Data center energy consumption: tracking global power demand, water use, and carbon intensity

Why It Matters

Global data centers consumed an estimated 460 TWh of electricity in 2025, roughly 1.8% of total worldwide electricity demand, and the International Energy Agency projects that figure will surge past 800 TWh by 2028 under its base-case scenario (IEA, 2025). That electricity footprint now rivals the annual consumption of France. Behind the numbers lies a collision between two megatrends: the exponential growth of AI workloads, which require orders of magnitude more compute than traditional cloud services, and the urgent need to decarbonize the global energy system. Every ChatGPT query, every AI-generated image, and every large-language-model training run draws power from grids that, in many regions, still rely heavily on fossil fuels.

Water consumption is equally consequential but less visible. A single large hyperscale campus can withdraw millions of gallons of water per day for cooling, straining local aquifers and competing with agricultural and municipal supply. Google disclosed that its global data center operations consumed approximately 6.1 billion gallons (23.1 billion liters) of water in 2024, a 20% increase year-over-year driven primarily by AI infrastructure expansion (Google, 2025). Microsoft reported a similar trajectory, with water consumption rising 34% between 2022 and 2024 (Microsoft, 2025).

For sustainability professionals, tracking data center energy, water, and carbon metrics is essential because digital infrastructure underpins virtually every sector's decarbonization strategy. If the infrastructure itself becomes a major emissions source, the net climate benefit of digitalization erodes.

Key Concepts

Power Usage Effectiveness (PUE). PUE is the industry-standard metric for data center energy efficiency. It is calculated by dividing total facility energy by IT equipment energy. A PUE of 1.0 would mean every watt goes directly to computing; real-world values include overhead for cooling, lighting, and power distribution. The global average PUE has plateaued at approximately 1.58 since 2020, according to the Uptime Institute (Uptime Institute, 2025). Hyperscale operators like Google and Meta report fleet averages between 1.08 and 1.12, but the long tail of enterprise and colocation facilities remains far less efficient.

Water Usage Effectiveness (WUE). WUE measures liters of water consumed per kilowatt-hour of IT energy. Evaporative cooling systems offer excellent PUE but at the cost of high WUE, particularly in hot, arid climates. The metric has gained prominence as regulators and communities scrutinize water-withdrawal permits for new data center developments.

Carbon intensity and carbon-free energy (CFE) matching. Carbon intensity measures grams of CO2 emitted per kilowatt-hour of electricity consumed. It varies dramatically by grid region: a data center in Norway running on hydropower may have a carbon intensity below 20 gCO2/kWh, while one in Poland powered largely by coal exceeds 700 gCO2/kWh. Google introduced the concept of 24/7 carbon-free energy matching in 2021, which goes beyond annual renewable-energy credit purchases to ensure that every hour of consumption is matched by carbon-free generation on the same grid (Google, 2024). This hourly matching approach has since been adopted as a target by Microsoft and Amazon.

Embodied carbon. Beyond operational energy, data centers carry significant embodied carbon in their construction materials, servers, and networking equipment. A 2024 study by researchers at Harvard and Meta estimated that embodied emissions account for 20% to 30% of a hyperscale facility's lifetime carbon footprint, a share that grows as operational grids become cleaner (Gupta et al., 2024).

AI compute scaling laws. Training a frontier large language model in 2025 required approximately 10 to 100 times more compute than an equivalent model in 2022. Epoch AI estimates that compute used in the largest AI training runs is doubling roughly every six months, far outpacing efficiency gains in hardware and software (Epoch AI, 2025). Inference workloads, which run continuously after a model is deployed, are projected to consume five to ten times the energy of training within two years of deployment at scale.

What's Working and What Isn't

Progress. Hyperscale operators have driven meaningful efficiency gains at the facility level. Google's fleet-wide PUE of 1.10 and Meta's 1.08 represent near-theoretical-minimum overhead for air-cooled and evaporative-cooled designs (Google, 2025; Meta, 2025). Liquid cooling adoption is accelerating: Dell'Oro Group estimates that the liquid cooling market for data centers grew 60% year-over-year in 2025, driven by the thermal demands of GPU-dense AI clusters that can exceed 100 kW per rack (Dell'Oro Group, 2025).

