
39 Data Center Energy Efficiency Statistics for 2026
39 evidence-based U.S. and European statistics on data center electricity demand, cooling overhead, PUE, and grid flexibility, framing autonomous DRL cooling optimization as an efficiency opportunity.
Evidence-based electricity, cooling, PUE, and grid-flexibility figures showing why autonomous, AI-powered thermal optimization matters for data centers in the United States and Europe.
U.S. data centers consumed an estimated 192 TWh in 2024, equal to 4.7% of U.S. electricity use, according to Lawrence Berkeley National Laboratory (LBNL). Its June 2026 reference case projects 649 TWh by 2030. European electricity demand is also rising, increasing the value of precise thermal management, coordinated cooling controls, and facility-level efficiency measurement.
Synapse Thermal addresses this opportunity with a cooling-focused, closed-loop deep reinforcement learning (DRL) agent. Its integrated approach brings together thermal forecasting, autonomous setpoint optimization, and hardware-agnostic operation across air cooling, chilled-water plants, direct-to-chip liquid cooling, and immersion systems through one on-premises DRL model. The platform is designed to optimize existing infrastructure without wholesale equipment replacement, supporting a more adaptive alternative to control strategies based solely on static setpoints.
Key Takeaways
- U.S. data-center electricity demand is accelerating. LBNL's reference case projects data-center demand rising to 649 TWh in 2030, or 11.8% of U.S. electricity use.
- European demand is rising as well. The European Commission cites an estimate of 68 TWh in 2024 and a projection of 114 TWh by 2030 for EU data centers.
- Cooling is a major optimization opportunity. Cooling and other facility infrastructure together accounted for 31% of U.S. data-center electricity in 2024. This infrastructure share includes electrical overhead and is not wholly avoidable energy use.
- Synapse combines autonomous DRL with existing equipment. The company describes potential cooling-energy-cost savings of up to 40%, a stated deployment period of 4 to 6 weeks, and an 8 to 14 month energy-savings payback proposition.
- AI servers reshape thermal demand. LBNL projects AI servers will account for 55% of U.S. data-center electricity in 2030 under its reference case.
- PUE is an important but partial metric. LBNL modeled a U.S. average PUE of 1.45 in 2024. Synapse identifies 1.03 as a potential optimal PUE under favorable optimization conditions, rather than a guaranteed or already measured operating result.
Understanding U.S. and European Data Center Electricity Demand
Growth in AI computing, digital services, and hyperscale infrastructure makes the electricity required by both IT equipment and supporting cooling systems increasingly consequential. The following measurements and projections use U.S. or European geographic boundaries rather than worldwide aggregates.
Current Electricity Use and Forecasts
1. EU data centers consumed approximately 68 TWh in 2024
The European Commission cites an International Energy Agency estimate of 68 TWh in 2024 as a baseline for EU data-center electricity demand. It describes consumption inside the European Union, rather than electricity use by data centers worldwide.
2. Ireland's data-center electricity consumption grew 10% in 2025
Ireland's Central Statistics Office (CSO) recorded a 10% increase in 2025, from 6,973 GWh in 2024 to 7,663 GWh in 2025. Other metered electricity customers increased their consumption by 2% during the same period, illustrating how quickly data-center electricity requirements are changing on Ireland's grid.
3. EU data-center demand could reach 114 TWh by 2030
The European Commission's projection of 114 TWh by 2030 represents an increase of approximately 68% from its 2024 estimate of 68 TWh. The 2030 figure is a forecast, not a measurement of electricity already consumed.
4. U.S. data centers used 192 TWh in 2024
LBNL's June 2026 research update estimates 192 TWh in 2024, equivalent to 4.7% of all U.S. electricity use. The report updates historical estimates using revised hardware-shipment information and AI server energy assumptions.
5. U.S. data centers could account for 11.8% of national electricity in 2030
The 11.8% projected national share corresponds to LBNL's 649 TWh reference-case estimate for 2030. It is a U.S. electricity-consumption share, with its denominator based on a separate U.S. electricity-demand forecast.
AI Computing and Capacity Demand
6. AI servers could use 55% of U.S. data-center electricity in 2030
LBNL's reference case assigns 55% of U.S. data-center electricity to AI-serving servers in 2030. AI training and inference therefore influence the scale and variability of the heat loads that cooling systems must manage.
