Deep dive: Green IT & sustainable data centers — the fastest-moving subsegments to watch

The global data center industry consumed approximately 460 TWh of electricity in 2025, representing roughly 2% of total worldwide electricity demand, a figure that has nearly doubled since 2020 and is projected to reach 800-1,000 TWh by 2030 according to the International Energy Agency. This trajectory has transformed data center sustainability from a corporate social responsibility exercise into an operational imperative with direct implications for grid reliability, carbon reduction targets, and regulatory compliance. Within this rapidly evolving landscape, several subsegments are moving faster than the rest, attracting disproportionate capital, generating measurable environmental gains, and reshaping the competitive dynamics of the broader IT infrastructure market.

Why It Matters

The convergence of three forces is accelerating the Green IT transformation at a pace that even industry participants did not anticipate. First, artificial intelligence workloads have fundamentally altered data center energy profiles. Training a single large language model consumes 1,000-10,000 MWh of electricity, equivalent to the annual consumption of 100-1,000 US households. Inference workloads, which scale with user adoption, are growing even faster than training loads. NVIDIA's latest GPU accelerators consume 700-1,000 watts per chip, and hyperscale facilities now deploy tens of thousands of these processors in configurations that generate heat densities of 40-80 kW per rack compared to the 5-10 kW per rack typical of traditional enterprise computing.

Second, regulatory pressure has intensified across major markets. The EU's Energy Efficiency Directive requires data centers above 500 kW to report energy performance metrics starting in 2024, with binding efficiency standards expected by 2027. California's Title 24 building energy standards now include specific provisions for data center cooling efficiency. The SEC's climate disclosure rules require publicly traded companies to report Scope 1 and Scope 2 emissions from owned or controlled data centers, and Scope 3 emissions from cloud computing usage when material. These requirements are creating compliance costs that make efficiency investments financially compelling even without environmental motivation.

Third, the physical constraints of water and power availability are forcing design innovation. Major data center markets including Northern Virginia, Dublin, Singapore, and Amsterdam have experienced grid capacity constraints that have delayed or blocked new facility development. Water scarcity has become equally limiting: a typical 100 MW data center using evaporative cooling consumes 1-3 million gallons of water daily, prompting municipalities in water-stressed regions to restrict new data center approvals. These constraints are pushing the industry toward fundamentally different approaches to cooling, power, and site selection.

Key Concepts

Power Usage Effectiveness (PUE) remains the primary efficiency metric for data centers, calculated as total facility energy divided by IT equipment energy. A PUE of 1.0 represents theoretical perfection where all energy powers computing. The industry average PUE has improved from 1.58 in 2018 to approximately 1.40 in 2025, though hyperscale operators routinely achieve 1.10-1.20. However, PUE has significant limitations as a sustainability metric because it does not account for the carbon intensity of energy sources, water consumption, embodied carbon in construction, or the efficiency of the IT hardware itself.

Water Usage Effectiveness (WUE) measures liters of water consumed per kilowatt-hour of IT energy. Traditional evaporative cooling systems produce WUE values of 1.2-2.5 L/kWh. Air-cooled systems achieve 0-0.3 L/kWh but at higher energy costs. Liquid cooling systems can achieve near-zero WUE while simultaneously improving energy efficiency, representing a rare technology that improves both metrics simultaneously.

Carbon-Free Energy (CFE) Matching goes beyond purchasing annual Renewable Energy Certificates (RECs) to match electricity consumption with carbon-free generation on an hourly basis within the same grid region. Google pioneered this approach with its "24/7 Carbon-Free Energy" initiative, achieving 64% average hourly CFE matching across its global fleet in 2023, with a target of 100% by 2030. Hourly matching is significantly more challenging and expensive than annual matching because it requires either on-site generation, battery storage, or access to diverse renewable sources that collectively cover all hours.

Embodied Carbon refers to the greenhouse gas emissions associated with manufacturing, transporting, and installing data center infrastructure, including servers, networking equipment, cooling systems, and building materials. Studies estimate that embodied carbon represents 15-25% of a data center's total lifecycle emissions, a share that increases as operational energy becomes cleaner. For facilities powered entirely by renewable energy, embodied carbon can represent 50-70% of total lifecycle impact.

Fastest-Moving Subsegments

Liquid Cooling for High-Density AI Workloads

Liquid cooling has transitioned from a niche technology used in high-performance computing to a mainstream requirement driven by AI accelerator power densities that exceed the practical limits of air cooling. The global data center liquid cooling market reached $4.8 billion in 2025 and is projected to exceed $15 billion by 2029, representing the fastest-growing subsegment in data center infrastructure.

