4

Module 4  ·  Track 1: Technical Foundation

Impact Domains: Where the Levers Actually Sit

Different domains have different dominant drivers. GreenOps requires domain fluency.

Duration20–25 minutes
TrackTechnical Foundation

What you will take from this module

The estate is a system, not a list of categories.

Each domain in an IT estate has its own dominant driver. End-user devices are dominated by embodied carbon. Data centres by energy and cooling overhead. Cloud by utilisation and region. Software by runtime efficiency and data volume. AI by model selection and inference volume. Hardware by lifecycle and circularity. Networks by data movement patterns.

The interactions are where waste compounds. A badly designed application running on over-provisioned cloud infrastructure in a high-carbon region, accessed through devices replaced on a rigid three-year cycle, generates waste across four domains at once. Single-domain fixes miss those interactions.

Domain fluency means two things: the specific levers in each domain, and how decisions in one domain create or remove waste in another. The explorer below covers both.

The measurement question to carry through every domain

For each domain: what data do you have at Tier 1 (measured), what are you modelling at Tier 2, and what is still Tier 3 (proxy)? The data quality hierarchy from Module 3 applies here, domain by domain. If you cannot answer that question per domain, your baseline has gaps you cannot see.

The seven domains of operational impact.

Select each domain to explore the levers, waste patterns, and metrics that matter. Different domains need different governance approaches.

End-User Services

The core insight: The footprint of end-user services is dominated by embodied impact. A laptop's manufacturing footprint (energy, materials, mining, fabrication, logistics) can outweigh its operational electricity across the entire working life. The biggest levers are lifecycle and procurement decisions, not monitor-switching campaigns.

Key waste patterns

Death by a thousand defaults. Extra devices issued by course. Peripheral sprawl: two monitors regardless of need. Docking stations accumulating in storage rooms. Laptops air-freighted because onboarding was planned at the last minute. Each decision is small. Across an estate, they add up to material footprint.

The device refresh trap: Three-year mandatory cycle applied uniformly. Evidence-led approach replaces this with role-based and security-justified timing.
Virtualisation shift: Browser-and-document-based roles move to thin clients on shared infrastructure. Device becomes lightweight terminal; embodied impact shifts to shared, longer-lived infrastructure.
Recovery friction: If returning kit is harder than ordering new, replacement will always win. Frictionless recovery process is the only governance that sticks.

Operational Levers

  • Device standards by role: accounting for power draw, repairability, and expected lifespan, not just performance
  • Power management defaults enforced centrally rather than left to individuals
  • Recovery and redeploy processes that are easier than ordering replacement devices
  • Accessory management treating cables, chargers, and docking stations as assets with lifecycle, not consumables
  • Device count governance: how many people have both laptop and desktop when one would do?
  • Virtual desktop infrastructure for role-based workloads, shifting to well-utilised shared servers
  • Explicit AI usage governance with model selection and inference controls

Data Centres

The core insight: Data centre footprint is concentrated and measurable. This is also where poor measurement creates credibility failures. PUE matters, but PUE alone is dangerous. A lightly loaded data centre can show respectable PUE and still be wasteful in absolute terms.

Core metrics and their meaning

PUE
Poor ↓ Good
Server Utilisation
Low ← → High
Infrastructure Efficient, Compute Wasteful
Good PUE + Low Util
Cooling efficiently, but servers are idle. Hidden waste pattern: capacity sitting cold but empty.
Example: PUE 1.15, utilisation 8%
Efficient
Good PUE + High Util
Target state. Facility cooling is efficient and equipment is doing productive work at scale.
Example: PUE 1.15, utilisation 75%
Worst Case
Poor PUE + Low Util
Both facility and equipment inefficient. Urgent action required. Cooling overhead is high with minimal productive work.
Example: PUE 2.0, utilisation 12%
Compute Efficient, Facility Wasteful
Poor PUE + High Util
Equipment productive and well-used. Infrastructure overhead is high. Targeted cooling improvements needed.
Example: PUE 1.8, utilisation 80%

