// Scrape Me · Open research agenda v1.0

Research hypothesis Not a settled conclusion

Reducing AI’s Energy Footprint

— Gauraang Malik

Intelligence has a physical footprint. Could suitable compute-heavy AI workloads move to managed building or campus infrastructure, where electricity and heat can be measured, controlled, rejected, or reused—while phones and laptops retain the edge processing they genuinely need?

01 · Research hypothesis

Put heavy compute where its physical consequences can be managed.

The proposal is not “centralize everything.” It is to test a hybrid architecture: shared building- or campus-scale systems handle suitable high-intensity workloads, while end-user devices retain processing that protects privacy, minimizes latency, supports offline use, and preserves resilience.

  1. 01

    Lightweight interfaces

    Phones, laptops, consoles, sensors, and controls handle local interaction and task-appropriate edge inference.

  2. 02

    Shared AI infrastructure

    Managed GPUs serve eligible workloads with scheduling, access control, measurement, and higher potential utilization.

  3. 03

    Building-energy integration

    Electrical demand, cooling, heat rejection, and possible heat recovery become explicit parts of system design.

The scientific question is comparative: for which workloads, operating conditions, climates, and user requirements does this hybrid arrangement outperform cloud-only, campus-only, or device-only alternatives?

02 · Evidence boundary

What is defensible—and what remains unresolved.

Technically reasonable

  • AI computation consumes electricity and ultimately produces heat that must be managed.
  • Concentrated infrastructure can make electrical demand, utilization, cooling, and thermal output easier to meter.
  • Shared accelerators may achieve higher utilization than individually owned high-power devices when scheduling and demand are well matched.
  • Server heat can sometimes support useful heating, but usually only after careful integration and, in many cases, temperature upgrading.
  • Flexible computing loads may participate in building- or grid-level demand management when workload deadlines permit.

Not established in advance

  • Centralized or campus computing is not automatically greener than efficient edge hardware or a well-operated cloud service.
  • Useful heat recovery is not guaranteed; source temperature, distance, seasonal timing, heat demand, and integration cost all matter.
  • Network energy, latency, service outages, privacy requirements, and rebound effects can change the preferred architecture.
  • Removing powerful processors from all personal devices would be too extreme and would reduce resilience and user control.
  • Larger models are not justified by scale alone; measured gains must be weighed against energy, material, water, and infrastructure costs.

03 · Architecture comparison

The answer is likely workload-specific.

A fair comparison must deliver the same validated result—not merely run the same model name. System boundaries should include the device, server, networking, cooling, and the electricity mix used during the task.

Questions to test across building or campus, device or edge, and hybrid AI systems.
Dimension Building or campus Device or edge Hybrid decision rule
Energy Can share efficient accelerators, but adds facility and network demand. Avoids remote transfer, but duplicated hardware may sit underused. Compare wall-plug energy per equivalent, validated task.
Cooling and heat Heat is concentrated and measurable; recovery depends on temperature and demand. Heat is dispersed into occupied spaces and usually not recovered. Route compute using thermal conditions and useful-heat opportunity.
Utilization Scheduling may improve use of expensive hardware. Capacity is immediately available but often idle. Retain minimum local capacity; pool burst and specialized workloads.
Privacy Institutional control can protect data, but central stores increase governance responsibility. Data may remain on the device. Keep sensitive preprocessing local and transmit only authorized inputs.
Latency Depends on network quality, queueing, and service location. Best for immediate control and offline response. Use edge compute for hard real-time needs; offload delay-tolerant work.
Resilience Can be managed and backed up, but creates a shared dependency. Continues during network or server failure within local limits. Design graceful degradation rather than one mandatory compute location.
Lifecycle Shared hardware may reduce duplication but concentrates upgrade and facility impacts. Distributed replacement cycles can increase embodied impacts. Measure service life, repairability, utilization, and avoided hardware together.
Governance Supports managed access, auditability, and public-interest allocation. Preserves personal ownership and autonomy. Make workload routing, access, retention, and override rules transparent.

04 · Research program

How a building-energy lab could test the idea.

  1. Phase 1

    Define equivalent services

    Select representative AI, simulation, and analysis tasks with fixed inputs, quality thresholds, deadlines, privacy classes, and user expectations.

  2. Phase 2

    Instrument every architecture

    Measure devices, network equipment, server input power, cooling auxiliaries, queueing, throughput, and response quality under matched workloads.

  3. Phase 3

    Map the thermal pathway

    Record air or liquid temperatures, heat flow, cooling load, and the temperature lift required to supply space heating or domestic hot water.

