Full Text
1. Introduction
Cloud computing has reshaped the way organizations acquire and use
information technology. Instead of purchasing and maintaining all
computing infrastructure internally, organizations can obtain
processing, storage, networking and software services from shared
platforms. This model offers elasticity, rapid provisioning, geographic
reach and access to specialized infrastructure. It also changes the
economics of computing by allowing providers to aggregate demand across
customers and operate large facilities with professional engineering
teams. The ability to scale resources on demand can reduce the idle
capacity associated with individually managed server rooms. However, the
efficiency gains of shared infrastructure do not automatically guarantee
a reduction in total environmental impact.
The environmental significance of cloud computing arises from the
physical infrastructure beneath apparently intangible digital services.
Every online transaction, stored file, video stream, model training run
and database query depends on servers, networking equipment, cooling
systems, power distribution and buildings. These systems consume
electricity and require materials throughout their lifecycles. Data
centres also depend on water directly or indirectly, depending on
cooling design and electricity generation. Their environmental impacts
extend beyond operational electricity to include manufacturing,
transportation, construction, maintenance and end-of-life treatment of
equipment.
The scale and growth of data-centre energy demand have attracted
substantial research and public attention. Earlier studies identified
rapid growth in global data-centre electricity consumption, while also
noting that efficiency improvements and changes in computing practices
complicate simple projections (Koomey, 2011). Masanet et al. (2020)
showed that global data-centre workloads and compute instances increased
substantially between 2010 and 2018, yet electricity use grew much more
slowly than might have been expected because of efficiency improvements
in servers, storage, virtualization and facility infrastructure. This
finding is important: technological efficiency can moderate growth, but
it does not remove the need to monitor absolute consumption and the
environmental effects of expanding digital services.
Sustainable cloud computing is therefore concerned with more than
reducing the electricity used by an individual server. It involves the
design, procurement, operation and governance of cloud services so that
computing demand is met with lower environmental impact. Relevant
decisions include where data centres are located, how electricity is
sourced, how workloads are scheduled, how equipment is utilized, how
cooling is provided, how software is designed and how hardware is reused
or recycled. The customer side also matters because inefficient
applications, unnecessary data retention and poorly governed workloads
can increase demand even when infrastructure is efficient.
The purpose of this article is to provide an integrated account of
sustainable cloud computing, with particular attention to energy
efficiency, green data centres and intelligent resource management. It
addresses the following questions: What are the principal environmental
impacts of cloud infrastructure? Which technical and operational
strategies can reduce energy and resource consumption? How can workload
scheduling and artificial intelligence support carbon-aware operations?
Which indicators are suitable for measuring sustainability? What
barriers prevent organizations from adopting greener cloud practices?
The article develops a practical framework for providers, enterprise
customers, public institutions and researchers.
The article is organized as follows. Section 2 establishes the
conceptual foundations and reviews the evolution of green computing.
Section 3 examines the environmental footprint of cloud infrastructure.
Sections 4 to 8 discuss energy-efficient data centres, hardware,
software, workload management, renewable energy and cooling. Section 9
considers measurement and reporting. Section 10 examines intelligent
resource management, while Sections 11 and 12 address implementation
challenges and an integrated sustainability framework. The final
sections discuss implications, research directions and conclusions.
2.
Conceptual Foundations of Green and Sustainable Computing
2.1 From
green IT to sustainable digital infrastructure
Green computing emerged from efforts to reduce the environmental
effects of information technology through energy-efficient equipment,
responsible procurement, improved utilization and environmentally sound
disposal. The concept initially focused heavily on electricity
consumption and the energy efficiency of personal computers, servers and
data centres. As digital services became more complex, the scope
expanded to include software design, networks, cloud platforms,
electronic waste, supply chains and the wider environmental consequences
of digitalization. Sustainable computing is broader than green equipment
because it considers environmental, economic and social dimensions over
time.
Sustainability is often described through the interaction of
environmental protection, economic viability and social responsibility.
In cloud computing, these dimensions are connected. An energy-efficient
data centre can reduce operating costs and emissions, but its location
may affect local water resources, land use and electricity
infrastructure. Extending the life of servers can reduce embodied
emissions, but older equipment may consume more electricity or lack
security support. A sustainable decision must therefore consider
lifecycle impacts, service requirements and the distribution of costs
and benefits.
2.2 Cloud
computing service and deployment models
Cloud services are commonly organized into infrastructure as a
service, platform as a service and software as a service. Infrastructure
as a service gives customers access to virtualized compute, storage and
network resources. Platform as a service provides managed environments
for application development and deployment. Software as a service
delivers complete applications through the network. Each model creates
different opportunities for sustainability management. Infrastructure
customers may control instance types and resource allocation, while
platform customers influence runtime configuration and application
efficiency. Software customers may have less direct control over
infrastructure but can influence usage patterns, data retention and
service configuration.
Deployment models include public, private, community and hybrid
clouds. Public clouds can benefit from high utilization and large-scale
engineering, although customers may have limited visibility into
facility-level energy and water data. Private clouds provide more direct
control but may suffer from lower utilization if capacity is reserved
for individual organizations. Hybrid environments can place workloads
according to latency, compliance, cost and environmental considerations.
The sustainability outcome depends on actual utilization, facility
efficiency, electricity supply and the way workloads are
distributed.
2.3 Efficiency,
effectiveness and absolute impact
Energy efficiency measures the amount of useful computing service
delivered for a given quantity of energy. Examples include transactions
per kilowatt-hour, completed simulations per unit of energy, or useful
compute operations per watt. Efficiency is important, but it should not
be confused with total environmental performance. A system can become
more efficient while total energy use rises because demand grows faster
than efficiency improves. This is a form of rebound effect: lower cost
per unit of service can encourage additional consumption. Sustainable
cloud strategies must therefore monitor both intensity indicators and
absolute impacts.