Metrics Beyond PUE

The measurement maturity chain
  • PUE: Total facility power ÷ IT equipment power. Context: ~1.1 is exceptional; most enterprise facilities materially higher.
  • Utilisation: Actual server load as percentage of capacity. Follow-up question to any PUE claim: what is IT utilisation and the consolidation plan?
  • WUE (Water Usage Effectiveness): In water-stressed regions, non-negotiable. Links cooling approach to water dependency and resilience risk.
  • CUE (Carbon Usage Effectiveness): Operational energy to carbon. Bridge metric. Moves with facility change, energy sourcing, or grid intensity.
  • REF & ERF: Renewable Energy Factor and Energy Reuse Factor. Only credible with stated definitions, metering coverage, and evidence.
The metric theatre problem: Most organisations fail not from lacking metrics, but from using them as theatre. Design PUE presented as operational reality. Annual averages hiding seasonal collapse. Portfolio averages masking worst sites. No hall-level visibility. Renewable claims without boundary definitions. Test every metric with: can this team explain the calculation, boundary, and instruments used?

Cloud and Platform Operations

The core insight: Cloud is not inherently more sustainable. Hyperscalers achieve better utilisation than most enterprise data centres, but cloud sustainability is not automatic. The decisions you make about how you use cloud determine whether efficiency benefits are realised or whether waste scales faster.

The biggest waste category

Idle or over-provisioned compute. Instances running continuously with low utilisation. Resources provisioned generously and never right-sized. Non-production environments left running. Storage accumulating without retention policy. Resilience provisioned at most expensive defaults because defaults were never challenged.

The continuous control loop: Five-part discipline

1

Visibility and Attribution

If you cannot attribute consumption, you cannot govern it. Every material resource needs an owner, environment label, and service identifier. Tagging is not housekeeping. It is accountability. Untagged resources are non-compliant by definition. Nobody is responsible; nothing gets optimised.
2

Baselining Unit Economics

Total cost and consumption figures suit annual reporting. Operations need unit measures: carbon per VM-hour, idle percentage, utilisation distributions, storage growth by service. For top services, carbon per transaction or API call. That is where the conversation becomes actionable.
3

Remediation Hierarchy

Start low-risk: enforce time-to-live on non-production, kill zombie resources, delete orphaned storage, rightsize based on observed utilisation. Move to structural levers: consolidate platforms, revisit availability patterns to match actual requirements, get serious about Kubernetes resource requests, limits, bin packing, and node lifecycle.
4

Guardrails for Prevention

If you fix waste once and it returns next month, you have done clean-up, not improvement. Guardrails prevent recurrence: policy-as-code for tagging and ownership, default sizing limits with clear exception routes, mandatory termination for non-production, conservative retention defaults unless justified, platform templates encoding efficient defaults.
5

Verification and Institutionalisation

Verify savings not only in cost but in consumption drivers: is idle percentage falling? Is carbon per VM-hour improving? For key services, is carbon per transaction moving right? Bake the loop into governance: FinOps reviews, platform reliability reviews, architecture review boards. GreenOps becomes normal operating governance.

Region selection as a lever

Cloud providers operate in regions with substantially different grid carbon intensity. Equivalent workloads in renewable-powered regions versus fossil-heavy grids can mean a tenfold difference. For flexible workloads (no strict data residency, no sub-second latency constraint), region selection is a real, practical GreenOps lever.

Software and Data

The core insight: Software behaviour is one of the most underestimated sustainability levers. How software is designed determines computational work, data generation, retention, and unnecessary movement. These decisions compound over a service's lifetime.

The Software Carbon Intensity model

Carbon is driven by energy, which is driven by computational work, multiplied by grid carbon intensity. Reduce compute per transaction; reduce energy per transaction; reduce carbon per transaction. The SCI formula prevents the efficiency-growth trap: a service can emit more in total while improving in intensity. Governance requires understanding which direction is moving.