Renewable energy procurement by the sector is substantial. Amazon, Microsoft, and Google collectively held over 40 GW of contracted renewable capacity by end of 2025, making the technology industry the largest corporate buyer of clean energy globally (BloombergNEF, 2025). Several operators have moved beyond annual matching toward 24/7 CFE targets, with Google achieving 64% hourly carbon-free energy matching across its global fleet in 2024 (Google, 2025).

Regulatory pressure is mounting in constructive ways. The EU's Energy Efficiency Directive now requires data centers above 500 kW to report PUE, WUE, and renewable energy share annually, creating the first mandatory sustainability disclosure regime for the sector in a major market (European Commission, 2024). Ireland and the Netherlands have imposed moratoriums or conditional permitting requirements on new facilities in grid-constrained regions, forcing operators to invest in on-site generation and storage.

Challenges. Despite facility-level efficiency gains, absolute energy and water consumption continue to rise because workload growth outpaces efficiency improvements. The IEA projects that data center electricity demand could double between 2025 and 2030 under aggressive AI adoption scenarios (IEA, 2025). This "rebound effect" means that better PUE alone cannot bend the emissions curve downward.

Water stress is intensifying. In 2025, several proposed data center campuses in Arizona, Chile, and Spain faced community opposition and regulatory delays over water-withdrawal concerns (S&P Global, 2025). Operators increasingly face a trade-off: air-cooled designs protect water resources but have higher PUE and energy consumption, while evaporative cooling delivers lower PUE but at significant water cost.

Scope 3 emissions from the semiconductor supply chain remain largely unaddressed. Manufacturing a single advanced AI chip generates an estimated 20 to 25 kg of CO2e, and the industry is projected to ship over 10 million AI accelerators annually by 2027 (Semiconductor Climate Consortium, 2025). Few operators currently include chip manufacturing emissions in their carbon footprint disclosures.

Transparency gaps persist. While hyperscalers publish annual sustainability reports, the colocation and enterprise segments, which account for roughly 60% of global data center capacity, provide limited or no public environmental data (Uptime Institute, 2025). This makes sector-wide benchmarking incomplete and hampers policy design.

Key Players

Established Leaders

• Google (Alphabet) — Operates at 1.10 fleet PUE and leads on 24/7 carbon-free energy matching (64% globally in 2024). Published granular water and carbon data since 2021.
• Microsoft — Committed to being carbon-negative by 2030 and water-positive by 2030. Investing in next-generation cooling and small modular reactor partnerships for data center power.
• Amazon Web Services (AWS) — Largest corporate renewable energy buyer globally with 30+ GW contracted. Building custom AI chips (Trainium, Graviton) designed for energy efficiency.
• Meta — Achieves 1.08 fleet PUE. Pioneered open-source data center designs through the Open Compute Project, raising efficiency standards across the industry.

Emerging Startups

• Crusoe Energy — Builds modular data centers powered by stranded natural gas and renewable energy, targeting AI workloads with reduced grid impact.
• Nautilus Data Technologies — Uses water-based cooling with closed-loop systems that eliminate evaporative water loss, achieving PUE below 1.15 with zero water consumption.
• ZutaCore — Develops dielectric liquid cooling solutions for high-density AI racks, enabling rack densities above 100 kW with minimal energy overhead.
• Lancium — Operates flexible data centers that modulate power consumption in response to grid conditions, acting as controllable load resources for renewable-heavy grids.

Key Investors/Funders

• Breakthrough Energy Ventures — Backed by Bill Gates, invests in next-generation cooling, nuclear microreactors, and energy-efficient compute hardware for data centers.
• Infrastructure capital (Blackstone, Brookfield, KKR) — Collectively deployed over $50 billion into data center assets between 2023 and 2025, increasingly with sustainability-linked financing covenants.
• U.S. Department of Energy — Funds research into advanced cooling, waste-heat recovery, and grid-interactive data center controls through national laboratory partnerships.