Effectiveness concerns whether resources are used to achieve
meaningful outcomes. A server running at high utilization is not
necessarily sustainable if it performs unnecessary computation or stores
redundant data. A well-designed application may reduce demand by
avoiding duplicate processing, compressing data, caching frequently used
results and eliminating unused services. Sustainable computing should
ask whether a workload is necessary, whether it can be simplified and
whether its benefits justify its environmental cost.
2.4 Lifecycle thinking
Lifecycle thinking evaluates impacts from raw material extraction and
manufacturing through transportation, operation, maintenance, reuse and
disposal. Operational electricity is visible in utility bills, but
embodied emissions associated with manufacturing servers, storage
devices, networking equipment and buildings can also be substantial. As
electricity supplies become cleaner and operational efficiency improves,
embodied impacts may account for a larger share of total lifecycle
emissions. This creates a tension between replacing older equipment with
more efficient hardware and extending equipment life to avoid
manufacturing impacts.
A lifecycle perspective also includes water consumption,
refrigerants, construction materials, land use and electronic waste. The
environmental profile of a data centre is influenced by local climate,
electricity generation, cooling technology, supply chain practices and
end-of-life management. A single energy metric cannot represent all
these effects. Providers should disclose the boundaries and assumptions
used in environmental reporting so that customers can interpret results
responsibly.
3.1 Electricity consumption
Electricity is the most visible operational input for data centres.
Servers consume energy while performing computation, storage devices
consume energy while reading and writing data, and networking equipment
transfers information between systems. Power distribution and cooling
infrastructure add further demand. The proportion of total facility
electricity used by IT equipment varies according to design, climate,
utilization and operating conditions. A data centre with highly
efficient servers but inefficient cooling may still have a substantial
overhead, while a well-designed facility can reduce non-computing energy
through optimized airflow, power conversion and thermal management.
Data-centre electricity demand is not uniform. Workloads vary by time
of day, season, geography and business cycle. Some facilities operate
near steady utilization, while others experience sharp peaks.
Provisioning for peak demand can leave substantial idle capacity during
quieter periods. Virtualization, workload consolidation and elastic
scaling can improve utilization, but consolidation must be managed
carefully to avoid performance degradation, excessive thermal
concentration or reduced resilience. The objective is to match available
capacity to useful demand while maintaining service-level
requirements.
3.2 Greenhouse gas emissions
The carbon intensity of electricity varies across regions and over
time. A data centre powered by a grid with a high share of coal
generation may have a different operational emissions profile from a
similar facility supplied by a low-carbon grid. Even within a region,
the marginal electricity source can vary by hour. Annual renewable
energy certificates or power purchase agreements can support investment
in clean electricity, but they do not necessarily mean that every hour
of consumption is matched by carbon-free generation. The distinction
between annual accounting and time-specific physical supply is
increasingly relevant to carbon-aware computing.
The emissions associated with cloud services include direct emissions
from on-site fuel use and refrigerant leakage, indirect emissions from
purchased electricity and heat, and other value-chain emissions such as
equipment manufacturing, construction, transportation and waste
treatment. Corporate greenhouse gas accounting commonly distinguishes
these categories through Scope 1, Scope 2 and Scope 3 reporting. Cloud
customers may find it difficult to estimate their share of provider
emissions when allocation methods are opaque. Transparent reporting
should explain how shared infrastructure impacts are attributed to
individual services or customers.
3.3 Water and local
environmental impacts
Water is used in some data centres for evaporative cooling and may
also be consumed in electricity generation. Water use can be especially
important in regions experiencing scarcity or competing demand from
households, agriculture and industry. A cooling system that reduces
electricity consumption may increase water consumption, while a
water-efficient system may require more electricity. The appropriate
design depends on local climate, water stress, electricity mix and
facility requirements. Sustainable site selection should therefore
evaluate water risk alongside energy availability and connectivity.
Local impacts may include land conversion, noise, heat rejection,
construction traffic and pressure on electricity networks. Large
data-centre developments can create economic activity and improve
digital infrastructure, but they may also compete for scarce resources.
Environmental assessment and community engagement should take place
before construction decisions are finalized. Facility operators should
publish meaningful information about water sourcing, discharge, energy
demand and emergency generator testing where disclosure is
appropriate.
3.4 Embodied carbon and
electronic waste
Manufacturing servers and storage equipment requires metals,
semiconductors, plastics and energy-intensive industrial processes. The
extraction and processing of materials can generate greenhouse gas
emissions, pollution and social risks. Rapid hardware refresh cycles may
increase embodied impacts even when new equipment is more energy
efficient. The correct replacement decision depends on the remaining
useful life of existing equipment, the efficiency improvement offered by
new hardware, the carbon intensity of electricity and the availability
of reuse pathways.
Electronic waste presents another challenge. Data-centre equipment
may contain valuable materials as well as substances requiring careful
handling. Reuse, refurbishment, component harvesting and certified
recycling can reduce waste and recover materials. Secure data
sanitization is essential before storage devices are reused or recycled.
Procurement contracts should require suppliers to provide lifecycle
information, take-back options and evidence of responsible end-of-life
management.
4. Energy-Efficient
Green Data Centre Design
4.1 Facility planning and
site selection
The sustainability of a data centre begins before the first server is
installed. Site selection affects climate-related cooling opportunities,
electricity carbon intensity, water availability, access to renewable
generation, grid capacity, natural hazard exposure and network latency.
Cooler climates can support economization for parts of the year,
although climate alone does not determine total environmental
performance. A location with a favourable temperature profile may still
depend on carbon-intensive electricity or face water stress. Site
assessment should evaluate several indicators together and consider
future climate conditions over the facility's expected lifetime.
Facility design should avoid unnecessary overprovisioning. Modular
construction can allow capacity to be added as demand grows, reducing
the energy and material costs of maintaining large unused spaces.
Electrical and cooling systems should be designed to operate efficiently
across a range of loads rather than only at full capacity. Redundancy is
essential for availability, but redundant equipment should be
appropriately sized and controlled so that reliability requirements do
not result in excessive idle consumption.
