An AI data center's environmental and financial trajectory is largely fixed by a few early decisions, according to a new arXiv paper. The authors argue that choices such as connecting to the grid or building on-site gas generation, selecting evaporative or dry cooling, and deciding which county to file in determine the facility's capital costs, water draw, and carbon emissions for years to come.
The paper introduces a benchmark called "Watts If?" to evaluate whether large language models can reason about these counterfactual scenarios. Since LLMs are increasingly being asked to weigh in on infrastructure planning, the benchmark tests their ability to trace how discrete upstream choices cascade into downstream consequences.
The abstract does not report specific results, but the framing suggests the authors expect gaps in current model performance. The work points to a need for better evaluation tools as AI systems are consulted on high-stakes, long-horizon decisions like data center siting and energy sourcing.