Artificial intelligence is often measured through GPU performance, model size, training time, and inference speed. Those measures describe computation, but they do not describe the complete physical system. AI workloads depend on servers, memory, networking, storage, cooling, power systems, and buildings. Each layer requires materials, manufacturing capacity, transportation, electricity, maintenance, and eventual replacement. These activities create environmental impacts across different stages of the infrastructure lifecycle. The International Telecommunication Union supports lifecycle-based assessment for information and communication technologies. Its approach considers production, use, and end-of-life stages within environmental assessment. AI workloads also increase the importance of electrical capacity and thermal management. Lawrence Berkeley National Laboratory estimated U.S. data centers consumed about 176 TWh during 2023. The figure covers data centers broadly rather than AI facilities alone. This distinction matters when evaluating the environmental impact of AI computing.

Why the GPU is only one part of the equation

A GPU performs computation, but it depends on several supporting systems. Servers require processors, memory, storage, circuit boards, power components, and physical chassis. Networking systems connect processors and move data between computing resources. Storage systems retain datasets, checkpoints, models, logs, and operational information. Power systems include transformers, switchgear, UPS equipment, batteries, and distribution components. Cooling systems remove heat from processors and other electronic components. Different facilities use different combinations of air and liquid cooling technologies. Their environmental impacts vary with location, utilization, design, lifetime, manufacturing, and electricity supply. ISO/IEC 30134-8 defines Carbon Usage Effectiveness as a data-center carbon metric. CUE relates use-phase carbon emissions to the energy demand of IT equipment. PUE instead compares total facility energy with IT equipment energy. A complete assessment must therefore examine the infrastructure surrounding the accelerator.

Electricity is only the operational starting point

Electricity is one of the most measurable operational environmental variables in AI computing. Data processing requires electricity, while cooling and facility systems create additional demand. Facility electricity also supports lighting, controls, power conversion, and other auxiliary functions. The share consumed by IT equipment changes with workload intensity and infrastructure design. Power Usage Effectiveness provides a common metric for evaluating facility efficiency. A lower PUE indicates less overhead energy for each unit of IT energy. PUE does not show whether the electricity itself comes from a low-carbon source. Carbon intensity depends on the electricity generation mix serving a facility. That mix can vary between regions and across different periods. The International Energy Agency identifies data centers as a growing electricity demand category. Lawrence Berkeley National Laboratory projects continued growth in U.S. data-center electricity consumption through 2030. Those projections cover data centers broadly and should not represent AI electricity alone.

The grid connection changes the emissions profile

The location of an AI facility affects its operational emissions profile. Electricity generation differs significantly between regional power systems. A facility using lower-carbon electricity can have lower operational emissions than an equivalent facility elsewhere. Renewable energy contracts can also affect reported market-based emissions accounting. Physical electricity consumption still interacts with the regional electricity system. Grid congestion can influence how large new loads connect to available generation. Transmission capacity can also affect the timing and location of additional demand. Generation availability can change the emissions profile of electricity consumed at different times. ITU environmental guidance considers energy, carbon, and water among relevant AI impacts. Operational efficiency therefore does not eliminate emissions from carbon-intensive electricity supplies. Cleaner electricity also does not remove embodied impacts from equipment and construction. A complete assessment should keep operational and embodied impacts as separate categories.

Cooling creates another layer of demand

High-density computing converts electrical energy into heat. Cooling systems must continuously remove that heat from computing equipment. Traditional data centers commonly rely on air-based thermal management. These systems can include fans, heat sinks, air handlers, chillers, and related equipment. Higher accelerator density can increase the heat load within individual racks. Direct-to-chip liquid cooling transfers heat through liquid loops near high-power components. This design can reduce reliance on some air-based cooling infrastructure. Microsoft reports that its newer design can avoid more than 125 million liters of water annually per facility. The figure applies to Microsoft’s applicable zero-water cooling design. Water consumption still depends on facility architecture, climate, and operating conditions. Cooling decisions can create trade-offs involving electricity, water, equipment, and supporting infrastructure. Engineers therefore need to assess cooling as an integrated facility system.

Water and carbon require separate measurements

Water consumption and greenhouse gas emissions represent different environmental pressures. A single metric cannot capture both impacts accurately. Some cooling systems can reduce electricity demand while increasing water consumption. Other designs can reduce water use while requiring additional electricity. The result depends on cooling architecture and facility operating conditions. Data-center water use can originate from cooling and other facility activities. Electricity generation can also create indirect water impacts outside the data center. Water availability and water stress differ substantially between geographic regions. ITU environmental assessment work includes water among relevant AI impact categories. Carbon intensity also varies according to regional electricity generation. These differences make geographic context important for environmental assessment. Multiple metrics provide a stronger basis for evaluating AI infrastructure performance.

Embodied emissions begin before the server reaches the rack

AI infrastructure creates environmental impacts before equipment enters a data center. Semiconductor manufacturing requires specialized facilities, equipment, chemicals, and controlled production environments. Semiconductor production also requires significant quantities of water and electricity. Server manufacturing adds memory, processors, circuit boards, storage, cables, and power components. Data-center construction introduces concrete, steel, glass, insulation, and mechanical systems. Transportation adds another stage before equipment begins operational service. These activities contribute to embodied emissions within the infrastructure lifecycle. Microsoft has reported significant Scope 3 emissions associated with construction and hardware. Its disclosures identify semiconductors, servers, and racks among relevant hardware components. Embodied impacts depend on materials, manufacturing processes, energy sources, transportation, and equipment lifetime. Lifecycle assessment therefore needs to begin before equipment reaches the operational rack.

