New paper proposes regenerative AI infrastructure roadmap beyond Nvidia-centric stacks

A new academic paper accepted for the 2026 IEEE ICE/ITMC Conference takes a hard look at the environmental and social costs of today's AI scaling. Authors Han-Teng Liao and Karen Ang argue that current Generative AI infrastructure follows a linear "stack" model that prioritizes performance density while externalizing thermodynamic and material costs. They identify three major technology gaps: Scope 3 emissions (supply chain emissions), e-waste recycling, and social tensions from unsustainable scaling. The paper proposes a "Regenerative Socio-Technical Roadmap" that reframes AI infrastructure as a system-of-systems governed by planetary limits.

Central to their approach is a "metabolic circuit framework" that integrates the IEEE International Roadmap for Devices and Systems (IRDS) sustainability considerations for semiconductor facilities. The framework centers human values and needs within production-consumption loops, and explicitly critiques Nvidia-centric roadmaps as insufficient. The authors propose a competing reference architecture that enables a spontaneous order of resource parsimony and planetary accountability. This provides an actionable pathway for regulators and industry to build a truly circular digital economy.