Direct-to-chip liquid cooling, where coolant circulates through cold plates mounted directly on processors, has become the standard approach for AI training clusters. NVIDIA's GB200 NVL72 rack configuration, which integrates 72 Blackwell GPUs consuming approximately 120 kW total, ships with liquid cooling as the default option rather than an alternative. This OEM-driven shift has eliminated the adoption hesitancy that previously slowed liquid cooling deployment, because operators purchasing the latest AI hardware must implement liquid cooling regardless of their sustainability priorities.

Immersion cooling, where servers are submerged in dielectric fluid, represents the more radical approach with potentially larger efficiency gains. GRC (Green Revolution Cooling) and LiquidCool Solutions have deployed systems achieving PUE values of 1.02-1.05, essentially eliminating cooling energy overhead entirely. The technology also extends server component lifetimes by 20-40% by eliminating thermal cycling stress and airborne contamination, reducing embodied carbon through longer hardware refresh cycles.

The economics have tipped decisively in liquid cooling's favor for new AI-oriented facilities. A 2025 analysis by Uptime Institute found that liquid-cooled AI clusters achieve 25-40% lower total cost of ownership compared to air-cooled equivalents when accounting for reduced cooling energy, higher compute density per square foot, and extended hardware lifetimes. For retrofit applications, direct-to-chip systems can be deployed with minimal facility modifications, while immersion cooling typically requires purpose-built containment infrastructure.

Rear-door heat exchangers represent a transitional technology enabling existing facilities to manage moderate density increases (15-25 kW per rack) without full liquid cooling conversion. Vendors including Vertiv, Schneider Electric, and CoolIT Systems offer rack-level liquid cooling solutions that integrate with conventional raised-floor environments, providing a migration path for operators not yet ready for full immersion or direct-to-chip deployment.

On-Site and Behind-the-Meter Clean Energy

The shift from purchasing RECs to deploying dedicated clean energy generation represents the second fastest-moving subsegment. Hyperscale operators have signed over 40 GW of power purchase agreements (PPAs) for renewable energy, but grid congestion and interconnection delays (averaging 5-7 years in US markets) have pushed operators toward on-site and behind-the-meter solutions that bypass grid constraints entirely.

Fuel cells running on natural gas with carbon capture, or increasingly on green hydrogen, have emerged as a viable on-site power source for data centers. Microsoft's agreement with Plug Power for 12 MW of green hydrogen fuel cells at its Cheyenne, Wyoming facility represents the largest such deployment to date. Bloom Energy's solid oxide fuel cells are operational at data centers for Equinix, CyrusOne, and several hyperscale customers, providing 24/7 baseload power at grid-parity costs in markets with high electricity prices.

Small modular reactors (SMRs) represent the longer-term on-site power play. Amazon Web Services has signed agreements with Dominion Energy for SMR capacity in Virginia, and Google has contracted with Kairos Power for 500 MW of advanced nuclear capacity. While commercial SMR deployments remain 3-5 years away, the magnitude of these commitments signals that nuclear will play a significant role in data center power strategies.

On-site solar paired with battery storage provides supplemental clean energy, though the land requirements for solar (approximately 5-7 acres per MW) limit its contribution at dense urban data center campuses. Edge and smaller colocation facilities in suburban or rural locations have greater opportunity for meaningful on-site solar contribution.

Waste Heat Recovery and District Energy Integration

Data center waste heat recovery has accelerated from pilot projects to commercial-scale deployments, particularly in Northern European markets where district heating infrastructure creates ready buyers for low-grade thermal energy. Data centers reject 100% of the electrical energy they consume as heat, typically at temperatures of 25-45 degrees Celsius for air-cooled facilities and 50-70 degrees Celsius for liquid-cooled systems.

In Stockholm, Interxion's data center supplies waste heat to the Fortum Varme district heating network, providing thermal energy to approximately 35,000 apartments. The arrangement generates revenue of $5-8 per MWh of rejected heat, transforming a waste stream into a revenue source while displacing fossil fuel consumption in the heating network. Similar deployments are operational in Helsinki (Equinix), Paris (Interxion/Data4), and Dublin (Echelon Data Centres).

The economics of waste heat recovery depend critically on temperature and proximity. Air-cooled data centers produce low-temperature waste heat that requires heat pumps to reach the 70-90 degree Celsius supply temperatures needed for district heating, reducing net energy savings. Liquid-cooled facilities, particularly those using immersion cooling, produce higher-temperature waste heat that can be used directly or with minimal boosting. This synergy between liquid cooling adoption and waste heat viability is creating a positive feedback loop: operators deploying liquid cooling for AI workloads simultaneously unlock waste heat recovery revenue that further improves project economics.

