Researchers Propose Regenerative Roadmap for AI Infrastructure
A new arXiv paper (2606.10544, submitted June 9, 2026, authors Han-Teng Liao and Karen Ang) argues that AI infrastructure planning needs a system-of-systems redesign built around planetary resource limits, not just compute performance. The paper, titled "From Stacks to Circuits," proposes a Regenerative Socio-Technical roadmap that folds a Sustainable Production and Consumption system map and IEEE IRDS semiconductor-facility sustainability metrics into AI lifecycle planning, and introduces a metabolic circuit framework centered on values and needs rather than throughput. For practitioners and infrastructure planners, the core claim is that current Nvidia-centric scaling roadmaps externalize thermodynamic and material costs, including Scope 3 emissions and e-waste, that a system-of-systems accounting would surface. The work is slated for presentation at the 2026 IEEE ICE/ITMC conference.
The practical takeaway for infrastructure and sustainability teams is not the paper's terminology but the accounting gap it points to: current AI scaling roadmaps measure compute density and cost per FLOP, but rarely account for Scope 3 emissions or semiconductor-facility material flows in the same framework, which makes it hard to compare the true resource footprint of competing infrastructure strategies.
What happened
Researchers Han-Teng Liao and Karen Ang posted "From Stacks to Circuits: A Regenerative Socio-Technical Roadmap for AI Infrastructure within Planetary Boundaries" to arXiv on June 9, 2026 (2606.10544), slated for the 2026 IEEE International Conference on Engineering, Technology, and Innovation (ICE/ITMC). The paper reframes AI infrastructure as a system-of-systems constrained by planetary boundaries rather than a linear supply-side "stack," repurposing a Sustainable Production and Consumption system map and integrating IEEE IRDS sustainability considerations for semiconductor facilities.
Technical context
The paper's central construct is a metabolic circuit framework that centers "Values and Needs" within production-consumption loops, connecting thermodynamic and material flows to governance and industrial-design decisions. It contrasts this with what it characterizes as Nvidia-centric roadmaps that prioritize performance density while externalizing costs such as Scope 3 emissions and e-waste.
For practitioners
Infrastructure planners and sustainability teams evaluating data center or chip-fabrication roadmaps can use the paper's framing as a checklist for what current lifecycle accounting tends to omit, particularly Scope 3 emissions, e-waste recycling paths, and semiconductor-facility resource metrics drawn from IEEE IRDS. It is a conceptual/academic contribution rather than a deployed standard, so it is best read as an input to roadmap and reporting discussions rather than an adopted framework.
What to watch
As a working paper heading into a 2026 engineering-management conference rather than a top-tier ML venue, its influence will show up, if at all, in whether standards bodies or corporate sustainability roadmaps begin referencing system-of-systems or metabolic-circuit language, or incorporate IEEE IRDS-style semiconductor facility metrics into AI lifecycle reporting.
Key Points
- 1A June 2026 arXiv paper proposes a system-of-systems framework tying AI infrastructure scaling to planetary resource limits.
- 2The paper argues current Nvidia-centric roadmaps externalize Scope 3 emissions and e-waste costs that lifecycle accounting should capture.
- 3As an academic conference paper rather than an adopted standard, its impact depends on future uptake by standards bodies or industry roadmaps.
Scoring Rationale
A verified, well-specified academic framework paper (confirmed via search: authors Han-Teng Liao and Karen Ang, arXiv 2606.10544) addressing a genuine gap in AI infrastructure sustainability accounting, but it is a working paper headed to an engineering-management conference rather than a top ML venue or industry standard, so impact is solid rather than notable pending any real-world uptake.
Sources
Public references used for this report.
Practice interview problems based on real data
1,625 SQL & Python problems across 15 industry datasets — the exact type of data you work with.
Try 250 free problems
