From Stacks to Circuits: A Regenerative Socio-Technical Roadmap for AI Infrastructure within Planetary Boundaries

Digital Circular Economy for Deep Tech AI Presentation

Han-Teng Liao & Karen Ang

2026-06-11

From Stacks to Circuits

A Regenerative Socio-Technical Roadmap for AI Infrastructure within Planetary Boundaries

• SS08-SJ-2C: Digital Circular Economy III, ICE 32nd Ed.

• Tuesday, 23/June/2026: 4:40pm - 6:00pm; Location: Room João

Han-Teng Liao 0000-0003-1081-5599
Independent Researcher

Karen Ang 0009-0008-5923-0106
Independent Researcher

Welcome everyone. My name is Han-Teng Liao. Allow me to start with a personal perspective on the impact of the generative AI on the environment.

🌱 Personal Background and Angle 1

pie title Taiwan First Carbon Fee Revenue 
    "Semiconductors" : 2200
    "Electricity Supply" : 635
    "Steel" : 400
    "Concrete" : 130
    "Others (Remaining Sectors)" : 1605

🌱 Personal Background and Angle 2

AI & Net-Zero Dual Transformation Forum

🧭 Section I: Introduction

The Unprecedented Scaling of Generative AI

• Linear Efficiency Focus: “Electrons to Tokens” optimization
• Performance Density: Prioritizing computational speed
• Externalized Liabilities: Material and thermodynamic costs
• Socio-Technical Tension: Disconnection from regional carrying capacities

Today we look at the rapid scaling of Generative AI. The current industry standard is a linear, performance-at-all-costs paradigm. This approach focuses strictly on optimizing from electrons to tokens, but it completely externalizes major physical and ecological liabilities.

Collision with Physical Realities

• Resource Strains: Intense energy and cooling water extraction
• The Grid Paradox: Local infrastructure application caps
• The Twin Transition: Merging digital innovation with material circularity
• Core Hypothesis: RST frameworks can reconcile compute growth with environmental boundaries

This intensive scaling creates a massive collision with physical reality. In places like Taiwan, regional power distribution grids are already capping data center power expansions due to severe supply imbalances. Our study explores how a Regenerative Socio-Technical framework can prevent these sudden infrastructure crashes.

📚 Section II: Literature Review

II.A Electrons to Tokens: Scaling Through Orchestration

TABLE I: The 5-Layer AI Production Stack (adapted from Nvidia [1])

Layer	Functional Domain	Industrial Role
L5	Applications	Enterprise agents, robotics, smart apps
L4	Models	Cognitive engines (LLMs)
L3	Infrastructure	AI / data centers (compute)
L2	Chips	GPU / TPU / NPU acceleration
L1	Energy	Power generation & cooling

• The Factory Model: Electricity as raw material, the token as final product — “electrons to tokens.”
• The Commoditization Trap: Energy treated as a static, infinite utility input, decoupled from regenerative limits.
• The Jevons Paradox: Per ITU-T L.1480, cheaper tokens trigger non-linear, System-of-Systems-level demand rebounds that outrun local grid adaptation.

Nvidia’s five-layer cake — Energy, Chips, Infrastructure, Models, Applications — is the dominant industry blueprint, treating the data center as a factory: electricity goes in as raw material, and the token comes out as the final product. The problem is that this linear “electrons to tokens” model treats energy as a static, commoditized input rather than a finite, regenerative resource. Viewed through a System-of-Systems lens and under ITU-T Recommendation L.1480, this narrow framing fails systemic auditing: as token unit costs fall, the resulting AI agent boom triggers Jevons Paradox rebound effects, expanding aggregate macro-energy demand faster than local power networks can structurally adapt.

II.E-F Semiconductor Facilities & The IEEE IRDS Lens

• IEEE IRDS ESSF Guidelines: Facility-level resource parameters
• Key Pillars: Water intensity, energy efficiency, and chemical safety
• SSbD Framework: Safe and Sustainable by Design electronics
• The Knowledge Gap: Merging real-time AI operation with static facility roadmaps

IRDS/ISRDS Workshop in Granada

To address this, we turn to the 2024 IEEE IRDS semiconductor facility roadmap. It details quantitative targets for water and chemical safety. However, a major knowledge gap persists: the industry lacks a way to bridge these static manufacturing parameters with live, operational AI workloads.

