Overview
As AI technologies escalate demand for data center expansions, material cost increases and intensifying political headwinds are making buildouts more challenging to navigate than ever. Planners must weigh these variables against an already diverse range of factors including capital timing, electricity and water availability, land use, and other risks.
Modeling data center expansion decisions becomes even more complicated when one considers that demand and supply may co-evolve. Capacity availability can accelerate adoption and enable new workloads, or it can change the timing at which latent demand becomes observable. This makes the expansion problem more difficult than long-term infrastructure forecasting based on factors like exogenous climate projections.
But what if we treated data center expansion as a sequential decision process instead? This NSDPI paper uses this approach to formulate a finite-horizon Markov decision process (MDP) in which a developer or planner decides when to build low-density capacity, when to start and complete high-density capacity, and when to expand electricity, water, and land resources. Because demand capacity follows stochastic S-curves, the model allows demand timing, saturation level, and uncertainty to be varied explicitly.
By making the data center expansion process explicit, this MDP empowers developers to better model and plan their buildouts around demand uncertainty, supply availability, and idle capacity.
Key Takeaways
- Data center service quality and stranded capacity are in fundamental tension. Every experiment confirmed the same tradeoff: Earlier high-density commitment improves service but increases idle exposure.
- Demand beliefs are observable in expansion behavior. For example, planners trained on deterministic demand forecasts build earlier and more aggressively than those trained under uncertainty.
- The separation between training beliefs and evaluation conditions in the MDP mean it also can be used as a diagnostic structure for intelligence assessments. Analysts can ask: given the expansion signals observed to date, which training distribution best explains the observed action sequence?