In North America, where district heating infrastructure is less prevalent, data center waste heat is finding applications in adjacent commercial greenhouses, aquaculture facilities, and industrial processes. QTS Realty Trust's Richmond, Virginia campus directs waste heat to a co-located greenhouse operation, and Nautilus Data Technologies has proposed floating data centers where waste heat supports aquaculture in temperature-controlled marine environments.

Sustainable IT Hardware and Circular Lifecycle Management

The hardware lifecycle subsegment is gaining momentum as operators recognize that server manufacturing emissions are a significant and growing share of total data center carbon footprint. Extending server lifetimes from the traditional 3-year refresh cycle to 5-6 years can reduce lifecycle emissions by 15-25%, and several operators have implemented extended lifecycle programs.

Dell Technologies and HPE have introduced server product lines with expanded recycled content (30-50% post-consumer recycled plastics) and design-for-disassembly features that increase end-of-life material recovery rates from 65% to over 90%. Circular computing models, where operators refurbish and redeploy servers internally or through certified partners, have reduced procurement costs by 30-40% for non-performance-critical workloads such as development environments, backup systems, and archival storage.

ITRenew (now Iron Mountain Data Centers' asset lifecycle management division) processes over 4 million drives and 500,000 servers annually, extending useful life and recovering materials from equipment that would otherwise enter waste streams. Their operations demonstrate that circular approaches are economically viable at scale: refurbished enterprise SSDs sell at 40-60% of new pricing while meeting performance requirements for many enterprise workloads.

The rise of "server-as-a-service" models from companies like Penguin Computing and Scale Computing enables operators to shift from owning hardware (and bearing disposal responsibility) to consuming compute capacity with lifecycle management handled by the provider. These models align economic incentives with sustainability outcomes because providers retain ownership and are motivated to maximize hardware useful life and end-of-life value recovery.

What's Working

Liquid cooling deployments at AI-focused facilities are delivering measurable efficiency improvements that justify capital investment. Facilities operating immersion cooling at scale report PUE values of 1.03-1.06, compared to 1.25-1.40 for air-cooled AI clusters, translating to 15-25% reduction in total energy consumption. The simultaneously eliminated water consumption from evaporative cooling systems represents a dual sustainability benefit.

Corporate 24/7 CFE matching commitments from Google, Microsoft, and Amazon are driving investment in technologies (long-duration storage, advanced nuclear, enhanced geothermal) that benefit the broader energy transition beyond data centers. The procurement volumes involved, measured in gigawatts rather than megawatts, create market-shaping demand signals.

European waste heat recovery programs are demonstrating that data center thermal output has quantifiable economic and environmental value. Facilities receiving revenue for waste heat achieve 8-15% improvement in overall business economics while displacing fossil fuel heating equivalent to 500-2,000 tonnes of CO2 annually per megawatt of IT capacity.

What's Not Working

PUE as a standalone metric increasingly misleads. A facility with excellent PUE (1.10) running on coal-fired electricity produces dramatically more carbon than a facility with mediocre PUE (1.50) running on 100% renewable energy. The industry lacks a universally adopted holistic metric that captures carbon intensity, water consumption, embodied carbon, and operational efficiency in a single comparable figure.

Scope 3 emissions from cloud computing remain extremely difficult to measure and attribute. Cloud providers publish average carbon intensity figures, but workload-specific emissions depend on server utilization, hardware generation, cooling method, and grid carbon intensity at the specific time and location of computation. The GHG Protocol's guidance on cloud computing emissions is insufficient for the granularity that corporate reporting increasingly requires.

Small and medium enterprise data center operators lack the scale to justify the capital investments in liquid cooling, on-site generation, and waste heat recovery that hyperscale operators are deploying. The resulting efficiency gap between hyperscale and enterprise facilities is widening, which may accelerate workload migration to cloud but creates stranded asset risk for smaller operators.

Embodied carbon measurement lacks standardized methodologies and supply chain transparency. Most server manufacturers do not publish component-level carbon data, and lifecycle assessment methodologies vary significantly across providers, making meaningful comparison impossible. The absence of standardized Product Environmental Footprint data for IT equipment remains a significant gap.

Key Players

Google leads in 24/7 CFE matching with 64% average hourly clean energy matching and published methodologies enabling industry adoption.

Microsoft has committed to being carbon-negative by 2030 and has signed the largest corporate nuclear PPA (Kairos Power, 500 MW via Google; Constellation for Three Mile Island restart).

Equinix operates the largest global colocation portfolio with industry-leading sustainability reporting and over 96% renewable energy coverage.

GRC (Green Revolution Cooling) and LiquidCool Solutions lead in immersion cooling technology with deployments across hyperscale and enterprise customers.

Schneider Electric provides comprehensive data center infrastructure management including power distribution, cooling, and sustainability monitoring software.

Iron Mountain (ITRenew) operates the largest IT asset disposition and circular lifecycle management platform processing millions of devices annually.