Dark data and data minimisation

Dark data (stored, protected, backed up, replicated, but never accessed) is common and easily remediable. It accumulates because default retention is permissive, because storage was historically cheap, and because no process questions whether data still serves a purpose. Start with data minimisation: collect only what you have clear use for, define retention at creation, treat storage growth as a performance metric.

Telemetry excess: Logs retained at maximum resolution for maximum retention period regardless of operational need. Metrics streamed continuously when periodic sampling would serve the purpose. At enterprise scale, material storage and processing with no business value.
Retention by default: Databases replicated to dev with production data; never cleaned. S3 buckets from deprecated services still paying monthly fees. Simple process change (retention rules and automated cleanup) recovers material waste.

Impactful Software Sustainability Decisions

Key considerations
  • Computation required to perform the task (optimise algorithms; avoid unnecessary processing)
  • Data generation, retention, and replication (dark data is one of the fastest-remediable waste categories)
  • Background process frequency (is it needed at that cadence?)
  • Data movement between services and regions (intentional or unconsidered architecture?)
  • Telemetry and logging intensity (most organisations collect more than they actively use)

Hardware Lifecycle and Circular Economy

The core insight: Embodied carbon in hardware is substantial. For laptops, it typically represents the majority of lifetime footprint. For servers, the balance depends on utilisation and grid intensity, but embodied impact is always material. Procurement and end-of-life decisions matter as much as operational efficiency.

Avoiding trade-off traps

Common mistakes to avoid
  • The extension trap: Extending asset life is often right. But if older hardware is dramatically less energy-efficient for the workload, keeping it longer increases operational energy. Decision depends on workload, utilisation, and grid intensity.
  • The premature refresh trap: Refreshing early in the name of efficiency without accounting for embodied impact increases e-waste and lifecycle carbon.
  • The comparison trap: A server with higher product footprint might be more efficient per unit of delivered compute if it does the same work with fewer machines. Compare on functional basis relevant to workload, not headline numbers.

The full lifecycle chain

Selection
Deployment
Use
Redeploy
Repair
Refurbish
End-of-Life
Selection

Procurement driven by performance and price with sustainability as afterthought. Wrong approach. Include: full lifecycle footprint, repairability score, spare parts availability and support for expected lifespan, manufacturer take-back programme and evidence of actual refurbishment. Supply chain risk: rare earth minerals, conflict minerals, labour standards (Responsible Business Alliance Code of Conduct).

Deployment

When and how devices reach users. Air-freighted because onboarding was left to the last minute amplifies embodied impact. Logistics planning is a sustainability decision.

Use

Operational energy draw and utilisation. For laptops, typically smaller than embodied impact. For servers, balance depends on grid intensity and usage patterns.

Redeploy

If device is viable, move to lower-demand user tier before retirement. Requires real asset tracking and a process easier than ordering new. Otherwise intention will not survive reality.

Repair

Treat repairability as procurement requirement, not aspiration. Spare parts, repair SLAs, battery replacement, upgrade paths. If device cannot be repaired economically, refresh cycle is forced regardless of policy position.

Refurbish

Extend life through certified refurbishment. Only credible with evidence and standards compliance.

End-of-Life

ITAD is an evidence-generating process, not a disposal contract. Chain-of-custody documentation, verified treatment routes, data sanitisation certificates to recognised standards, reporting supporting audit and ESG disclosure. If you cannot trace what happened, you cannot credibly claim sustainability.

AI as a Cross-Cutting Multiplier

The core insight: AI is not a separate sustainability topic. It is a multiplier across everything else. The main risk for most organisations is not headline-grabbing training runs. It is inference demand scaling silently across hundreds or thousands of users, embedded in products, left ungoverned because it feels like productivity.