🛠️ Section III: Methodology

Repurposing Sustainable Production & Consumption (SPaC)

Fig 1: The “Atom-to-Values” System of Systems (SoS) loops for AI compute (adapted and expanded from the Sustainable Production and Consumption (SPaC) framework by J. Barber in [18]

• The Core Capital Loop: Extraction, Production, Distribution, Consumption, Investment
• Ecological Ceiling: Enforcing biophysical limits at the core
• Socio-Technical Governance: Centering “Values and Needs”
• System Dynamics: Interconnecting operational loops with environmental boundaries

Our methodology bridges this gap by repurposing the Sustainable Production and Consumption framework into a system dynamics model. This maps the lifecycle of compute into a Triple Bottom Line taxonomy, ensuring that the inner economic loops are structurally anchored by socio-technical governance and ecological limits.

🛠️ Section III: Methodology

Design Science & Case Sensitivity

• Design Science Research: Information systems architecture development
• Green Digital Deep Tech: Advanced engineering acting as systemic control loops
• Empirical Grounding: Real-world metrics from regional power dynamics
• Analytical Integration: Linking chip micro-telemetry to macro-systemic impacts

We apply design science research principles to build a circular reference architecture. By analyzing the structural limits of high-performance hardware, we transform sustainability requirements into real-world engineering constraints rather than treating them as standard, retroactive corporate social responsibility checklists.

📊 Section IV: Findings

IV.A Boundary Inputs for Metabolic Governance

TABLE II: Three-Tiered System Constraints

Layer	Primary Data Source	Homeostatic Threshold Example
Physical Infrastructure	IEEE IRDS ESSF Roadmaps	< 0.5 L/kWh compute; PFAS & emission controls
Operational Compute	Inference efficiency literature	Thermodynamic efficiency collapse at low concurrency; KV-cache memory sprawl
Policy & Governance	EU SSbD/CBAM; Magnifica Humanitas	5 MW grid ceiling north of Taoyuan; zero-exploitation supply chains

• Unified Boundary Matrix: Physical limits, chip telemetry, and policy mandates synthesized into one real-time governance baseline — not three disjointed layers.
• Operational Closure: These thresholds define the conditions under which the computing fabric must achieve metabolic balance.

Our first major result is the formal translation of multi-scalar telemetry into a single, unified boundary matrix for automated system regulation. Rather than treating physical infrastructure limits, chip-level telemetry, and international policy mandates as three disjointed operational layers, the RST architecture synthesizes them into one real-time baseline. Physical infrastructure constraints come from the IEEE IRDS ESSF roadmaps — for example, water intensity under half a liter per kilowatt-hour of compute. Operational compute constraints come from inference-efficiency literature, where thermodynamic efficiency collapses sharply at low concurrency. And policy constraints come from EU SSbD and CBAM regulation, alongside the Papal encyclical’s social mandates — concretely instantiated by Taipower’s strict five-megawatt grid allocation ceiling north of Taoyuan. Table II shows how these three tiers, taken together, define the homeostatic thresholds the RST architecture must operate within.

📊 Section IV: Findings

IV.B Upstream & Downstream Externalities

TABLE III: The 5-Layer Stack’s Externalized Liabilities

Layer	Functional Domain	Key Externalized Liability
L5	Applications	Uncapped token demand drives the Jevons Paradox rebound
L4	Models	Unmitigated HBM capacity sprawl from context scaling
L3	Infrastructure	Cooling water depletion & unbuffered grid volatility
L2	Chips	Embodied carbon of advanced lithography & e-waste
L1	Energy	Extractive reliance on municipal grids, delaying green transitions

• The Contrast: Table III is what the linear stack ignores; Table II is what the RST architecture instead enforces.
• The Reframe: Energy and materials shift from infinite exogenous inputs to localized, accountable boundaries.

Having established the boundary matrix in Table II, we now contrast it against the prevailing linear stack model, which focuses narrowly on unidirectional electrons-to-tokens conversion. Table III shows what that model overlooks at every layer. Applications drive uncapped token demand and the Jevons Paradox rebound. Models cause unmitigated high-bandwidth-memory sprawl as context windows grow. Infrastructure triggers cooling water depletion and unbuffered regional grid volatility. Chips carry embodied carbon from advanced lithography and short e-waste lifecycles. And energy remains an extractive draw on municipal grids that delays green digital transitions. Instead of treating energy and materials as infinite, exogenous inputs, the RST framework accounts for all of these as localized, socio-technical, and thermodynamic boundaries — directly governed by the thresholds we just saw in Table II.

📊 Section IV: Findings

The Thermodynamic Micro-Architecture Bottleneck

• The Memory Wall: Data transit bottlenecks between HBM and compute core
• Minimum Operational Batch Threshold (\(B_{min}\)): \[B_{min} \geq \frac{\alpha \cdot N}{L \cdot d_{model}}\]
• Arithmetic Underutilization: Sub-optimal batches draw peak baseline power
• Metabolic Entropy: Idle hardware expending energy on memory-bus transfers

The root of these environmental liabilities is actually found at the micro-architectural silicon layer. The memory wall forcing model weights between high-bandwidth memory and registers creates severe thermodynamic friction. Falling below the minimum operational batch threshold causes the hardware to waste high baseline power while idling.

