Data center siting is locked in for years: one multi-objective RL agent lets planners retune water, power, carbon and reliability at inference time
Synopsis
AI data center siting, power procurement and cooling are still settled once by static single-point scoring on day-zero numbers, though water, power and carbon drift over a 10- to 20-year asset life and commitments bind for years. The build-out becomes a tiered Markov decision process: strategic commitments occur only in discrete windows, carry switching costs and take effect after source-dependent lead times, dispatch uses only capacity that exists, and a preference-conditioned multi-objective agent returns a policy family steerable at inference time across cost, carbon, water and reliability. On real site data, preference conditioning gives a confirmed, steerable trade-off on both objective pairs across three commitment regimes, and lock-in covers the frontier more evenly.
Interpretation
A planner no longer has to accept the single score computed at day zero: one family of policies can be pushed along cost, carbon, water and reliability at inference time, spreading the candidates into a comparable trade-off curve. Siting, power procurement and cooling have been compared with static single-point scoring tools, evaluated once on day-zero numbers, while a site's water, power and carbon profile drifts over a 10- to 20-year asset life and the commitments are hard to undo. On real site data, imposing lock-in gives a more evenly covered frontier and a more consistent rate of anticipatory retrofit ahead of a binding water-stress constraint than the freely reversible formulation.
The decision is split into two layers: strategic commitments that occur only in discrete windows, carry switching costs and energize only after source-dependent lead times, and operational dispatch that can use only capacity that physically exists. Solving it uses a preference-conditioned multi-objective agent built as a semi-Markov manager and worker hierarchy, meaning decisions hold for variable durations and are split between an upper and a lower level; one solve returns a family of policies that a planner selects from with preference settings at inference time. Evaluation spans the irreversible, reversible and commit-once commitment regimes on real interconnection-queue, water-stress and grid-emissions data, and the environment, agents, baselines and evaluation harness are released.
Perspective
The demonstrated scope is AI data center siting, power procurement and cooling, using real interconnection-queue, water-stress and grid-emissions data, tested under irreversible, reversible and commit-once commitment regimes with a 10- to 20-year asset life; the environment, agents, baselines and evaluation harness are released.
Long-horizon forecasting on the water-stress axis is the stated weak point: the compact counterfactual dynamics model follows slow-moving state well but is measurably outperformed by simple persistence there, a limitation the authors set out. This report rests on the summary, so method details and the full set of experimental numbers were not checked item by item.