Key governance areas
  • Model selection: Substantial energy difference between small, fine-tuned model and general-purpose frontier model. Right-size the model to the task. Use the smallest model that reliably achieves the outcome.
  • Prompt and interaction design: Every token sent and received drives compute. Longer prompts, longer outputs, extended context windows increase inference load. Intentional use, not habitual.
  • Caching and architecture: If same queries are made repeatedly, cache responses. If AI is embedded in a product, minimise unnecessary model calls. Pre-screen whether query genuinely requires inference.
  • Unit economics accountability: Product owners accountable for AI unit economics: value per inference, cost per inference, guardrails preventing unnecessary model calls.

Stop-doing and start-doing lists

Stop Doing (Waste Prevention)

Stop using biggest model by default
If smaller model is fit for purpose, use it. Make that the default, not the exception. Most routine tasks do not require frontier-class inference.
Stop stuffing context with full documents
Extract relevant excerpt. If you cannot explain why full document is needed, it probably is not. Context window overhead is real compute cost.
Stop letting chat threads grow indefinitely
Long-running threads carry significant context overhead. Start fresh when context is no longer relevant. Reset discipline saves compute and clarifies thinking.
Stop vague questions, then iterate five times
Be specific. Constrain output. Tell model what good looks like from start. Iterative refinement is expensive; clear framing is efficient.
Stop generating outputs you will not use
Want bullet summary? Ask for bullet summary. Want short answer? Enforce length limit. Long outputs are compute waste if they are discarded.
Stop using AI where search or templates work
AI is a tool for specific use cases, not default replacement for thinking. Search queries and rules-based workflows are cheaper and faster.
Stop running non-urgent jobs in peak carbon periods
If it can run later, schedule it later. Batching AI workloads into lower-carbon periods reduces absolute footprint without reducing capability.
Stop treating AI consumption as ungoverned
If you cannot see usage, you cannot manage it. Track calls, tokens, and cost by team and use case. Visibility is the foundation of governance.

Start Doing (Governance)

Start right-sizing models by use case
Simple tiering: small for routine tasks, mid-range for reasoning, frontier-class only where it changes outcome. Default to smallest model that works reliably.
Start measuring AI consumption properly
At minimum: calls per user, tokens per request, cost per call. If embedded in service, track tokens per transaction so you see whether intensity is drifting.
Start designing prompts that enable efficient behaviour
Use templates. Default to short outputs. Make it easy to provide only relevant excerpt rather than paste whole documents. Efficient defaults win over education.
Start batching and caching wherever sensible
Repeatedly asked questions should not trigger repeated inference every time. Cache stable answers and standard outputs. Batch back-office workloads where latency is not critical.
Start treating non-urgent workloads as flexible load
Embeddings, classification, summarisation pipelines, evaluation, reporting. Schedule for lower-carbon periods. Flexibility reduces absolute footprint.
Start putting lightweight user training in place
Not a lecture. Short guide: be specific, set constraints, avoid context stuffing, avoid long outputs, restart threads when context no longer relevant. Reinforce through interface.
Start making product owners accountable for unit economics
If AI is a feature, owner should explain value per unit and cost per unit, not just say users like it. Accountability drives rightsizing discipline.
Start building guardrails early, while controllable
Once behaviour is entrenched, much harder to unwind. Policy-as-code, usage limits, cost alerts, and model defaults prevent silent scale.

Networks

The core insight: Networks are invisible infrastructure with frequently omitted sustainability impact. Data is hard to access; control points are indirect. That does not make them unimportant. It makes them easy to avoid. The practical answer is that egress cost and egress footprint move together.

The data movement footprint

Every API call transferring data unnecessarily. Every cross-region synchronisation serving no operational purpose. Every video stream at resolution higher than the connection will display. Every chatty microservice architecture generating constant internal traffic. These aggregate into real network energy consumption at scale.

Egress as a sustainability signal: If cloud egress charges are significant, data is moving in patterns that were not intentionally designed. Egress cost and egress footprint move together. One of the areas where cost and carbon genuinely align without requiring additional justification.