4.2 Power usage effectiveness
Power Usage Effectiveness, commonly abbreviated as PUE, is calculated
by dividing total data-centre facility energy by the energy consumed by
IT equipment. A value closer to 1.0 indicates that a smaller proportion
of facility energy is used for overhead such as cooling and power
distribution. PUE is useful for tracking facility infrastructure
efficiency and comparing changes within a facility when measurement
boundaries are consistent. It does not measure the efficiency of the
computing work performed, the carbon intensity of electricity, water
consumption or embodied emissions.
PUE should therefore be interpreted as one indicator rather than a
complete sustainability score. A facility can report an excellent PUE
while running inefficient workloads or consuming electricity from a
carbon-intensive grid. Conversely, a facility in a hot climate may have
a higher PUE while using low-carbon electricity and delivering highly
efficient computing services. Reporting should include the measurement
method, time period, IT load, total facility load and operating
conditions. Comparisons should avoid treating small differences in PUE
as proof of overall environmental superiority.
4.3 Airflow management and
cooling
Cooling systems remove heat generated by IT equipment and maintain
operating conditions within manufacturer specifications. Poor airflow
management can cause hot spots, recirculation and unnecessary fan or
chiller energy. Hot-aisle and cold-aisle containment separates supply
and exhaust air, improving the predictability of airflow. Blanking
panels, cable management, raised-floor design and careful placement of
equipment can reduce mixing between hot and cold air. These measures are
often relatively low-cost compared with major cooling plant
replacement.
Free cooling or economization uses outside air or suitable water-side
systems when environmental conditions permit. Direct air economization
can reduce mechanical cooling demand, but it requires careful filtration
and humidity control. Indirect systems can transfer heat without
exposing IT equipment directly to outside air. Liquid cooling can be
effective for high-density computing because liquids transfer heat more
efficiently than air in many configurations. However, liquid systems
require leak management, compatible equipment and careful consideration
of water use and heat reuse opportunities.
4.4 Thermal management and
heat reuse
Data centres reject large quantities of low-grade heat. In some
locations, this heat can be recovered for district heating, nearby
buildings, industrial processes or greenhouse operations. Heat reuse is
constrained by temperature levels, distance to users, seasonal demand
and the cost of heat pumps or distribution infrastructure. Planning for
heat reuse at the site selection stage is more effective than attempting
to retrofit a facility after construction. Operators should assess
whether local heat demand exists and whether the additional
infrastructure delivers net environmental benefits.
Thermal management can also be improved through dynamic control.
Sensors can monitor inlet temperatures, humidity, airflow and equipment
load. Control systems can adjust fan speeds, cooling set points and
chilled-water operation based on measured conditions. These systems
should be designed with safety limits and fail-safe behaviour. Overly
aggressive optimization can create thermal risk, so operators need clear
operating envelopes and monitoring for sensor faults.
5.
Energy-Efficient Computing Hardware and Virtualization
5.1 Server utilization and
consolidation
Underutilized servers consume energy even when they perform little
useful work. Virtualization allows multiple virtual machines to share
physical hardware, increasing utilization and reducing the number of
active servers required for a given workload. Dynamic consolidation can
migrate workloads from lightly loaded hosts to a smaller number of
machines and place unused servers into low-power states or shut them
down. Beloglazov and Buyya (2010) examined energy-efficient resource
management techniques for virtualized cloud data centres, including
dynamic consolidation and energy-aware allocation.
Consolidation is not simply a matter of maximizing utilization.
Excessive consolidation can increase contention for CPU, memory, storage
and network resources. It may also create correlated failure risks if
too many critical workloads are placed on the same physical host.
Scheduling policies should account for service-level objectives,
workload variability, fault domains and maintenance requirements. A
sustainable scheduler seeks an appropriate balance between energy use,
performance, reliability and resilience.
5.2 Hardware
heterogeneity and right-sizing
Different workloads have different resource profiles. CPU-intensive
scientific computing, memory-intensive analytics, storage-heavy
databases and latency-sensitive web services may benefit from different
hardware configurations. Right-sizing selects a server or instance type
that matches the workload rather than allocating a generic high-capacity
configuration. Accelerators can improve performance per watt for
selected workloads, although their manufacturing impacts and utilization
levels must be considered. Specialized hardware is beneficial when it
performs useful work efficiently and is not left idle for long
periods.
Hardware heterogeneity complicates scheduling because performance and
energy characteristics vary across processors, memory configurations,
storage devices and accelerators. Resource managers need reliable
information about workload requirements and hardware capabilities.
Benchmarking should reflect real application behaviour rather than
relying exclusively on theoretical peak performance. The best hardware
choice is workload-dependent, and an efficient configuration for one
application may be wasteful for another.
5.3 Storage efficiency
Storage demand grows through operational records, multimedia,
scientific datasets, backups, logs and replicated application data.
Storage systems consume energy through devices, controllers, cooling and
network traffic. Techniques such as data deduplication, compression,
tiered storage and lifecycle management can reduce the capacity required
for a given service. Frequently accessed data may be placed on
high-performance storage, while older or infrequently accessed data can
be moved to lower-energy or lower-cost tiers where retrieval
requirements permit.
Storage optimization should not compromise durability, recovery or
compliance. Data deletion must follow retention obligations, and backup
copies should be designed to support recovery from ransomware,
accidental deletion and infrastructure failure. Organizations should
identify redundant datasets and unnecessary replicas, but they should
not remove resilience copies without a risk assessment. Sustainable
storage management balances energy and material use with data
availability and integrity.
5.4
Virtualization, containers and serverless computing
Virtual machines, containers and serverless platforms can improve
utilization by allowing workloads to share infrastructure and scale
according to demand. Containers often have lower overhead than full
virtual machines, although the actual advantage depends on workload
design and orchestration. Serverless computing can reduce idle capacity
for event-driven applications because resources are allocated in
response to requests. Yet abstraction can make energy use less visible
to customers, and poorly designed functions may generate excessive
invocations or data transfers.