Examples

Google's 24/7 Carbon-Free Energy program. Google set a target to run every data center on carbon-free energy every hour of every day by 2030. By 2024, its global fleet averaged 64% hourly CFE matching, with individual campuses in Finland, Denmark, and Oregon exceeding 90%. The program has driven the development of new procurement instruments, including time-matched energy certificates (T-EACs), and catalyzed a broader industry shift from annual to hourly matching frameworks (Google, 2025).

Microsoft's water replenishment projects. After disclosing a 34% increase in water consumption between 2022 and 2024, Microsoft launched water-replenishment projects targeting 1.5 times its operational consumption by 2030. Projects include aquifer recharge in Arizona, watershed restoration in South Africa, and rainwater harvesting infrastructure in India. By the end of 2025, Microsoft reported replenishment commitments covering 45% of its global data center water use (Microsoft, 2025).

Equinix's fuel cell deployment. Equinix, the world's largest colocation provider, deployed solid-oxide fuel cells manufactured by Bloom Energy at multiple U.S. campuses. The 60 MW of installed fuel cell capacity provides reliable baseload power with 60% electrical efficiency and produces waste heat that is captured for campus heating. Equinix reports that fuel-cell-powered facilities achieve a combined PUE of 1.16 while reducing grid dependence during peak demand periods (Equinix, 2025).

Open Compute Project (OCP) sustainability track. Founded by Meta, the Open Compute Project publishes open-source server, rack, and facility designs optimized for efficiency. The OCP's sustainability track, launched in 2024, added standardized metrics for embodied carbon, circularity, and water impact. Over 200 organizations now contribute to OCP designs, and Meta estimates that OCP-derived hardware has collectively saved over 100 TWh of energy since the project's inception (Open Compute Project, 2025).

Action Checklist

• Measure and disclose. Track PUE, WUE, carbon intensity, and Scope 1/2/3 emissions for every facility. Align reporting with the EU Energy Efficiency Directive requirements, even if not yet legally obligated, to prepare for expanding regulatory mandates.
• Set hourly CFE targets. Move beyond annual renewable energy certificates toward 24/7 carbon-free energy matching. Prioritize procurement in regions with high grid carbon intensity, where each megawatt-hour of clean energy displaces the most emissions.
• Adopt advanced cooling. Evaluate direct liquid cooling, rear-door heat exchangers, and immersion cooling for high-density AI workloads. Assess the water-energy trade-off explicitly: in water-stressed regions, prefer closed-loop or air-cooled systems even at modest PUE penalties.
• Address embodied carbon. Require server and networking vendors to provide product carbon footprint data. Extend hardware refresh cycles where performance requirements allow, and implement certified e-waste recycling or refurbishment programs.
• Engage on grid planning. Participate in regional grid-planning processes to ensure new data center load is matched with new clean generation and transmission capacity. Support policies that link data center permitting to renewable energy commitments.
• Implement waste-heat recovery. Explore district heating partnerships to channel data center waste heat into nearby buildings, greenhouses, or industrial processes. Projects in Stockholm, Helsinki, and Dublin have demonstrated that waste-heat reuse can offset 10% to 30% of total facility energy costs.
• Benchmark against peers. Use publicly available data from hyperscaler sustainability reports and Uptime Institute surveys to benchmark facility performance and identify improvement opportunities.

FAQ

How much energy does a single AI query consume compared to a traditional web search? A traditional Google search uses roughly 0.3 Wh of electricity, while a generative AI query (such as a ChatGPT response) consumes an estimated 3 to 10 Wh, depending on model size, response length, and hardware efficiency (IEA, 2025). This roughly ten-fold difference explains why AI workload growth is the primary driver of accelerating data center energy demand. As inference optimization improves and purpose-built AI chips mature, this gap is expected to narrow, but aggregate demand will still rise as AI use cases proliferate.