📊 Section IV: Findings

The Context Horizon Memory Sprawl

• Key-Value Cache Expansion (\(M_{KV}\)): \[M_{KV} = 2 \cdot B \cdot L \cdot h \cdot d_{byte}\]
• Linear Biophysical Liability: Cache memory scales directly with context length
• Hardware Fragmentation: Sharding models across clusters just to pool memory
• Capacity Over-Provisioning: Low-concurrency regimes driving building booms

This issue is amplified by long context windows, where Key-Value cache memory demands scale linearly. To accommodate this memory sprawl, operators are sharding workloads across massive, multi-GPU clusters purely to pool high-bandwidth memory space, resulting in low arithmetic utilization and intense resource consumption.

📊 Section IV: Findings

Connecting the Micro-Bus to the Macro-Grid

• Structural Contradiction: Maximizing batch efficiency causes memory faults
• Down-batching Penalty: Preserving memory traps hardware in idling states
• The Cycle of CapEx: Building redundant facilities that inherit the same flaws
• The Ecosystem Ceiling: The immediate need for an automated metabolic governor

This creates an acute engineering contradiction. Maximizing batch efficiency causes hardware memory faults, while down-batching traps the system in a high-entropy idling state. This drives a cycle of massive capital expenditure to build redundant hyperscale facilities that inherit the exact same operational flaws.

📊 Section IV: Findings

IV.C Introducing the Systemic Volatility Indicator (SVI)

• The Automated Metabolic Governor: \[SVI = \max\left(0, \, \frac{B_{roof} - B}{B_{roof}} \cdot \gamma_{grid}\right)\]
• \(B\) vs. \(B_{roof}\): Instantaneous operating batch size against the hardware roofline threshold.
• \(\gamma_{grid}\): Localized grid utilization relative to institutional caps — concretely, Taipower’s 5 MW regional allocation ceiling north of Taoyuan.
• The Homeostatic Target: A cybernetic steering engine driving the indicator toward zero, not a static compliance score.

To resolve the structural contradiction between batch efficiency and memory safety, the architecture implements an automated metabolic governor: the Systemic Volatility Indicator, or SVI. It is bounded at zero to prevent negative indexing. The first term captures how far the instantaneous operating batch size falls below the hardware roofline threshold — that’s the micro-architectural side. The second term, gamma-grid, captures the localized power grid utilization ratio relative to institutional ceilings — concretely, this is Taipower’s strict five-megawatt regional allocation cap north of Taoyuan. The SVI acts as a cybernetic steering engine, not a one-off compliance score, dynamically bundling fragmented requests or shifting workloads to off-peak micro-grids whenever it drifts away from zero.

📊 Section IV: Findings

IV.D SVI Operational States

• State A — Maximum Volatility (\(SVI \to 1\)): Under-batched workloads (\(B \ll B_{roof}\)) run during peak regional grid stress (\(\gamma_{grid} \to 1\)); high eco-debt accumulates for low computational throughput.
• State B — Architectural Homeostasis (\(SVI \to 0\)): Full batch saturation (\(B \to B_{roof}\)) or off-peak execution; computation is safely buffered by the biophysical capacity of the ecosystem.
• The Lesson: Local latency optimization, pursued in isolation, can quietly enforce an aggregate energy penalty across the macro-grid.

The SVI formalizes two operational states. State A is maximum volatility: under-batched workloads run while regional grid stress is already at its peak, so the indicator approaches one and signals heavy eco-debt accumulation for very little computational throughput. State B is architectural homeostasis: the system reaches full batch saturation, or it shifts non-latent work to off-peak windows, and the indicator resolves to zero — computation is then safely buffered by the biophysical capacity of the surrounding ecosystem. The broader lesson here is that local optimization for client-side latency, pursued in isolation, quietly enforces an aggregate energy penalty across the macro-grid — which is exactly the blind spot the linear electrons-to-tokens model cannot see.

💬 Section V: Discussion

Restoring the Atom-to-Values Closed Loop

• Reframing the Paradigm: Shifting from extractive to regenerative governance
• Internalizing Entropy: Treating waste heat, water, and carbon as liabilities
• From CSR to RegTech: Real-time engineering metrics replacing static reports
• Systemic Realignment: Forcing tokenomics to respect finite planetary boundaries

Our reference architecture directly answers our primary research question by shifting the AI infrastructure paradigm from an extractive linear stack to an atom-to-values closed loop. Rather than treating waste heat and thermal entropy as ignored externalities, they are tracked as active balance-sheet liabilities.

