI becomes more efficient, the energy crisis might be less severe than predicted—but so would the profits for the energy and infrastructure companies betting on the boom.
Ultimately, this view suggests the bubble isn’t in the AI technology itself, which is real and revolutionary, but in the hardware build-out, which may be repeating the mistakes of the past.