Can data centers realistically achieve net-zero emissions? Operational net-zero is achievable for individual facilities through a combination of 24/7 carbon-free energy matching, on-site renewable generation, and high-quality carbon removals for residual emissions. However, achieving net-zero across the full lifecycle, including embodied carbon from construction and hardware manufacturing, is considerably harder. The most credible pathway combines aggressive operational decarbonization with supply-chain engagement, extended hardware lifecycles, and investment in low-carbon semiconductor manufacturing processes.

Why has the global average PUE stalled at around 1.58? The global average is heavily influenced by the large installed base of older enterprise and colocation facilities that were designed before efficiency became a priority. Many of these facilities use raised-floor cooling, oversized power distribution, and legacy UPS systems that waste 30% to 50% of input energy on non-IT loads. Retrofitting these facilities is expensive, and many operators lack the capital or incentive to invest. New hyperscale builds regularly achieve PUE below 1.2, but they have not yet displaced enough legacy capacity to move the global average significantly (Uptime Institute, 2025).

What role does nuclear energy play in data center power strategies? Several hyperscalers are pursuing nuclear power as a 24/7 carbon-free baseload source. Microsoft signed a power-purchase agreement with Constellation Energy to restart the Three Mile Island Unit 1 reactor specifically for data center load, while Amazon acquired a campus adjacent to Talen Energy's Susquehanna nuclear plant. Small modular reactors (SMRs) are a longer-term option: NuScale and X-energy are developing designs targeting data center applications, though commercial deployment is not expected before 2030.

How can sustainability professionals evaluate a colocation provider's environmental performance? Request PUE, WUE, and carbon intensity data for the specific facility, not just fleet averages. Ask for the renewable energy procurement strategy (certificates vs. direct PPAs vs. hourly matching) and check whether Scope 3 emissions are disclosed. Review water-source documentation, particularly in water-stressed regions. Certifications such as ISO 50001 (energy management) and the EU Code of Conduct for Data Centres provide third-party assurance, as does participation in the Climate Neutral Data Centre Pact for European operators.

Sources

• IEA. (2025). Data Centres and Data Transmission Networks: Tracking Report 2025. International Energy Agency.
• Google. (2025). 2024 Environmental Report: Carbon-Free Energy, Water Stewardship, and Circular Economy. Google / Alphabet.
• Google. (2024). 24/7 Carbon-Free Energy: Methodology and Progress. Google.
• Microsoft. (2025). 2024 Environmental Sustainability Report. Microsoft.
• Uptime Institute. (2025). Global Data Center Survey 2025: PUE Trends, Staffing, and Outages. Uptime Institute.
• Gupta, U. et al. (2024). "Chasing Carbon: The Elusive Environmental Footprint of Computing." IEEE Micro, 44(2), 30-41.
• Epoch AI. (2025). Compute Trends in Machine Learning: 2025 Update. Epoch AI.
• Dell'Oro Group. (2025). Data Center Physical Infrastructure Quarterly Report: Liquid Cooling Market Growth. Dell'Oro Group.
• BloombergNEF. (2025). Corporate Clean Energy Buying: 2025 Market Outlook. BloombergNEF.
• European Commission. (2024). Commission Delegated Regulation on Data Centre Energy Performance Reporting Under the Energy Efficiency Directive. European Commission.
• S&P Global. (2025). Data Center Water Risk: Permitting Challenges and Community Opposition in Water-Stressed Regions. S&P Global Market Intelligence.
• Semiconductor Climate Consortium. (2025). First Annual Greenhouse Gas Report: Semiconductor Manufacturing Emissions Baseline. Semiconductor Climate Consortium.
• Equinix. (2025). 2024 Sustainability Report: Clean Energy, Cooling Innovation, and Circular Operations. Equinix.
• Open Compute Project. (2025). OCP Sustainability Track: Metrics Framework and Member Impact. Open Compute Project Foundation.
• Meta. (2025). 2024 Sustainability Report. Meta Platforms.