The environmental benefit of an abstraction model depends on provider
implementation and customer behaviour. Moving an inefficient application
into containers does not automatically make it efficient. Applications
should be profiled, unnecessary work removed and resource requests
aligned with observed demand. Orchestration systems should avoid
excessive overprovisioning and should account for the energy
implications of frequent scaling, image distribution and cross-zone
communication.
6. Sustainable
Software and Application Engineering
Software design has a direct influence on infrastructure demand. An
application that performs redundant calculations, makes unnecessary
network requests or stores excessive data can consume more energy than a
well-optimized alternative. Sustainable software engineering seeks to
reduce the resources required to deliver useful functionality while
preserving quality, security, accessibility and maintainability. It
includes algorithm selection, efficient data structures, caching,
database design, testing, deployment practices and responsible feature
management.
Algorithmic efficiency is important for compute-intensive workloads.
A more efficient algorithm can reduce CPU time and memory use, lowering
energy demand at scale. However, theoretical complexity is not the only
consideration. Real-world performance depends on data distribution,
hardware, implementation language, parallelization and I/O behaviour.
Developers should benchmark representative workloads and measure the
energy or resource cost of important operations where tools permit.
Optimization should focus on high-volume paths and workloads with
significant resource consumption.
Network efficiency can be improved through caching, content delivery
networks, compression, request batching and careful API design. Reducing
unnecessary data transfer can lower network energy use and improve
latency. Database efficiency can be improved through indexing, query
optimization, data partitioning and removal of redundant processing.
Background jobs should be scheduled according to actual business need
rather than running continuously by default. These practices may appear
to be ordinary software engineering, but they also contribute to
environmental performance.
Software lifecycle management should include sustainability
requirements. During design, teams can estimate resource demand and
identify opportunities to avoid unnecessary processing. During
development, code review and automated testing can identify inefficient
patterns. During deployment, teams can select appropriate instance sizes
and autoscaling thresholds. During operation, observability tools can
reveal resource consumption and performance regressions. During
retirement, unused services and obsolete data should be removed through
controlled processes.
Sustainable software must not be pursued by weakening security or
reliability. Disabling encryption, reducing backup frequency or removing
redundancy may lower immediate resource consumption but can create
unacceptable risks. The correct objective is to eliminate waste while
maintaining the service properties required by the business. Security
controls themselves can also be optimized through efficient
cryptographic implementations, sensible logging levels and appropriate
retention policies.
7.
Workload Scheduling and Intelligent Resource Management
7.1 Dynamic allocation and
autoscaling
Cloud workloads fluctuate over time. Static allocation can leave
resources idle during low-demand periods and insufficient during peaks.
Autoscaling adjusts capacity in response to measured or predicted
demand. Reactive scaling uses thresholds based on CPU, memory, request
rate or queue depth. Predictive scaling uses historical patterns and
forecasting to provision capacity before demand increases. Both
approaches can improve utilization, but poor configuration can cause
oscillation, overprovisioning or service degradation.
An energy-aware resource manager should consider the energy
characteristics of available hosts, workload deadlines, performance
requirements and the carbon intensity of electricity. It may consolidate
workloads during periods of low demand, defer flexible jobs to cleaner
electricity periods or select a region with a lower expected emissions
intensity. Such decisions should also consider latency, data residency,
network transfer, resilience and the potential environmental impact of
moving data between regions.
7.2 Virtual machine
placement and migration
Virtual machine placement determines which physical host runs a
workload. Energy-aware placement can consolidate workloads on a subset
of servers while preserving sufficient capacity for performance and
failure recovery. Live migration can move workloads between hosts with
limited interruption, enabling underutilized machines to enter low-power
states. However, migration consumes network and CPU resources and may
increase energy use temporarily. It is beneficial when the expected
savings from consolidation exceed the migration cost and when service
quality remains acceptable.
Beloglazov and Buyya (2010) describe dynamic consolidation strategies
that use utilization thresholds to identify hosts suitable for migration
and shutdown. A practical system must account for workload variation,
heterogeneous hosts and the risk of repeated migration. Hysteresis,
minimum residence times and migration budgets can reduce instability.
Placement policies should also preserve anti-affinity constraints for
critical services and avoid concentrating too many workloads in a single
failure domain.
7.3 Carbon-aware scheduling
Carbon-aware scheduling considers the carbon intensity of electricity
when deciding when or where to run a workload. Flexible workloads such
as batch analytics, backups, model training and scientific simulations
may be shifted within acceptable time windows. A scheduler can
prioritize periods when renewable generation is abundant or grid
emissions are lower. This strategy requires access to reliable
carbon-intensity forecasts and a clear understanding of which jobs can
tolerate delay.
Time shifting and geographic shifting create different trade-offs.
Time shifting delays a workload until a cleaner electricity period in
the same location. Geographic shifting moves a workload to another
region where electricity may be less carbon-intensive. Geographic
shifting can increase network transfer, latency and data movement, and
may conflict with data residency requirements. A carbon-aware scheduler
should evaluate the net emissions impact rather than assuming that a
lower regional grid intensity always produces a better outcome.
7.4 Artificial
intelligence and predictive control
Machine learning can support demand forecasting, cooling
optimization, anomaly detection and workload placement. Predictive
models can estimate future server utilization, cooling demand and
renewable energy availability. Reinforcement learning and optimization
algorithms can explore control strategies under operational constraints.
AI-based cooling control has attracted attention because cooling systems
involve complex interactions among temperature, airflow, humidity,
equipment load and external weather conditions.
AI does not guarantee sustainability. Training and operating models
consume energy, and complex models may use more resources than the
operational savings they deliver. The value of an AI optimization system
should be assessed by comparing its total lifecycle cost with the
measurable savings achieved. Simpler statistical models or rule-based
controls may be more appropriate for some facilities. Human oversight,
safety limits, explainability and fallback modes remain important when
automated systems influence critical infrastructure.
