Efficiency Will Not Lead to Sustainable Reasoning AI
This addresses the sustainability problem for AI researchers and policymakers by highlighting the need for explicit limits in optimization and governance, and it is incremental in building on existing concerns about AI's environmental impact.
The paper argues that efficiency improvements alone cannot ensure the sustainability of reasoning AI, as performance scales exponentially with compute investments in training and inference, unlike historical trends in computing where efficiency gains stabilized energy use.
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.