Efficiency Will Not Lead to Sustainable Reasoning AI

TL;DR

This paper argues that the unbounded compute demand of reasoning AI threatens sustainability because efficiency gains no longer guarantee energy reductions due to rebound effects and the absence of natural saturation. It reviews advances in structured prompting and reinforced reasoning that push reasoning capabilities with increasing compute, and analyzes the accelerating expansion of AI infrastructure. The authors advocate for measurement frameworks, governance mechanisms, and policy tools—such as caps, taxes, and compute budgets—and emphasize prioritizing earth-aligned applications to curb environmental impact. The work provides a roadmap for aligning progress in reasoning AI with planetary boundaries, informing researchers and policymakers about practical pathways for sustainable deployment.

Abstract

AI research is increasingly moving toward complex problem solving, where models are optimized not only for pattern recognition but for multi-step reasoning. Historically, computing's global energy footprint has been stabilized by sustained efficiency gains and natural saturation thresholds in demand. But as efficiency improvements are approaching physical limits, emerging reasoning AI lacks comparable saturation points: performance is no longer limited by the amount of available training data but continues to scale with exponential compute investments in both training and inference. This paper argues that efficiency alone will not lead to sustainable reasoning AI and discusses research and policy directions to embed explicit limits into the optimization and governance of such systems.