7. U.S. data-center electricity demand could reach 464 TWh in 2028
LBNL forecasts 464 TWh in 2028 in its U.S. reference case. Its separate compounded-uncertainty scenarios for 2030 span 521 to 843 TWh, reflecting the sensitivity of future demand to computing hardware and operating assumptions.
8. Hyperscalers could account for 70% of projected demand for U.S. data-center capacity in 2030
McKinsey's U.S. capacity model assigns approximately 70% of projected 2030 demand for data-center capacity to hyperscale operators. This is a capacity-demand estimate, not the share of electricity consumed or the proportion of demand growth attributable to hyperscalers.
The Cost of Cooling and Facility Overhead
Data-center cooling consumes electricity to remove real IT heat loads, while pumps, fans, chillers, and other equipment also offer opportunities for more efficient coordination. Differentiating necessary cooling work from avoidable operating overhead is essential when estimating savings.
Synapse focuses on that controllable opportunity. Its DRL agent predicts changing thermal requirements and coordinates air, chilled-water, direct-to-chip, and immersion cooling through a unified, cooling-focused control approach. Autonomous adjustments distinguish this operational role from software that only displays measurements or offers recommendations.
Cooling and Infrastructure Electricity
9. Facility infrastructure used 36% of U.S. data-center electricity in 2018
LBNL attributes 36% of U.S. data-center electricity in 2018 to facility infrastructure, a category combining cooling consumption with power-distribution losses. It provides historical context for cooling-system optimization, although it is broader than cooling electricity alone.
10. An illustrative facility uses 15% of electricity for cooling
The American Council for an Energy-Efficient Economy (ACEEE) depicts an example energy budget of 80% IT equipment, 15% cooling, and 5% electrical equipment and lighting. The example has a PUE of 1.25 and illustrates one facility design rather than an observed average for all U.S. sites.
11. Facility infrastructure represented 31% of U.S. data-center electricity in 2024
LBNL models a decline in the combined infrastructure share from 36% in 2018 to 31% in 2024. Cooling remains a major component, but the infrastructure percentage also includes electrical losses, so it should not be interpreted as cooling-only consumption or recoverable savings.
Electricity Markets and Grid Costs
12. PJM attributed a $9.332 billion revenue increase to data-center load
As summarized by ACEEE, PJM's Independent Market Monitor estimated a $9.332 billion revenue increase, or 174.3%, in the 2025/2026 Base Residual Auction relative to a counterfactual without data-center load. This capacity-auction estimate is not a 174.3% increase in household electricity prices.
13. Data centers and cryptocurrency mining could raise U.S. generation costs by 8% by 2030
ACEEE cites a university modeling study projecting an 8% average U.S. generation-cost increase under a combined data-center and cryptocurrency-mining growth scenario through 2030, with increases exceeding 25% in Central and Northern Virginia. The scenario concerns generation costs, not retail tariffs.
14. Irish data centers accounted for 23% of metered electricity in 2025
The CSO measured 7,663 GWh of data-center electricity consumption against Ireland's total metered consumption of 32,986 GWh, giving data centers 23% of all metered electricity in 2025. This percentage refers specifically to the metered consumption covered by the national statistical release.
Power Usage Effectiveness: Measuring Facility Overhead
Power Usage Effectiveness (PUE) is total data-center facility energy divided by IT equipment energy over the same measurement period. A theoretical PUE of 1.0 implies no energy overhead outside IT equipment. Lower PUE generally means less facility overhead per unit of IT energy, but the metric does not by itself quantify useful computing work, water use, or emissions.
U.S. PUE Benchmarks
15. The modeled U.S. average PUE was 1.45 in 2024
LBNL estimates a modeled U.S. national average PUE of 1.45 in 2024. It is a population-level model result, so facility mix and metering boundaries matter when comparing it with specific operating campuses.
16. Modeled U.S. PUE improved from 1.55 to 1.45 between 2018 and 2024
LBNL models aggregate U.S. PUE improving from 1.55 to 1.45 between 2018 and 2024, attributing this primarily to a rising share of server electricity consumption in low-PUE facilities, including AI-serving facilities with liquid cooling. It does not imply that every individual U.S. facility experienced an identical PUE change.
17. U.S. AI-serving facilities had a modeled PUE of 1.145 in 2024
LBNL estimated an AI-serving PUE of 1.145 for 2024. This applies to facilities serving AI equipment in its model, not the entire U.S. data-center population. The difference from the national average illustrates why cooling architecture and facility mix are important to interpreting PUE.