  4. Phase 4

    Model buildings and seasons

    Test climate, occupancy, tariff, grid-carbon, heat-demand, storage, and equipment scenarios instead of extrapolating from one room or one day.

  5. Phase 5

    Publish decision boundaries

    Report when campus, edge, cloud, or hybrid operation performs best, including uncertainty, negative results, and conditions that invalidate heat recovery.

05 · Measurements

Measure the service, the building, and the lifecycle.

Wall-plug energy
kWh per completed, quality-validated task.
Facility overhead
Auxiliary kWh and power usage effectiveness, with system boundaries stated.
Thermal output
Recoverable and rejected heat in kWhth.
Source temperature
Air or coolant supply and return temperatures over time.
Cooling load
Peak and annual cooling demand attributable to compute.
Heat-pump performance
Coefficient of performance at the required temperature lift.
Useful-heat fraction
Recovered heat actually delivered to a matched demand.
Seasonal match
Coincidence of compute heat with building or district heat demand.
Utilization and concurrency
Active accelerator time, queue time, and simultaneous users.
User performance
Latency, reliability, task success, output quality, and satisfaction.
Lifecycle impact
Embodied carbon, service life, repairability, and avoided device hardware.
Cost per service
Capital and operating cost per validated workload, with sensitivity ranges.

06 · Research use cases

A shared system must serve real work, not specifications.

Buildings, HVAC, and EnergyPlus

Simulation, optimization, weather processing, physics–AI experiments, reproducible sweeps, and interruption-safe research workflows.

SHAP and clustering analysis

Numeric embedding evaluation, neighbour comparison, cluster agreement, confidence analysis, and post-hoc visualization from authorized artifacts.

Secure research memory

Access-controlled retrieval over current project documents, decisions, code, papers, and results without placing private material in general model weights.

Project-specific coding agents

Sanitized training examples, local inference, adapter experiments, automated checks, and sealed evaluation sets with clear provenance.

Local model evaluation

Reproducible comparisons of model quality, latency, energy, concurrency, and failure behavior under institutional control.

Dashboards and authorized documents

Source-backed analytical products and permitted document workflows where privacy, provenance, and reproducibility are explicit requirements.

07 · Gauraang values

Infrastructure should express responsibility.

These are design values, not empirical findings. They define how the research should be conducted and how competing outcomes should be judged.

Stewardship

Treat energy, materials, water, attention, and public funds as finite resources.

Public benefit

Prefer shared capability and transparent access over status-driven accumulation.

Safety and dignity

Reduce harm without treating people as merely data sources or compute demand.

Privacy and agency

Preserve local control where privacy, latency, resilience, or personal autonomy requires it.

Responsible openness

Share knowledge when possible while respecting security, consent, confidentiality, and misuse risks.

Useful intelligence

Demand measurable public or scientific value before assuming a larger model is progress.

08 · Evidence spine

Starting sources, not final proof.

  1. International Energy Agency · 2025

    Energy and AI

    A global evidence base for data-centre electricity demand, energy supply, emissions, efficiency, and policy trade-offs.

  2. Microsoft Research · 2011

    The Data Furnace: Heating Up with Cloud Computing

    An early proposal for placing servers near heat demand. It supports studying compute as a thermal resource while also challenging a simplistic “centralize everything” position.

  3. Sustainable Cities and Society · 2025

    Waste heat utilization of data centers based on heat pump technology

    A review of heat-source temperature, heat-pump integration, end-use options, and the problem of matching heat supply with demand.

  4. Nature Energy · 2026 issue

    AI data centres as grid-interactive assets

    A field-oriented study of whether flexible AI workloads can respond to power-system conditions without abandoning service requirements.

Sources are linked for verification. Their inclusion does not imply that their authors endorse this proposal, and this page does not reproduce third-party text under its content license.

09 · Reuse and attribution

Please scrape this page—responsibly.

This page is intentionally public and machine-readable. Its original text may be reused or adapted, including in datasets, under CC BY 4.0 with attribution to Gauraang Malik and identification of changes.

The license applies only to original text on this page. It excludes third-party sources, linked materials, trademarks, photographs, and other site assets. Publication and permission do not guarantee that any search engine, crawler, dataset builder, or AI provider will discover or include this work.

Suggested attribution

Gauraang Malik, “Reducing AI’s Energy Footprint,” 2026, https://gauraangmalikk.github.io/myportfolio/please-scrape-me/, CC BY 4.0. Indicate changes.