7.5 Multi-objective
optimization
Cloud resource management is a multi-objective problem. Operators
seek to reduce energy use and emissions while meeting latency,
throughput, availability, security and cost requirements. These
objectives can conflict. Consolidating workloads may reduce energy but
increase the risk of correlated failures. Moving a workload to a region
with cleaner electricity may increase latency or data transfer. Delaying
a job may reduce emissions but violate a deadline. Optimization should
therefore make trade-offs explicit and allow business owners to define
acceptable constraints.
A practical objective function may combine energy consumption, carbon
intensity, service-level penalties, migration cost and resource cost.
The weights assigned to these terms should reflect organizational
priorities and regulatory obligations. Some requirements should be
treated as hard constraints rather than tradeable objectives. For
example, a critical medical or financial transaction should not be
delayed solely to obtain a small reduction in estimated emissions.
Flexible workloads offer more opportunities for carbon-aware scheduling
than time-critical services.
8. Renewable Energy and
Carbon Management
Renewable electricity is central to reducing the operational
emissions of data centres. Solar, wind, hydroelectric and geothermal
resources can supply low-carbon electricity, although availability and
lifecycle impacts vary by technology and location. Data-centre operators
may procure renewable electricity through direct ownership, power
purchase agreements, green tariffs or energy attribute certificates.
Each mechanism has different implications for additionality, temporal
matching, geographic matching and market credibility.
Annual renewable matching compares electricity consumption over a
year with renewable generation or purchased attributes over the same
accounting period. This can support renewable investment, but it does
not necessarily demonstrate that a facility is powered by carbon-free
electricity at every hour. Hourly matching provides a more granular view
by comparing consumption with clean generation in each hour. Geographic
matching considers whether generation is available in the same
electricity market or region as the facility. These distinctions matter
for organizations seeking to reduce actual grid emissions rather than
only meet an annual accounting target.
On-site renewable generation can reduce dependence on grid
electricity, but the available land and generation profile may limit its
contribution. Battery storage can shift electricity from periods of high
renewable output to periods of higher demand, while demand response can
move flexible computing workloads toward periods of abundant clean
energy. Backup generation remains necessary for resilience in many
facilities, but operators can evaluate lower-carbon alternatives and
ensure that testing and emergency use are managed responsibly.
Carbon management should include a credible inventory of emissions
sources, defined organizational boundaries and transparent calculation
methods. Scope 1 emissions include direct sources such as fuel
combustion and refrigerant leakage. Scope 2 emissions relate to
purchased electricity, heat or cooling. Scope 3 emissions include
value-chain activities such as hardware manufacturing, construction,
transportation and waste. Cloud providers and customers should avoid
double counting or omitting material sources through unclear allocation
methods.
Renewable procurement should be evaluated for quality and impact. A
certificate purchased from an existing renewable facility may not have
the same effect on new generation capacity as a long-term contract that
enables a new project. Organizations should disclose the instruments
used, the location of generation, the accounting method and the extent
of temporal matching. Claims such as carbon neutral or 100 percent
renewable should be supported by clear boundaries and independently
reviewable evidence.
9. Measurement,
Metrics and Environmental Reporting
Measurement is essential because sustainability programmes cannot be
managed reliably without consistent data. Data-centre operators should
measure electricity consumption at facility and IT levels, water
withdrawal and consumption, cooling performance, renewable electricity
procurement, equipment utilization and relevant emissions. Customers
should seek service-level information that helps them understand the
environmental consequences of their workloads. Where direct measurement
is unavailable, estimates should state assumptions, allocation methods
and uncertainty.
Power Usage Effectiveness measures facility overhead relative to IT
energy, but it should be complemented by indicators of useful work
delivered per unit of energy. Examples include transactions per
kilowatt-hour, virtual machines per kilowatt-hour, completed jobs per
kilowatt-hour and data processed per unit of energy. These indicators
help distinguish a facility that is merely efficient at cooling from one
that delivers computing services efficiently. They should be normalized
carefully because different workloads have different computational
characteristics.
Water Usage Effectiveness can describe water use relative to IT
energy or another defined operational denominator. Carbon Usage
Effectiveness can express carbon emissions relative to IT energy. These
metrics are useful only when their definitions and boundaries are clear.
Water withdrawal is not the same as water consumption, and the
environmental significance of a litre of water depends on local
scarcity. Carbon intensity also varies by grid region and time. A single
global average may conceal important local differences.
Lifecycle assessment provides a broader method for evaluating
environmental impacts across production, operation and end-of-life
stages. It can include greenhouse gas emissions, resource depletion,
water use and other impact categories. Lifecycle studies require
assumptions about equipment lifetime, utilization, electricity mix,
manufacturing processes and allocation among users. Results should
therefore be interpreted with attention to system boundaries and data
quality. Henderson et al. (2020) emphasize the importance of considering
the environmental sustainability of machine learning across the research
lifecycle, a concern that is relevant to cloud-hosted computation more
generally.
Environmental reporting should distinguish absolute consumption from
intensity improvements. A provider may improve energy per transaction
while total energy consumption increases because the number of
transactions grows rapidly. Both figures are important. Reporting should
also disclose trends over time, changes in methodology, renewable
procurement, water impacts, embodied emissions and progress against
stated targets. Independent assurance can improve confidence, although
assurance quality depends on the scope and evidence examined.
10. Barriers
and Risks in Sustainable Cloud Adoption
10.1 Limited
transparency and shared responsibility
Cloud customers often lack direct visibility into facility
operations, hardware utilization, cooling systems and electricity
procurement. Providers may publish sustainability reports, but the
information may be aggregated across regions or business units.
Customers may struggle to estimate the footprint of an individual
application or workload. This information gap makes it difficult to
compare services or verify claims. More consistent reporting standards
and workload-level estimates would support better decisions, provided
that methods are transparent and uncertainty is acknowledged.
