Data centers house the IT infrastructure (i.e. cloud computing, data storage, etc.) for organizations. They require a lot of power, space, redundancies, and a huge amount of capital. Physically they look like large commercial buildings with rows and racks of servers with switches and cables running between them.
Helpfully, SemiAnalysis categorizes data centers into three buckets:
Recently, there’s been a large push to build large data centers by “AI hyperscalers” due to increasing demands for AI training and inferencing. Hyperscalers are the giant tech companies that invest, own, and operate in massive data centers that are needed to deploy AI workloads. The top five are Amazon, Microsoft, Google, Meta, and Apple/Oracle (depending on the metric). A few are building data centers for their own products (like X) to improve ROI, but they also face a large demand from AI labs like OpenAI, Anthropic, and xAI for capacity and platforming.
If we distill AI data center demand into training and inference, we should think of the training demand as something that is lumpy and is the driver for quickly constructed data centers; inference demand is more product-driven and thus creates the need for more reliable and efficient data centers. In the long run, the latter should drive the load of data center investments; several sources believe by 2030, the majority of the load will likely shift towards inference (NREL, JLL).
Capital Formation → Power → Site Selection → Facility Infrastructure and IT Equipment → Operations → Product Demand