Practical Levers for Network Sustainability

Key interventions
  • Design systems that minimise unnecessary data movement (architecture discipline)
  • Reduce cross-region replication to what is operationally justified, not default paranoia
  • Cache at appropriate layer to avoid repeated data transfers for static or slowly changing content
  • Compress payloads where bandwidth is the constraint
  • Review video conferencing policies where high-resolution streaming is default regardless of context
  • Monitor egress costs as signal of unintentional data movement patterns

End-user device footprint

What typically represents the majority of a laptop's lifetime environmental footprint?
A) Operational electricity consumption during the device's 5-year lifespan
B) Manufacturing, mining, fabrication, and logistics. The embodied carbon before it reaches the user's desk
C) End-of-life recycling and disposal processes
D) Network and cloud infrastructure used to provision device software

PUE and utilisation

A data centre reports a PUE of 1.15 (which is good). Server utilisation is 8%. What does this tell you?
A) The data centre is highly efficient and requires no improvement
B) The facility is cooling efficiently, but the IT equipment being cooled is not producing much useful work. A hidden waste pattern
C) The utilisation is irrelevant; PUE is the only metric that matters
D) Server consolidation and decommissioning should be the immediate priority

Cloud control loop

Your cloud platform team discovers idle instances and right-sizes them, reducing consumption. One month later, idle instances reappear. What is the missing step in the control loop?
A) More aggressive right-sizing recommendations
B) Guardrails: policy-as-code for default sizing limits, mandatory termination for non-production, platform templates encoding efficient defaults
C) Switching to a different cloud provider with lower costs
D) Increasing the frequency of manual clean-up reviews

AI model selection

An organisation is using frontier-class AI models for routine email composition, short summaries, and routine document formatting. The best GreenOps intervention is:
A) Restrict AI use entirely for cost reasons
B) Right-size the model tier by use case: small models for routine tasks, mid-range for reasoning, frontier-class only where it changes the outcome
C) Increase the context window to allow more information in a single query
D) Monitor token usage but allow current model selection to continue

⏸ Pause & Reflect

Which domain in your estate has the most obvious governance gap? Where could your organisation pull a single lever that would move multiple metrics at once? That is where GreenOps begins.

Module 4: Key Takeaways

Embodied impact dominates end-user devices.

Procurement and lifecycle decisions matter more than switch-off campaigns. Recovery and redeploy processes must be easier than ordering new.

PUE without utilisation is incomplete.

Good facility efficiency with low server utilisation is hidden waste. Follow PUE with the utilisation question and a consolidation plan.

Cloud requires continuous control, not one-off projects.

Visibility, baselining, remediation, guardrails, verification. Fix waste once, then prevent recurrence with policy-as-code and platform defaults.

Software and network decisions compound at scale.

Compute per transaction, data retention, egress patterns, and network movement aggregate into material footprints. Data minimisation and architecture discipline pay off across domains.

Hardware lifecycle is seven stages, not one procurement decision.

Selection, deployment, use, redeploy, repair, refurbish, end-of-life. Circular economy requires governance across all seven, not just the first.

AI governance must start before adoption scales.

Model tiering, caching, prompt efficiency, and usage accountability prevent inference from becoming the fastest-growing ungoverned cost in the estate.

This completes Track 1: Technical Foundation. You now have the operational reality (Module 1), a clean definition of digital sustainability (Module 2), the measurement vocabulary (Module 3), and domain fluency across the estate (Module 4).

Track 2: Operations starts in Module 5 and turns from understanding to doing. Module 5 covers how reporting translates into operational practice, including the rebound effect, demand management, and the difference between efficiency gains that stick and those that get consumed by growth. Module 6 applies these ideas to software. Module 7 to AI. Module 8 to supply chain and procurement.

The foundation is set. From here, the course builds operational discipline on top of it.

← PreviousM3: How Digital Sustainability Is Measured