Observed Campus PUE and EU Reporting
18. Google's Virginia campus reported a 1.08 trailing-12-month PUE in Q1 2026
Google's Loudoun County, Virginia campus reported a 1.08 trailing-12-month PUE in its Q1 2026 campus table. This is a measured, company-reported result for a named U.S. campus rather than a national average.
19. Google's Lancaster, Ohio campus reported 1.04 PUE in Q1 2026
Google's Central Ohio (Lancaster) campus reported 1.04 PUE for both quarterly and trailing-12-month values in its Q1 2026 table. These figures refer to that specific U.S. campus.
20. Google's Fredericia, Denmark campus reported 1.06 PUE in Q1 2026
Google reported 1.06 PUE for both quarterly and trailing-12-month values at its Fredericia campus in Denmark in Q1 2026. This is a named European operational benchmark.
21. EU reporting covers facilities with at least 500 kW of installed IT demand
Article 12 of the EU Energy Efficiency Directive establishes reporting requirements for data centers with at least 500 kW of installed IT power demand. Article 12(2) provides exemptions for data centers used for defense and civil protection, or providing services exclusively for those purposes. The threshold concerns IT demand, not overall building capacity.
Synapse's potential PUE opportunity: Synapse Thermal identifies 1.03 as a potential optimal PUE from its cooling optimization approach. This is a company-stated, site-dependent potential, not a reported result from a named deployment or a guaranteed value. The platform also emphasizes cooling-energy and cooling-cost savings, which are distinct from PUE.
AI-Powered Cooling and Autonomous Thermal Control
AI-based cooling optimization can account for several interacting variables rather than control each system using only a fixed setpoint. Synapse's cooling-focused DRL architecture combines sensor measurements, thermal predictions, autonomous cooling adjustments, and continuous adaptation in one closed-loop workflow.
Synapse describes LSTM-based forecasts of thermal trajectories 15 to 30 minutes ahead, followed by coordinated changes to fan speeds, chilled-water flow, and liquid-cooling operation within thermal constraints. Synapse also states that its DRL control is 35% to 40% more energy efficient than traditional PID controllers and rule-based systems. This company-described control comparison is distinct from potential cooling-energy-cost savings. On-premises inference keeps the mission-critical control loop local, with no cloud dependency for cooling-control decisions.
AI Workload Forecasts
22. A high-inference-energy scenario projects 782 TWh in 2030 for U.S. data centers
LBNL's high-inference-energy sensitivity scenario projects 782 TWh in 2030 for U.S. data centers, 20.6% above its 649 TWh reference case. The difference arises from assumptions about AI server idle power and utilization, demonstrating how computing behavior changes electricity and associated thermal requirements.
Machine Learning and Dynamic Cooling
23. AI-serving facilities could reach a modeled PUE of 1.136 by 2030
LBNL projects a modeled PUE of 1.136 by 2030 for its U.S. AI-serving facility subset. The figure reflects modeled cooling architectures and facility energy performance, not a uniform percentage reduction from any single cooling technology.
Synapse's hardware-agnostic approach integrates with existing HVAC equipment, computer room air conditioners (CRACs), chilled-water systems, direct-to-chip liquid cooling, and immersion cooling. Unlike a monitoring-only interface, its operational DRL agent is designed to act on cooling settings and learn from measured outcomes. The combination of plant-side, data-hall, and liquid-loop control is central to Synapse's positioning.
Cooling Savings, Facility Efficiency, and Grid Flexibility
Cooling efficiency has a direct financial component: electricity avoided in pumps, fans, chillers, and related equipment can lower cooling operating costs while maintaining thermal conditions required by IT equipment. Synapse states a potential cooling-energy-cost reduction of up to 40% and describes an 8 to 14 month investment payback through energy savings. These are company-stated potential outcomes, with results determined by the characteristics of each facility.
Efficiency Benchmarks and AI Demand
24. Google's Omaha, Nebraska campus reported 1.05 trailing-12-month PUE in Q1 2026
Google's Q1 2026 campus table reports 1.05 trailing-12-month PUE for its Omaha, Nebraska facility. Its quarterly PUE was also 1.05. These measurements describe energy overhead at that particular U.S. site, not changes in computing productivity.
25. AI servers could represent 84% of U.S. server electricity in 2030
Under LBNL's U.S. reference case, AI servers account for 84% of server electricity and 55% of total data-center electricity in 2030. These percentages use different denominators and should not be treated as interchangeable.
