AI training clusters, cloud platforms and scientific computing compete for the same scarce inputs: power, cooling capacity and silicon. As data center projects expand across the United States, the question for data center operators is no longer only how much computing capacity can be built. It is how much useful work each unit of that capacity delivers per kilowatt-hour, per rack and per ton of embodied carbon.

Sustainable computing is an approach to delivering computing performance while reducing energy use, hardware demand and environmental impact across the technology lifecycle. It treats the computing layer, including processors, software, schedulers and hardware lifecycles, as a source of efficiency in its own right, alongside the buildings that house it.

What is Sustainable Computing?

Sustainable computing aims to minimize the environmental impact of computing while maintaining useful performance, reliability and cost efficiency. It is not limited to buying renewable electricity or building efficient facilities. It asks how much energy and hardware a given computing task requires.

Six practices define it:

• Energy-efficient hardware: Processors and accelerators that complete more work per watt for the workloads actually run.

• Efficient software: Algorithms, code and data handling that reduce unnecessary computation and data movement.

• High utilization: Virtualization, containers and consolidation that reduce idle capacity.

• Power-aware scheduling: Placing and timing workloads according to power, thermal and grid conditions.

• Hardware durability and reuse: Longer service lives, repair, refurbishment and responsible end-of-life handling.

• Carbon and resource accounting: Tracking emissions, water and materials associated with computing, including manufacturing.

Let’s get into an illustrative example, assume two systems each serve 10 million requests per day at the same latency target. System A draws an average of 50 kW and System B draws 30 kW. System A therefore uses 1,200 kWh per day, or 0.12 Wh per request, while System B uses 720 kWh per day, or 0.072 Wh per request. 

The 480 kWh daily gap is about 175 MWh a year in IT energy alone, before cooling and power-conversion overhead. The outputs are identical, so the difference lies entirely in how efficiently the computing is done.

Sustainable Computing vs. Green Computing vs. Green Data Centers

Why Sustainable Computing Is Becoming Important for Data Centers

Growing AI and High-Performance Computing Workloads

AI training and inference place different demands on data centers. Training involves processing large datasets and updating model parameters across multiple processors. Inference uses trained models to generate responses, analyze information and support applications, often around the clock.

Both workloads require computing power, but their energy consumption depends on the model, hardware, utilization and amount of work performed. Sustainable computing helps operators deliver the required performance while reducing unnecessary energy use.

Rising Electricity Demand and Power Availability

Data center expansion depends on access to electricity, grid connections, substations and generation capacity. As more facilities come online, securing sufficient power becomes a major development challenge.

The US Energy Information Administration reported that Virginia's commercial electricity sales increased by nearly 30 million MWh between 2019 and 2025, with data centers a major driver. The International Energy Agency also estimates that data centers will account for nearly half of US electricity demand growth through 2030.

These trends highlight two related priorities: ensuring enough electricity is available and using that electricity efficiently. Measuring energy consumption per workload helps operators understand how effectively their infrastructure converts power into useful computing output.

Cooling Requirements of High-Density Computing

Processors convert most of the electricity they consume into heat. As computing density increases, data centers must remove more heat from smaller spaces without compromising performance.

Traditional air cooling may struggle to manage the heat generated by densely packed AI accelerators. Liquid cooling offers an alternative for high-density deployments, while effective workload placement and thermal management can help reduce unnecessary cooling demand.

Computing efficiency therefore depends on more than processor performance. Operators must also manage heat, maintain suitable operating temperatures and prevent thermal throttling.

Operating Costs and Infrastructure Utilization

Servers and accelerators can consume electricity even when they are underused. Poor workload allocation, idle hardware and inefficient software can increase energy consumption without delivering proportional computing output.

Improving utilization allows operators to make better use of existing infrastructure. However, they must balance utilization against application performance, response times and the need to maintain spare capacity for demand spikes or equipment failures.

Sustainable computing focuses on delivering more useful work from existing resources rather than simply adding more hardware.

Carbon Reporting and Environmental Accountability

Data center sustainability involves both operational and embodied emissions. Operational emissions are associated with the electricity used to run computing and cooling equipment. Embodied emissions arise from manufacturing hardware, constructing facilities and replacing equipment.

These impacts can move in different directions. For example, a data center operator may reduce emissions from electricity consumption while increasing supply-chain emissions through hardware purchases and construction.

For data center operators, sustainable computing is therefore not just about reducing electricity use. It also involves improving hardware utilization, managing cooling demand and accounting for the environmental impact of the infrastructure throughout its lifecycle.

Key Technologies Enabling Sustainable Computing

Energy-Efficient CPUs, GPUs, and AI Accelerators

Efficient hardware performs more computing work with less energy. Modern CPUs, GPUs and AI accelerators use specialized architectures, faster memory and power-management features to improve performance while controlling electricity consumption.