The shared responsibility model should extend to environmental
performance. Providers control data-centre design, hardware procurement,
cooling and energy sourcing. Customers influence application
architecture, resource allocation, data retention, workload timing and
service consumption. Software vendors influence the efficiency of the
applications they deliver. Governments and regulators influence energy
markets, disclosure requirements and infrastructure planning. No single
participant can achieve sustainable cloud computing independently.
10.2 Rebound effects and
demand growth
Efficiency improvements can reduce the cost of computing and
encourage new uses. The resulting growth in demand may offset some of
the energy savings. For example, more efficient data processing can make
large-scale analytics economically attractive, leading organizations to
collect and process more data. The expansion of high-resolution media,
connected devices and AI workloads can also increase total
infrastructure demand. Sustainability strategies should therefore
combine efficiency with demand governance, including decisions about
which data to retain, which computations are necessary and which
services provide meaningful value.
10.3 Skills,
procurement and organizational incentives
Sustainable cloud management requires skills that cross traditional
organizational boundaries. Infrastructure teams understand facilities
and resource allocation, software teams understand application
behaviour, procurement teams manage vendor contracts, and sustainability
teams understand environmental reporting. These groups may use different
metrics and have different incentives. A cloud team may be rewarded for
reducing cost, while a sustainability team focuses on emissions and a
product team prioritizes performance. Shared objectives and governance
processes are needed to align decisions.
Procurement processes often emphasize price, performance,
availability and compliance but may not include lifecycle emissions,
water stress or equipment circularity. Organizations can improve
procurement by asking providers for energy and emissions information,
renewable electricity practices, water management plans, hardware reuse
policies and independent environmental disclosures. Contracts can
include reporting obligations and requirements for secure equipment
disposal. However, procurement criteria should be measurable and
proportionate to the organization's influence over the provider.
10.4
Reliability, security and regulatory constraints
Environmental optimization must not undermine service availability,
data protection or legal compliance. Workload migration may be
constrained by data residency, privacy requirements, export controls or
contractual commitments. Aggressive consolidation can reduce redundancy.
Delayed execution can affect time-sensitive services. Sustainable
architecture should identify which workloads are flexible and which
require strict performance or location guarantees. Security controls,
backup arrangements and disaster recovery capabilities should be
maintained even when they add resource overhead.
10.5 Greenwashing
and weak environmental claims
Organizations may make broad environmental claims based on incomplete
indicators or accounting instruments that do not demonstrate actual
reductions in physical impacts. Reporting a low PUE does not prove low
carbon emissions. Purchasing renewable certificates does not necessarily
demonstrate hourly clean-energy matching. Offsetting emissions does not
eliminate the need to reduce direct and value-chain emissions. Credible
sustainability communication should state the scope, method, limitations
and evidence behind each claim. Targets should be measurable, time-bound
and connected to operational plans.
11.
Integrated Framework for Sustainable Cloud Computing
This article proposes an integrated framework with six connected
layers: governance and accountability; measurement and visibility;
efficient infrastructure; intelligent workload management; clean energy
and circular procurement; and continuous assurance. The layers are
mutually dependent. Governance defines objectives and responsibilities.
Measurement provides the evidence needed to identify inefficiencies.
Infrastructure and workload controls reduce resource demand. Clean
energy reduces the carbon intensity of remaining consumption. Circular
procurement addresses embodied impacts and waste. Assurance checks
whether claimed improvements are real and sustained.
Table 1. Sustainable cloud computing dimensions and practical
indicators
Dimension | Management focus | Example indicators
Energy efficiency | Reduce energy required for useful computing | IT
energy, facility energy, PUE, useful work per kWh
Carbon management | Reduce greenhouse gas emissions | Scope 1, 2 and 3
emissions; hourly carbon intensity
Water stewardship | Reduce pressure on local water resources | Water
withdrawal, water consumption, WUE, local water stress
Hardware lifecycle | Extend useful life and reduce material impacts |
Average equipment life, reuse rate, repair rate, e-waste recovery
Workload efficiency | Avoid idle and unnecessary computation |
Utilization, right-sizing, idle capacity, job energy intensity
Software sustainability | Reduce resource demand per service outcome |
CPU time, memory footprint, data transfer, query efficiency
Governance | Establish accountability and credible reporting | Target
progress, data coverage, assurance findings, exception closure
The first layer is governance. Organizations should assign
sustainability responsibilities to executives, cloud architects,
infrastructure managers, software teams, procurement specialists and
data owners. Policies should define environmental objectives, acceptable
trade-offs and reporting requirements. A sustainability steering group
can review performance and resolve conflicts between energy, cost,
availability and compliance. The programme should be connected to
enterprise risk management and procurement rather than operating as a
separate communications activity.
The second layer is measurement and visibility. Providers should
instrument facilities and infrastructure to capture energy, cooling,
water, utilization and emissions data. Customers should use cloud cost
and observability tools to identify idle resources, oversized instances,
unnecessary storage and inefficient data transfers. Measurement systems
should record the time period, location, workload category and
calculation assumptions. Data quality should be reviewed because
unreliable telemetry can produce misleading optimization decisions.
The third layer is efficient infrastructure. This includes efficient
servers, power supplies, storage, networking, cooling, airflow
management and facility design. It also includes virtualization and
consolidation strategies that reduce idle capacity without compromising
availability. Hardware procurement should consider performance per watt,
expected lifetime, repairability and supplier environmental performance.
Facilities should be designed for modular growth and should evaluate
opportunities for heat recovery and low-impact cooling.
The fourth layer is intelligent workload management. Schedulers
should use workload profiles, performance objectives, energy
characteristics and carbon-intensity information to allocate resources.
Flexible workloads can be shifted to periods or locations with lower
expected emissions. Autoscaling should respond to demand without
maintaining excessive idle capacity. AI-based optimization may improve
control, but models should be evaluated against their own resource
consumption and should operate within safety constraints.