💬 Section V: Discussion

Digital Product Passports (DPP) & Material Circularity

• Tracking Lifecycles: Managing the rapid obsolescence of advanced accelerators
• Reverse Supply Chains: Auditing e-waste streams and material recovery
• Regulatory Compliance: Automating adherence to EU CSDDD and CBAM mandates
• Operational Integration: Linking physical hardware telemetry with policy constraints

By deploying Digital Product Passports and regulatory technology, the reference architecture tracks the entire lifecycle of high-performance components. This allows operators to audit e-waste streams and automate compliance with strict international environmental laws like the EU’s Corporate Sustainability Due Diligence Directive.

💬 Section V: Discussion

Value Creation Centered on Human Needs

• The Core Anchor: Placing “Values and Needs” at the center of production loops
• Socio-Ethical Guidance: Adhering to standards of human dignity
• Supply Chain Auditing: Tracking data labeling, mineral extraction, and labor
• Resource Sovereignty: Protecting regional civil infrastructure from digital strain

Crucially, our framework centers value creation around real human needs. This allows socio-ethical specifications, such as institutional human rights frameworks, to guide compute deployment. By doing so, it protects regional communities from labor exploitation and defends public civil resource sovereignty.

💬 Section V: Discussion

Engineering Management Checklist

• Grid Integration: Respect regional grid caps and power community boundaries
• Thermodynamic Recovery: Use IRDS facility KPIs to route waste heat and water
• Lifecycle Circularity: Deploy DPPs to track hardware through reverse supply chains
• Asset Amortization: Extend operational lifetimes to balance embodied carbon

For technology strategists and engineering managers, we translate these insights into a highly practical compliance checklist. Managers must actively align their workloads with local energy community capacities, structurally route facility waste heat, and utilize digital passports to extend accelerator lifecycles.

🏁 Section VI: Conclusion

Framework Limitations & TRL Status

• Conceptual Stage (TRL 1–2): Pending formal adoption by official IEEE roadmapping bodies
• Micro-Telemetry Access: Needs deeper integration with proprietary hardware APIs
• Validation Requirements: Large-scale distributed pilots required to confirm data integrity
• Future Path: Bridging the gap between conceptual reference models and ground-truth physics

We acknowledge certain limitations in our current work. The framework sits at an early technology readiness level and needs formal integration into standard global industry roadmaps. Future testing must include extensive field validation to ensure metadata integrity across real-world, global supply chains.

🏁 Section VI: Conclusion

Summary of Main Contributions

• System-of-Systems View: Re-engineered AI infrastructure as an eco-bound system
• SVI Deployment: Provided a quantified mathematical governor for automated routing
• Internalized Debt: Formalized ecological footprints as active software constraints
• Paradigmatic Shift: Moved deep tech thinking from raw compute density to resource parsimony

In conclusion, this study contributes a system-of-systems reference architecture that tethers token generation to planetary boundaries. By introducing the Systemic Volatility Indicator, we prove that ecological debt can be successfully managed as an active software orchestration constraint.

🏁 Section VI: Conclusion

Real-World Friction: The Taiwan Precedent

• The Reality Check: Taipower’s 2026 data center grid application freezes
• The Structural Collision: Sovereign AI compute demands meeting binding energy limits
• The Lesson: Technical efficiency gains cannot bypass biophysical constraints
• The Imperative: Standardizing circular governance before hitting ecological walls

The real-world urgency of this research is highly visible. Right now, utilities are enforcing strict energy ceilings on new data centers due to grid imbalances. This confirms that technical efficiency gains cannot escape local biophysical limits, highlighting the urgent need for standardized circular governance.

🏁 Section VI: Conclusion

Strategic Roadmap for Future Alliances

• IEEE Integration: Linking HIR manufacturing tracks with systemic environmental accounting
• Policy Expansion: Advancing smart city frameworks into smart intelligence networks
• Final Vision: Deep technologies demand deep systems thinking anchored in planetary realities

Our future work calls for an integrated roadmap across global standards bodies, expanding smart city frameworks into smart intelligence networks. Ultimately, advanced digital technologies require deep systems thinking. To make generative AI truly sustainable, we must govern it within the finite limits of our planet. Thank you.

🙏 Thank You for Your Attention!

From Stacks to Circuits

A Regenerative Socio-Technical Roadmap for AI Infrastructure within Planetary Boundaries

• RS-MI-2B: AI, Generative AI & Digital Transformation, ICE 32nd Ed.

• Tuesday, 23/June/2026: 3:00pm - 4:20pm; Location: Room Miragaia

Han-Teng Liao 0000-0003-1081-5599
Independent Researcher

Karen Ang 0009-0008-5923-0106
Independent Researcher / Infineon