However, peak performance alone does not show how efficiently a chip handles real workloads. Operators need to compare hardware under similar conditions, including workload size, output accuracy and response time.

The June 2026 Green500 ranking placed the KAIROS supercomputer at 73.282 gigaflops per watt on the High-Performance Linpack benchmark. This measures performance on a specific computing task, not the energy efficiency of every AI workload. MLPerf Power provides a more relevant way to compare power consumption across machine-learning tasks, although results still depend on the benchmark and testing conditions.

Software Optimization and Efficient Algorithms

Software affects how much computing power a task requires. Efficient algorithms, smaller models and reduced data movement can help applications complete work using fewer resources. Techniques such as model quantization and pruning can also reduce computation in suitable AI workloads.

These improvements can lower electricity consumption and help organizations delay hardware upgrades. However, results vary by application. Operators must test whether an optimization reduces energy use without compromising accuracy, performance or reliability.

Virtualization, Containerization, and Server Consolidation

Virtualization and containerization allow multiple workloads to share physical infrastructure. This helps operators improve server utilization and reduce the number of machines needed to run applications.

Consolidation still requires careful planning. Overloading shared servers can increase latency and create resource conflicts. Operators must balance utilization against performance, security and availability requirements.

AI Workload Scheduling and Resource Orchestration

Workload scheduling determines where and when computing tasks run. Orchestration systems can allocate workloads based on available hardware, power demand, temperature and application requirements.

Carbon-aware scheduling adds electricity-related emissions to these decisions. Operators can shift flexible workloads to periods when the electricity supply has a lower carbon intensity. This approach works best for tasks that can tolerate delays, such as some batch-processing jobs. Real-time applications and workloads with strict deadlines have less flexibility.

The benefits also depend on the accuracy of carbon forecasts, data-transfer requirements and the availability of suitable computing resources. Carbon-aware scheduling can reduce emissions from flexible workloads, but it cannot replace improvements in hardware, software and power infrastructure.

Liquid Cooling and Thermal Optimization

As data center servers become more powerful, they generate more heat in increasingly compact spaces. Operators use airflow management, direct-to-chip liquid cooling and immersion cooling to maintain suitable operating temperatures.

Liquid cooling can remove heat more effectively in high-density environments, while thermal-aware workload placement helps prevent overheating and unnecessary cooling demand. The benefits depend on facility design, climate and the method used to reject heat. Lower cooling electricity consumption does not automatically mean lower water use.

Power Management and Dynamic Scaling

Processors do not need to operate at maximum speed for every task. Dynamic voltage and frequency scaling (DVFS) adjusts processor voltage and clock speed to match workload requirements. Power capping limits the electricity a server or accelerator can draw.

These techniques can reduce energy consumption when applications do not need full computing capacity. Operators must still maintain acceptable processing times, application performance and reliability.

The Zeus framework reported improvements of 15.3% to 75.8% in training energy efficiency across its tested configurations. The Perseus system reported up to 30% lower energy consumption while maintaining the same training time in selected tests.

Hardware Lifecycle Management and Circular Computing

Manufacturing servers, processors and other data center equipment creates emissions before the hardware begins operating. Extending equipment life, repairing components, reusing servers and recycling materials can reduce waste and limit demand for new hardware.

Operators must weigh the environmental cost of replacement against the energy savings offered by newer equipment. Older servers may consume more electricity for the same workload, while new hardware introduces additional manufacturing emissions. The better choice depends on the efficiency difference, electricity source, remaining equipment life and expected workload.

A lifecycle-based approach helps operators make replacement decisions based on total environmental impact rather than hardware age alone.

How Sustainable Computing Is Transforming Data Center Operations

Sustainable computing changes operating decisions more than it adds equipment. In practice it shows up as:

• Procurement by workload. Servers are chosen on measured efficiency for the intended workload, not peak specifications.

• Capacity planning on real utilization. Forecasts account for capacity that already sits idle.

• Dynamic placement and scheduling. Workloads move according to load, thermal headroom and, where feasible, grid conditions.

• Coordinated computing, cooling and power. Shared telemetry lets compute decisions respect thermal and electrical limits. See data center infrastructure.

• Multi-level monitoring. Energy is measured at server, rack, cluster and facility levels so gains can be attributed.

• Lifecycle-based upgrades. Refresh and retirement decisions weigh embodied and operational emissions together.

Real-World Examples of Sustainable Computing

1. Google: Reducing the Energy Use of Gemini Apps

Google studied how much energy its Gemini Apps use to answer user requests. The company measured power consumption across the serving system, including accelerators, host systems, idle capacity and data center overhead.