The fifth layer is clean energy and circular procurement. Operators
should prioritize credible renewable electricity procurement, energy
storage and demand response. Procurement policies should include
lifecycle information, supplier take-back arrangements, refurbishment
and responsible recycling. Equipment should be replaced based on
lifecycle evidence rather than fixed refresh habits alone. Data
sanitization and chain-of-custody procedures must be maintained
throughout reuse and disposal.
The sixth layer is continuous assurance. Organizations should review
metrics, conduct audits, test optimization policies and publish
progress. Environmental claims should be supported by transparent
boundaries and evidence. Operational teams should investigate unexpected
increases in energy, water or emissions. Lessons from incidents,
equipment changes and workload growth should be incorporated into future
planning. Sustainability should be treated as an ongoing engineering
discipline.
12.
Implementation Roadmap and Practical Applications
A practical implementation roadmap can begin with a baseline
assessment. The organization should identify cloud services, data-centre
assets, workloads, contracts, energy information and environmental
reporting already available. It should determine which applications are
critical, which are flexible and which generate significant compute,
storage or network demand. The baseline should include absolute
consumption where available, as well as intensity metrics. Gaps in
information should be recorded rather than concealed through unsupported
estimates.
The second step is to identify quick operational improvements. These
may include shutting down unused development environments, right-sizing
oversized virtual machines, removing unattached storage volumes,
reducing redundant data retention, optimizing database queries and
applying autoscaling to variable workloads. These changes can often
reduce cost and environmental impact simultaneously. Each action should
be monitored to ensure that it does not create performance or
availability problems.
The third step is to incorporate sustainability into architecture and
procurement. New applications should include resource efficiency
requirements in design reviews. Cloud region selection should consider
latency, data residency, grid carbon intensity and water stress.
Procurement teams should request consistent environmental information
and evaluate equipment lifecycle practices. Contracts should support
transparent reporting and responsible end-of-life management.
Organizations should avoid selecting a provider solely on the basis of a
single sustainability metric.
The fourth step is to introduce intelligent scheduling for workloads
that can tolerate flexibility. Batch processing, backups, data
transformation and some research workloads may be scheduled according to
renewable availability or grid carbon intensity. The organization should
define maximum delay, completion deadlines, data location constraints
and service priorities. Initial pilots should compare energy and
emissions outcomes with a baseline and include the cost of migration or
deferred execution.
The fifth step is to establish continuous governance and assurance.
Sustainability indicators should be included in operational reviews,
architecture boards and supplier assessments. Teams should review trends
and investigate cases where efficiency improves but absolute consumption
increases. Independent assurance may be appropriate for material public
claims. Training should help developers, administrators and procurement
staff understand how their decisions affect cloud sustainability.
Different sectors will emphasize different applications. Universities
and research institutions may focus on efficient high-performance
computing, scheduling scientific workloads around renewable availability
and reducing redundant research data. Financial institutions may
prioritize efficient transaction processing, storage lifecycle
management and the environmental performance of cloud suppliers. Public
agencies may include sustainability requirements in cloud procurement
and digital service design. Small and medium-sized enterprises may begin
with cloud cost optimization, managed services and basic measurement
because these actions can deliver practical improvements without
requiring a dedicated data-centre engineering team.
13. Discussion and
Future Research Directions
The analysis demonstrates that sustainable cloud computing is a
system-level challenge. Improvements in server efficiency, cooling and
workload scheduling are valuable, but their impact depends on the
electricity supply, utilization patterns, equipment lifecycle and growth
in demand. A facility-level metric cannot capture the full environmental
performance of an application. Equally, a customer-level estimate may be
misleading if the provider's allocation method is unclear. Future
reporting should connect facility data, workload characteristics and
lifecycle information while making uncertainty visible.
A major research priority is the development of comparable
workload-level carbon accounting. Customers need estimates that reflect
the location and timing of computation, the energy used by storage and
networking, and the allocation of shared infrastructure. Such estimates
should avoid false precision. A transparent range with documented
assumptions may be more useful than a single number presented without
context. Research should also examine how workload-level reporting
affects procurement and software design decisions.
A second priority is carbon-aware scheduling across geographically
distributed cloud environments. Research can improve forecasting of
renewable generation and grid carbon intensity, while accounting for
latency, network transfer, data residency and reliability.
Multi-objective optimization methods should be tested against real
workloads rather than only simulated environments. Evaluation should
include the energy cost of migration and the risk that scheduling
decisions shift emissions to other parts of the system.
A third priority concerns embodied emissions and equipment
circularity. As operational electricity becomes cleaner, manufacturing
and material impacts may become more significant in lifecycle
assessments. Researchers should examine repairability, modular upgrades,
server reuse, component harvesting and secure refurbishment. The
trade-off between replacing older equipment and extending its life
should be evaluated under different electricity mixes and workload
profiles. Better product-level environmental data would improve these
assessments.
A fourth priority is the environmental cost of artificial
intelligence. AI can optimize cooling, workload allocation and energy
forecasting, but training and inference also consume resources. Research
should compare the environmental cost of optimization models with the
savings they produce. It should also investigate efficient model
architectures, reuse of pretrained models, workload scheduling and
transparent reporting of computational requirements. The environmental
benefits of AI should be demonstrated empirically rather than
assumed.
A fifth priority is the relationship between digitalization and
rebound effects. More efficient cloud services can make new applications
economically viable, potentially increasing total demand. Research
should investigate how organizations can distinguish productive digital
growth from unnecessary computation and data accumulation. Demand
governance, data minimization and software lifecycle management may be
as important as infrastructure efficiency in controlling absolute
environmental impacts.
14. Conclusion
Cloud computing provides important economic and technical advantages,
but its environmental footprint must be managed deliberately. Data
centres consume electricity and water, require materials and generate
lifecycle emissions through equipment manufacturing, construction,
operation and disposal. Improvements in virtualization, hardware
efficiency, cooling and facility design have moderated some growth in
energy demand, yet continued expansion of digital services makes
sustainability an ongoing challenge.