26. Limited load curtailment could accommodate 76 GW of additional U.S. demand
A grid-flexibility study summarized by ACEEE models accommodation of 76 GW of additional load if eligible new loads can be curtailed for an average of 0.25% of operating hours. The study covers balancing authorities representing approximately 95% of U.S. peak demand. The estimate concerns grid accommodation through flexible consumption, not newly built generation.
27. Modeled U.S. average PUE could reach 1.36 in 2030
LBNL's U.S. reference case projects aggregate PUE of 1.36 in 2030, compared with 1.45 in 2024. This national result reflects projected shifts in the data-center fleet and equipment mix, rather than a predicted PUE for every individual facility.
Sustainability, Water Use, and Waste Heat
Reducing electricity required for cooling can support energy-efficiency and sustainability programs. Water and waste-heat outcomes depend on cooling design, local conditions, and heat-reuse infrastructure, which is why these indicators should be tracked separately from PUE.
European Sustainability Policy
28. Recovering half of European waste heat could serve 4 million households
The European Commission describes waste-heat recovery from roughly half of European data-center waste heat as having the potential to meet heating demand equivalent to 4 million households. This is an illustrative potential, not heat already delivered through district-heating networks.
29. Initial EU sustainability labels are expected in 2027
The European Commission expects the first sustainability labels in 2027. In September 2026, the Commission adopted a delegated act establishing a common data-center sustainability rating scheme. As of October 15, 2026, the act was subject to scrutiny by the European Parliament and the Council and had not yet entered into force.
30. European respondents' water-use tracking rose 5 percentage points in 2025
Uptime Institute's 2025 survey identifies a 5 percentage point increase in the share of European respondents collecting water-usage data. The statistic is specific to Europe, even though the survey report also covers other regions.
31. An illustrative 100 MW U.S. facility could use 2 million liters daily
A Congressional Research Service report cites an illustrative estimate of 2 million liters per day for direct and indirect water consumption at a 100 MW U.S. data center. Approximately 725,000 liters per day, or 190,000 U.S. gallons, would be consumed on-site. The full 2 million liters is not a cooling-tower or on-site water figure alone.
AI Servers and High-Density Thermal Management
AI server configurations can impose sustained and rapidly changing thermal loads. Effective data-center cooling therefore requires coordination among equipment operating limits, workload patterns, air management, and liquid cooling loops. Synapse's DRL strategy addresses these connected variables through predictive cooling control rather than isolated, static adjustments.
Server Electricity and Utilization
32. Modeled U.S. 8-GPU servers exceed 4 kW by 2026
LBNL's reference model estimates that average annual operating power for 8-GPU AI server configurations exceeds 4 kW by 2026. This is per modeled server, not an average rack power measurement.
33. Modeled 4-GPU and 2-GPU servers exceed 2 kW and 1 kW, respectively
LBNL's projected average operating-power trajectories peak just above 2 kW and 1 kW for 4-GPU and 2-GPU AI server configurations, respectively. Configuration mix, utilization, and cooling architecture influence the thermal demand generated by each installation.
34. Networking's share of U.S. data-center electricity rose from 3.4% to 4.5% between 2018 and 2024
LBNL models networking equipment's share of U.S. data-center electricity rising from 3.4% to 4.5% between 2018 and 2024. It links part of this shift to increased networking requirements, including InfiniBand switch deployment. Networking electricity contributes to IT thermal loads alongside the computing servers it supports.
35. The U.S. reference model assumes 80% training-server utilization and 20% inference-server utilization in 2030
The LBNL U.S. reference model assumes 80% training-server utilization and 20% inference-server utilization in 2030. These percentages are model assumptions, not measured average utilization across U.S. deployments. Different workloads produce distinct electricity and heat-load patterns, reinforcing the value of adaptive cooling control.
Deployment, Infrastructure Integration, and Grid Planning
Data-center efficiency solutions must work within real equipment, facility operating parameters, and changing energy demand. Synapse positions its platform around existing-equipment integration: one cooling-focused DRL agent operating across supported air, chilled-water, direct-to-chip, and immersion systems, with on-premises inference rather than a cloud-dependent mission-critical control loop.
Interconnection and Flexible Demand
36. U.S. data-center grid-interconnection capacity could reach 148 GW in 2030
LBNL estimates that 148 GW of interconnection capacity would be required to serve U.S. data centers in 2030, assuming 50% average utilization. Grid-interconnection capacity is not equivalent to electrical generation capacity or annual electricity usage.
