By Holt Hackney
The modern technology stack – cloud computing, artificial intelligence, advanced analytics, and high-performance computing – rests on a rapidly expanding network of data centers that now function as critical U.S. infrastructure. Over the past two decades, the growth of hyperscale and high-density compute facilities has driven unprecedented demand for electricity, cooling capacity, and water resources. With AI workloads accelerating power densities and stressing grid and thermal limits, data center efficiency is no longer a secondary optimization problem, it is a first-order systems challenge.
Lawrence Berkeley National Laboratory (Berkeley Lab) has played a central role in quantifying, modeling, and addressing these challenges. Through applied research, technology development, and close collaboration with hyperscalers, chip designers, utilities, and grid operators, Berkeley Lab, and other labs, are helping define how next-generation data centers are designed, cooled, powered, and integrated with the electric grid. Its work spans macro-level energy forecasting to chip-scale physics, providing both strategic insight and operational tools for an industry under rapid transformation.
Below are seven key areas where the Lab is influencing the reliability and efficiency of U.S. data centers.
Quantifying growth and forecasting system-level impacts
Berkeley Lab hosts the U.S. Department of Energy’s Center of Expertise for Data Center Energy (CoE), which serves as a national resource for data-driven analysis of data center energy use. A cornerstone of this effort is the United States Data Center Energy Usage Report, updated in December to reflect industry growth from 2014 through projected conditions in 2028.
The updated analysis confirms what operators and utilities are already experiencing: U.S. data center electricity consumption nearly tripled between 2016 and 2023. Under current trajectories, data centers could account for up to 12 percent of total U.S. electricity demand by 2028. These projections have major implications for capacity planning, transmission infrastructure, and grid reliability—particularly in regions experiencing rapid hyperscale clustering.
The Lab continues to refine these models using industry feedback and updated operational data, moving toward more frequent reporting cycles. For utilities, regulators, and large operators, these forecasts provide a common analytical baseline for load forecasting and infrastructure investment decisions.
Engineering energy efficiency into AI-scale supercomputing
At the facility level, Berkeley Lab has demonstrated how architectural and cooling choices can materially reduce the energy intensity of high-performance computing. At the Department of Energy’s National Energy Research Scientific Computing Center (NERSC), the Lab has partnered with vendors to deploy systems designed explicitly around efficiency constraints.
NERSC’s Perlmutter system, operational since 2021, and its successor Doudna, scheduled for deployment in late 2026, both rely on direct-to-chip liquid cooling combined with ambient air heat rejection. Rather than using traditional compressor-based refrigeration, the facility leverages outside air and cooling towers to reject waste heat directly to the environment.
A focused, two-year optimization initiative reduced non-IT power consumption—cooling, power distribution, and supporting infrastructure—by 42 percent. The result was annual savings exceeding 2 million kWh of electricity and roughly 500,000 gallons of water, with operating cost reductions of approximately $200,000 per year. These results provide a validated blueprint for other HPC and AI-focused facilities seeking to control non-compute energy overhead.
Advancing chip-scale efficiency and energy storage
While facility optimization delivers near-term gains, the Lab is also addressing efficiency at the silicon and materials level. Researchers have led advances in transistor architectures, energy storage components, and optoelectronic systems aimed at reducing energy per operation.
Recent work includes the development of microcapacitors with ultrahigh energy and power density, enabling more effective on-chip energy buffering and potentially reducing losses associated with power delivery and voltage regulation. Other research has demonstrated optoelectronic approaches that convert photons directly into information rather than images, bypassing energy-intensive signal processing pipelines used in conventional imaging and sensing workloads.
In parallel, Berkeley Lab has developed an open-source 3D simulation framework that models the atomistic behavior of electronic materials. This tool enables faster and lower-cost exploration of novel chip designs by linking physical phenomena at the materials level to device performance, providing industry with a powerful platform for accelerating energy-efficient chip development.
Establishing liquid cooling standards for AI workloads
As rack power densities climb well beyond what air cooling can reliably support, liquid cooling has shifted from niche deployment to core infrastructure strategy. However, broad adoption requires standardization across fluids, materials, operating conditions, and mechanical interfaces.
The Lab has collaborated with industry partners to develop specifications for liquid-cooled server racks and cabinets, addressing cooling architectures down to the chip level. Working with the Energy Efficiency High-Performance Computing Working Group – led by Lawrence Livermore National Laboratory – the Lab helped define standards that reduce vendor lock-in and enable interoperability.
These specifications include industry-standard guidance for liquid transfer fluids and system materials, later refined in collaboration with the Open Compute Project and issued as formal guidelines. As AI systems push thermal envelopes further, these standards are becoming foundational to scalable, reliable deployment.
Driving operational efficiency through diagnostic tools
Beyond system design, Berkeley Lab has developed practical tools that enable operators to quantify inefficiencies and prioritize upgrades. Through the CoE, the Lab provides a suite of diagnostic and assessment tools widely used across public and private data centers.
These include the DC Pro Tool for benchmarking overall energy performance, the Air Management Tool Suite for evaluating airflow and thermal containment strategies, and the Electrical Power Chain Tool for analyzing efficiency losses in UPS systems and power distribution architectures.
Operators using these tools have achieved targeted improvements such as refined airflow management, upgraded CRAH controls, and the adoption of thermosyphon-based cooling. Reported outcomes include approximately 8 percent reductions in cooling energy use and annual water savings exceeding 1 million gallons, along with improved fault tolerance and operational flexibility.
Optimizing performance with advanced simulation platforms
Simulation has become an essential capability for managing the complexity of modern data centers, and the Lab’s tools are widely deployed in industry. Meta uses the Lab’s Modelica Buildings Library—an open-source collection of dynamic simulation models—to optimize energy and water use across its facilities. Carrier applies Modelica models to operate co-located data centers and design cooling systems for hyperscale environments.
The Lab is also collaborating with the University of Maryland on MOSTCOOL (Multi-Objective Simulation Tool for Cooling Optimization and Operational Longevity), developed under the ARPA-E COOLERCHIPS program. MOSTCOOL integrates power, thermal, and reliability modeling to optimize data center design and operation, using the EnergyPlus engine to simulate cooling systems and waste heat recovery strategies.
Preparing the workforce and infrastructure for sustained growth
Finally, Berkeley Lab is supporting long-term industry readiness through education, standards development, and stakeholder engagement. Its Data Center Energy Practitioner Training Program equips engineers and operators with current best practices across IT equipment, air management, cooling, and electrical systems, with curriculum updates informed by ongoing industry consultation.
In October, nearly 150 industry professionals participated in a the Lab–hosted listening session with BP Castrol at the 2025 Open Compute Project Global Summit. Discussions focused on high-density power delivery, advanced cooling for AI hardware, and evolving equipment selection practices. The insights gathered will feed directly into updated national energy models and future research priorities.
As AI and cloud workloads continue to scale, Berkeley Lab’s work is helping ensure that the physical infrastructure underpinning the digital economy evolves with equal sophistication – balancing performance, efficiency, reliability, and grid integration at every level of the stack.