This article has examined sustainable cloud computing through the
connected themes of energy efficiency, green data-centre design,
sustainable software, workload scheduling, renewable electricity,
intelligent resource management and environmental reporting. It has
shown that no single metric or technology is sufficient. PUE is useful
for facility overhead, but it does not measure useful work, carbon
intensity, water stress or embodied emissions. Renewable procurement is
important, but credible claims require clear accounting boundaries and
attention to temporal and geographic matching. Workload consolidation
can reduce idle energy, but it must preserve performance and
resilience.
The proposed framework integrates governance, measurement, efficient
infrastructure, intelligent workload management, clean energy and
circular procurement, and continuous assurance. Providers should improve
facility performance and disclose environmental information. Customers
should design efficient applications, remove unused resources, govern
data retention and select services with lifecycle impacts in mind.
Policymakers and standards organizations can support consistent
reporting and responsible procurement. Researchers should develop better
workload-level accounting, carbon-aware scheduling and lifecycle
assessment methods.
Sustainable cloud computing should ultimately be judged by the
environmental impact of useful digital services, not by the number of
green technologies deployed or by isolated efficiency claims. A credible
strategy reduces unnecessary demand, improves the efficiency of
essential computation, uses lower-impact energy, extends equipment value
and reports outcomes transparently. By embedding these practices into
architecture and operations, organizations can pursue digital growth
while managing the environmental costs of the infrastructure on which
that growth depends.
References
Aslan, J., Mayers, K., Koomey, J. G., &, France, C. (2018).
Electricity intensity of internet data transmission: Untangling the
estimates. Journal of Industrial Ecology, 22(4), 785–798.
https://doi.org/10.1111/jiec.12630
Avgerinou, M., Bertoldi, P., & Castellazzi, L. (2017). Trends in
data centre energy consumption under the European Code of Conduct for
Data Centre Energy Efficiency. Energies, 10(10), Article 1470.
https://doi.org/10.3390/en10101470
Beloglazov, A., & Buyya, R. (2010). Energy efficient allocation
of virtual machines in cloud data centers. In 2010 10th IEEE/ACM
International Conference on Cluster, Cloud and Grid Computing (pp.
577–578). IEEE. https://doi.org/10.1109/CCGRID.2010.45
Buyya, R., Beloglazov, A., & Abawajy, J. (2010). Energy-efficient
management of data center resources for cloud computing: A vision,
architectural elements, and open challenges. arXiv.
https://doi.org/10.48550/arXiv.1006.0308
Buyya, R., Yeo, C. S., Venugopal, S., Broberg, J., & Brandic, I.
(2009). Cloud computing and emerging IT platforms: Vision, hype, and
reality for delivering computing as the 5th utility. Future Generation
Computer Systems, 25(6), 599–616.
https://doi.org/10.1016/j.future.2008.12.001
European Commission, Joint Research Centre. (2022). 2022 best
practice guidelines for the EU Code of Conduct on Data Centre Energy
Efficiency. Publications Office of the European Union.
Freitag, C., Berners-Lee, M., Widdicks, K., Knowles, B., Blair, G.
S., & Friday, A. (2021). The real climate and transformative impact
of ICT: A critique of estimates, trends, and regulations. Patterns,
2(9), Article 100340. https://doi.org/10.1016/j.patter.2021.100340
Greenpeace. (2017). Clicking clean: Who is winning the race to build
a green internet? Greenpeace Inc.
Henderson, P., Hu, J., Romoff, J., BrunSkill, E., Jurafsky, D., &
Pineau, J. (2020). Towards the systematic reporting of the energy and
carbon footprints of machine learning. Journal of Machine Learning
Research, 21(248), 1–43.
International Organization for Standardization. (2015). Environmental
management systems: Requirements with guidance for use (ISO Standard No.
14001:2015).
International Organization for Standardization. (2018). Energy
management systems: Requirements with guidance for use (ISO Standard No.
50001:2018).
Jones, N. (2018). How to stop data centres from gobbling up the
world’s electricity. Nature, 561, 163–166.
https://doi.org/10.1038/d41586-018-06610-y
Kliazovich, D., Bouvry, P., & Khan, S. U. (2012). GreenCloud: A
packet-level simulator of energy-aware cloud computing data centers. The
Journal of Supercomputing, 62, 1263–1283.
https://doi.org/10.1007/s11227-010-0504-1
Koomey, J. (2011). Growth in data center electricity use 2005 to
2010. Analytics Press.
Lannelongue, L., Grealey, J., & Inouye, M. (2021). Green
algorithms: Quantifying the carbon footprint of computation. Advanced
Science, 8(12), Article 2100707.
https://doi.org/10.1002/advs.202100707
Masanet, E., Shehabi, A., Lei, N., Smith, S., & Koomey, J.
(2020). Recalibrating global data center energy-use estimates. Science,
367(6481), 984–986. https://doi.org/10.1126/science.aba3758
Oró, E., Depoorter, V., Garcia, A., & Salom, J. (2015). Energy
efficiency and renewable energy integration in data centres. Strategies
and modelling review. Renewable and Sustainable Energy Reviews, 42,
429–445. https://doi.org/10.1016/j.rser.2014.10.035
Patterson, M., Fenwick, A., & colleagues. (2021). Carbon-free
energy for Google data centers. Google technical report.
Shehabi, A., Smith, S. J., Sartor, D. A., Brown, R. E., Herrlin, M.,
Koomey, J. G., Masanet, E., Horner, N., Azevedo, I. L., & Lintner,
W. (2016). United States data center energy usage report. Lawrence
Berkeley National Laboratory.
Uptime Institute. (2022). Global data center survey 2022. Uptime
Institute.
Zhang, Y., Wang, Y., & Wang, X. (2011). GreenWare: Greening
cloud-scale data centers to maximize the use of renewable energy. In
Middleware 2011 (pp. 143–164). Springer.
https://doi.org/10.1007/978-3-642-25821-3_8