37. Modeled curtailment events average 1.7 hours
For a scenario limiting annual curtailment of eligible new loads to 0.25% of operating time, the study summarized by ACEEE estimates a 1.7-hour average event duration. This is a demand-flexibility modeling result, not an observed outage statistic.
38. A 0.5% curtailment limit could accommodate 98 GW of new load in the United States
The modeled grid-accommodation headroom for the United States reaches 98 GW of additional load with average annual curtailment of eligible new loads limited to 0.5% of operating hours. This finding pertains to flexible demand and should not be described as permanent reductions in data-center energy consumption.
39. A 1% curtailment limit could accommodate 126 GW of new load in the United States
The same U.S. grid-flexibility study models accommodation of 126 GW of additional load at an average curtailment rate of 1% of operating hours. These values describe a conditional grid-planning scenario, separate from the operational cooling-energy savings targeted by Synapse.
Synapse Thermal's Integrated Deployment Approach
Techstars-backed Synapse brings several capabilities together in one cooling-focused architecture, supported by the company's data-center engineering and EPC relationships:
- Autonomous DRL control: Real-time, closed-loop adjustment of cooling operations, guided by thermal predictions and energy objectives.
- Comprehensive cooling coverage: Coordinated optimization across air cooling, chilled-water plants, direct-to-chip liquid cooling, and immersion cooling.
- On-premises edge inference: Local control decisions and data privacy benefits, with no cloud dependency in the mission-critical control loop; Synapse describes microsecond inference latency.
- Hardware-agnostic integration: Use of existing HVAC equipment, CRAC units, and liquid-cooling infrastructure without wholesale replacement.
- Defined commercial targets: Synapse states deployment in approximately 4 to 6 weeks, potential cooling-energy-cost savings of up to 40%, and an 8 to 14 month energy-savings payback proposition.
- Facility-specific evaluation: Synapse offers a complimentary cooling analysis centered on prospective efficiency and savings opportunities.
This combination provides a direct route from thermal measurement to predictive action, especially for operators managing varied cooling equipment and evolving AI workloads across U.S. and European facilities.
Frequently Asked Questions
What is the current average PUE for U.S. data centers?
LBNL estimates a modeled U.S. average PUE of 1.45 in 2024, down from 1.55 in 2018. Specific operating campuses can have lower values: Google's Loudoun County, Virginia facility reported a 1.08 trailing-12-month PUE in Q1 2026. Population-level estimates and site-level observations represent different measurement bases.
How much energy does data-center cooling consume?
Cooling's share varies by location, equipment, facility design, and operating conditions. ACEEE provides an illustrative 15% cooling share for a data center with PUE of 1.25. LBNL attributes 31% of U.S. data-center electricity to the broader facility-infrastructure category in 2024, which includes cooling and electrical losses. Neither figure is a universal cooling savings percentage.
How does Synapse's DRL cooling optimization work?
Synapse's agent analyzes temperatures, workloads, equipment conditions, and power measurements to forecast thermal demand and select cooling actions. Its closed-loop control approach coordinates fan speeds, chilled-water flow, and supported liquid-cooling systems, adapting as operating conditions change. The platform uses on-premises inference for local decisions and integrates with existing cooling infrastructure.
What savings and payback does Synapse describe?
Synapse cites up to 40% potential cooling-energy-cost savings and an 8 to 14 month payback proposition through energy savings. The company also identifies 1.03 as a potential optimal PUE under favorable conditions. These describe potential optimization and commercial outcomes, not guaranteed performance values or already achieved PUE at a named facility.
Does Synapse require replacement of existing cooling systems?
Synapse describes hardware-agnostic integration with existing HVAC, CRAC, chilled-water, liquid-cooling, and immersion systems. Its stated deployment timeframe is approximately 4 to 6 weeks, with the system designed to coordinate supported equipment through an on-premises cooling-focused DRL architecture rather than require wholesale infrastructure replacement.
Why do GPU workloads increase cooling requirements?
GPU-rich AI servers create concentrated, variable thermal loads. LBNL models average operating power exceeding 4 kW for 8-GPU servers by 2026 and projects AI servers will use 55% of U.S. data-center electricity in 2030. Adaptive cooling control helps align equipment operation with this changing thermal demand.
How does on-premises inference support cooling operations?
Synapse places inference at the facility for local cooling-control decisions, reducing reliance on external network connectivity in the mission-critical control loop. This edge architecture complements its autonomous DRL agent, which predicts thermal conditions and updates supported cooling settings as facility loads and ambient conditions evolve.