In this episode of 'The Inside Look,' Senior Commercial Real Estate Economist Xander Snyder examines the demand side of data centers and how competing language models, such as DeepSeek, may influence future demand.
Watch below and subscribe to the CRE Insider blog for additional First American Title NCS insights.
Transcript:
AI Has to “Live” Somewhere
Scalable, remotely accessible Generative AI models are changing the way we work. But running large language models (LLMs) like ChatGPT is costly in terms of computational requirements, and usually that computing happens in data centers.
Data centers are large, custom-built commercial properties that share many similarities to critical infrastructure in terms of operating requirements. They must be supplied with a substantial quantity of electricity, and water, and require extensive technical competency to build and maintain.
DeepSeek and Data Centers
Recently, a new large language model, called “DeepSeek”, was released that was supposedly trained at a fraction of the cost and with smaller electricity requirements than existing models. Many have speculated that this will negatively impact demand for data centers in the future, since, the argument goes, less data center capacity will be required to manage these new, more efficient LLMs.
But it’s still too early to know whether that’s the case. It’s not uncommon for new technologies to become more efficient, spurring broader adoption of them. When this happens, it often increases the net demand for the technology, because it’s cheaper and easier to access.
Efficiency Paradox
In economics, this phenomenon is common enough to have a name: Jevon’s paradox. Named after a 19th century economist, Jevon’s paradox describes a situation when a more efficient technology leads to an overall increase in demand.
- It was originally used to describe how a new type of steam engine improved the efficiency of the coal-fired steam engine, and as a result ended up increasing the demand for coal. However, this same concept could just as easily apply to Gen AI and computing capacity today.
As a result, I’d argue, more efficient LLMs could just as easily increase as decrease future demand for data centers capacity. However, the general push towards broader adoption of Gen AI, I think, will be a net positive contribution towards future demand.
Conclusion
Interested in learning more about data center fundamentals? Check out my recent blog post on the First American Economics Center called “Why DeepSeek May Signal Greater Data Center Demand, Not Less.” In this video, I’ve talked about the demand side of data centers, but in that blog post I also talk about the meaningful constraints to new data center supply.