Action Checklist

• Assess AI workload growth projections and develop a liquid cooling migration roadmap for racks exceeding 20 kW density
• Evaluate behind-the-meter clean energy options (fuel cells, solar plus storage, or PPA-backed dedicated generation) for facilities in constrained grid markets
• Investigate waste heat recovery partnerships with local district heating networks, greenhouses, or industrial neighbors
• Transition from annual REC purchasing to hourly CFE matching by engaging with providers offering granular carbon attribution
• Implement server lifecycle extension programs targeting 5-6 year refresh cycles for non-performance-critical workloads
• Adopt the EU Energy Efficiency Directive reporting framework proactively, even for facilities outside the EU, as a best-practice benchmark
• Require IT hardware suppliers to provide product-level embodied carbon data and recycled content percentages
• Establish WUE targets alongside PUE, with priority given to technologies achieving water-free cooling

FAQ

Q: What PUE should a modern data center target, and is PUE still a relevant metric? A: New facilities should target PUE of 1.20-1.30 for air-cooled designs and 1.05-1.10 for liquid-cooled facilities. Retrofit improvements typically achieve 1.30-1.45. PUE remains useful for comparing operational efficiency within similar facility types but is insufficient as a sustainability metric because it ignores energy source carbon intensity, water consumption, and embodied carbon. Organizations should track PUE alongside carbon use effectiveness (CUE), water use effectiveness (WUE), and increasingly, lifecycle carbon intensity per unit of useful computation.

Q: How should organizations account for Scope 3 emissions from cloud computing? A: Start with your cloud provider's published carbon data (AWS, Azure, and GCP all provide customer-specific emissions reports). Supplement with workload-level estimates based on instance types and regions, using tools like the Cloud Carbon Footprint open-source calculator. For material Scope 3 reporting, select cloud regions with the lowest grid carbon intensity and negotiate contractual provisions requiring providers to supply auditable emissions data at the workload level. The Science Based Targets initiative's ICT sector guidance provides methodological frameworks for cloud emissions accounting.

Q: Is liquid cooling practical for existing data centers, or only new builds? A: Direct-to-chip cooling and rear-door heat exchangers can be retrofit into existing facilities with moderate modifications (supplemental piping, coolant distribution units, and potentially structural reinforcement for increased rack weights). Immersion cooling typically requires more extensive renovation. The practical threshold is rack density: facilities consistently operating above 15-20 kW per rack will benefit economically from liquid cooling retrofit. Below that density, air cooling with improved airflow management (hot/cold aisle containment, blanking panels) remains cost-effective.

Q: What is the realistic timeline for nuclear power to contribute meaningfully to data center energy supply? A: NuScale's VOYGR small modular reactor received NRC design certification in 2023, but the first commercial deployment (for Utah Associated Municipal Power Systems) was cancelled due to cost escalation. Realistically, commercial SMR capacity serving data centers is 2029-2032 for early adopters. Meanwhile, existing nuclear plant life extensions (like the Three Mile Island Unit 1 restart for Microsoft) could provide nuclear capacity to data centers as early as 2028. Advanced geothermal (Fervo Energy) represents a potentially faster pathway, with initial commercial-scale projects expected in 2027-2028.

Q: How significant is water consumption as a data center sustainability concern? A: In water-stressed regions, it is a critical constraint. A 100 MW air-cooled data center using evaporative cooling may consume 300-500 million gallons of water annually, equivalent to the daily water needs of a city of 30,000-50,000 people. Singapore, parts of Arizona, and the Netherlands have restricted data center development partly due to water concerns. Operators in water-constrained markets should prioritize air-cooled or liquid-cooled (closed-loop) designs that achieve WUE below 0.5 L/kWh, or ideally zero water consumption through dry cooling or direct liquid cooling approaches.

Sources

• International Energy Agency. (2025). Data Centres and Data Transmission Networks: Tracking Report. Paris: IEA.
• Uptime Institute. (2025). Global Data Center Survey: PUE, Sustainability Metrics, and Cooling Technology Adoption. New York: Uptime Institute.
• BloombergNEF. (2025). Hyperscale Data Center Energy Procurement: Trends and Forecasts Through 2030. New York: Bloomberg LP.
• Google. (2025). 24/7 Carbon-Free Energy: 2025 Progress and Methodology Update. Mountain View, CA: Alphabet Inc.
• Lawrence Berkeley National Laboratory. (2025). United States Data Center Energy Usage Report: 2024 Update. Berkeley, CA: LBNL.
• European Commission. (2025). Energy Efficiency Directive: Data Centre Reporting Framework Implementation Guidance. Brussels: EC.
• Masanet, E. et al. (2025). "Recalibrating global data center energy-use estimates." Science, 382(6671), pp. 46-49.