Construction can become a significant part of the equation

AI expansion requires additional physical capacity for computing systems. New capacity requires buildings, electrical infrastructure, cooling systems, networking spaces, and supporting utilities. Construction therefore contributes environmental impacts before computing operations begin. Concrete and steel can represent important sources of embodied emissions. Their production requires energy-intensive industrial processes and substantial material quantities. Site development also involves excavation, transportation, installation, and commissioning activities. Microsoft has reported construction and hardware as significant contributors to Scope 3 emissions. Microsoft has also reported mass timber use within hybrid construction models for new data centers. The company states that this approach can reduce embodied carbon by up to 65% in applicable projects. The comparison uses typical precast-concrete construction as the reference. Building design can therefore influence the environmental profile of new computing capacity. The scale of construction impacts depends on facility design, materials, site conditions, and construction methods.

Networking and storage are part of the workload

AI clusters depend on high-speed communication between computing resources. Distributed training requires frequent data movement between processors and memory systems. Large-scale inference can also depend on efficient communication between system components. Network switches, optical transceivers, cables, and interfaces support this communication. These components consume electricity and require materials during manufacturing. Storage systems retain datasets, checkpoints, models, logs, and operational information. Storage hardware therefore creates both manufacturing and operational impacts. Network energy demand can vary with architecture, traffic, bandwidth, utilization, and equipment configuration. Storage impacts can vary according to workload, capacity, technology, and operating patterns. Network, storage, and compute components can follow different replacement cycles. Those cycles depend on architecture, workload requirements, reliability, and vendor roadmaps. Measuring accelerator electricity alone can therefore miss supporting infrastructure requirements.

Hardware replacement creates a lifecycle problem

AI hardware continues to evolve as accelerator architectures change. New generations can deliver higher performance or improved efficiency. Some systems can also provide greater memory capacity within similar physical constraints. Replacing equipment can therefore improve computational performance per unit of electricity. Replacement also introduces environmental impacts from manufacturing new equipment. Retired equipment creates additional decisions around reuse, refurbishment, recovery, and recycling. ITU environmental guidance recognizes hardware lifecycle impacts and electronic waste. Lifecycle accounting should separate operational savings from manufacturing impacts. A new accelerator does not automatically produce a lower total environmental impact. The outcome depends on energy savings, utilization, embodied emissions, and service life. Extending equipment life can reduce some lifecycle impacts in suitable circumstances. That decision must also consider reliability, security, performance, maintenance, and workload requirements. Hardware strategy is therefore both a technology decision and a lifecycle decision.

Supply-chain visibility remains a major constraint

AI hardware depends on complex supply chains across multiple manufacturing stages. Semiconductor fabrication can occur separately from equipment assembly and system integration. Memory, circuit boards, storage devices, and power components can involve different suppliers. Raw materials may also originate from different geographic regions. Transportation connects these manufacturing stages before equipment reaches a data center. Scope 3 accounting seeks to capture relevant indirect value-chain emissions. Companies may still lack detailed primary environmental data from every supplier. Different suppliers can also use different boundaries and assumptions in their reporting. These differences make precise comparisons between products more difficult. ITU has identified measurement and data gaps within AI environmental assessment. Standardized lifecycle methodologies can improve consistency between assessments. Procurement teams can request environmental data during hardware selection. Such data can include supplier emissions, materials, lifecycle assumptions, and environmental product information. Better data can make infrastructure trade-offs easier to evaluate.

Measuring useful computation rather than equipment alone

Environmental performance depends on the useful work produced by infrastructure. Equipment specifications alone cannot describe that performance. A faster accelerator may consume more resources elsewhere in the system. Low utilization can also reduce the productivity of installed infrastructure. Higher utilization can distribute some fixed infrastructure impacts across more computational output. Software optimization can change the amount of computation required for a task. Model architecture can influence memory, processing, and communication requirements. Batching and precision can also affect workload efficiency. Scheduling and workload placement can influence infrastructure utilization and electricity demand. ITU guidance connects AI environmental assessment with workload, hardware, energy, and location. Organizations can therefore evaluate facility energy alongside workload output and hardware utilization. CUE provides a facility-level carbon metric, but it does not replace lifecycle assessment. Operational efficiency and lifecycle efficiency should remain separate measurements.

The next stage is system-level accounting

AI infrastructure increasingly requires decisions across computing, energy, materials, and facility engineering. Expanding computational capacity requires additional physical infrastructure. That infrastructure includes electricity supply, cooling, buildings, networking, storage, and hardware. Semiconductor manufacturing and equipment logistics also contribute to the lifecycle profile. Equipment replacement creates another stage within the infrastructure lifecycle. No universal emissions value applies to every AI workload. Results vary with hardware, workload, location, electricity supply, utilization, and accounting boundaries. ITU-T L.1801 provides an LCA-based methodology for assessing AI environmental impacts. ITU-T L.1450 provides a broader lifecycle methodology for ICT environmental assessment. ISO/IEC 30134-8 provides CUE as an operational data-center carbon metric. These frameworks support measurement across different stages and environmental categories. Their combined use can provide a broader view of infrastructure performance. The remaining challenge is obtaining consistent and reliable data across the full supply chain. Better measurement can help infrastructure teams evaluate the environmental cost of useful AI computation.