According to Google's measurements, a median text prompt in May 2025 used 0.24 Wh of energy, produced 0.03 gCO₂e and consumed 0.26 mL of water. Google also reported that energy use per median prompt fell 33-fold over 12 months, while its carbon footprint fell 44-fold.

These figures cover median text prompts, not every type of AI request. They are company-reported measurements and do not represent the energy use of Google's entire AI infrastructure. Google also reported that its total electricity consumption increased by 37% in 2025.

2. Google: Scheduling Workloads Around Cleaner Electricity

Electricity-related emissions vary throughout the day. Google developed a system that shifts flexible computing tasks away from hours when electricity has a higher carbon intensity.

The system uses day-ahead carbon forecasts to adjust the computing capacity available to batch jobs that can finish within 24 hours. This allows Google to reduce demand during high-carbon periods without delaying tasks that require immediate processing.

In a two-month controlled experiment, Google reported a 1% to 2% reduction in average cluster power during the hours with the highest carbon intensity. In two clusters, flexible computing demand fell by about 50%, while total power consumption dropped by around 8% for three to six hours.

3. University of Michigan: Improving AI Training Efficiency

AI training jobs do not always need GPUs to run at their default settings. The University of Michigan developed two open-source systems, Zeus and Perseus, to help researchers reduce training energy without unnecessarily extending processing time.

Zeus tests different job and GPU configurations to find more energy-efficient settings. Perseus adjusts GPU speeds when certain GPUs are not critical to completing a training task.

In their evaluations, Zeus improved training energy efficiency by 15.3% to 75.8%. Perseus reduced energy consumption by up to 30% while maintaining the same training time in selected tests involving GPT-3, other large language models and a vision model. These results show how software can improve the efficiency of existing hardware. 

4. Google: Measuring the Carbon Impact of AI Hardware

Improving processor performance is only one part of sustainable computing. Manufacturing hardware also produces emissions, so operators need to consider both electricity use during operation and emissions from production.

Google researchers introduced Compute Carbon Intensity, which measures carbon emissions relative to computing output. They used this metric to compare different generations of Google's Tensor Processing Units (TPUs).

Their research reported an approximately threefold improvement in Compute Carbon Intensity from TPU v4 to Trillium. The analysis also found that operational emissions accounted for more than 70% of lifetime emissions, while embodied emissions remained a significant part of the total.

5. Microsoft: Reusing and Recycling Data Center Hardware

Data centers regularly retire servers and components as equipment ages or operational requirements change. Microsoft operates Circular Centers to test, refurbish and reuse equipment, while sending remaining materials for resale, donation or recycling.

Microsoft reported a 90.9% reuse and recycling rate for servers and components in 2024. The company also reported reusing more than 3.2 million components that year. Its Circular Centers program includes facilities in the United States, Europe and Singapore. Reusing equipment can reduce electronic waste and the need to manufacture replacement components.

Challenges and Limitations of Sustainable Computing

Several barriers slow adoption:

• Performance trade-offs and constraints. Power capping, consolidation and lower clock speeds can raise latency or cut throughput, and reliability, security and service-level commitments limit scheduling flexibility.

• Upfront cost and legacy software. New hardware requires capital, and legacy applications may not run on new architectures without rework.

• Measurement gaps. Energy is easy to meter at the rack and hard to attribute to one workload on shared hardware.

• Resource competition. GPUs, memory and power allocations are scarce, and efficiency work competes with capacity delivery.

• Embodied carbon. Frequent upgrades can raise manufacturing emissions even as operational energy falls.

• Shifting impacts. Moving work across hours or regions can move water use, electricity demand or network load elsewhere.

Conclusion

Sustainable computing works as a system. Efficient hardware sets the baseline, optimized software reduces the work required, scheduling and consolidation raise utilization, thermal management keeps dense equipment within limits, and lifecycle decisions govern embodied carbon. The goal is not simply to consume less electricity. It is to deliver useful computing output more efficiently while accounting for reliability, cost, carbon emissions and resource use.

The evidence also shows why measurement matters. Per-task gains are documented, yet total consumption at the largest operators keeps growing. As AI, cloud and hyperscale infrastructure expand, the operators best placed to manage power and carbon constraints will measure energy per workload, not only per facility, and treat sustainable computing as a core operating discipline.

Find the Latest Data Center Facility Projects Around the World with Ease

Are you seeking reliable and up-to-date insights into data center projects worldwide?

Explore the Global Project Tracking (GPT) platform by Blackridge Research, your go-to resource for the latest data center projects and tenders across all stages:

• Upcoming Projects

• Tender Notices

• Contract Awards

• Projects Under Construction

• Completed Projects

 Book a Free Demo today and see how the GPT platform can help you unlock opportunities and achieve your business goals.