For years, warnings circulated that the explosive growth of data centers would send electricity consumption soaring, threatening both power grids and climate goals. A major study published in the journal Science tells a very different story: computing output at global data centers increased roughly sixfold between 2010 and 2018, yet their total energy consumption rose by only about 6%.
What the Research Found
The study was a collaboration of five scientists from Northwestern University, Lawrence Berkeley National Laboratory, and an independent research firm, funded by the U.S. Department of Energy. Rather than extrapolating from demand curves, the researchers took a bottom-up approach—compiling detailed data on processors, storage hardware, networking equipment, software, and cooling systems across global data centers to estimate actual electricity consumption.
Their conclusion directly challenges frequently cited projections that data center energy use could triple or more over a decade. Lead author Eric Masanet described those projections as “simplistic extrapolations” that focus almost entirely on rising computing demand while ignoring substantial efficiency gains on the supply side.
The Cloud Migration Effect
The single biggest driver of improved efficiency is the shift from on-premises corporate data centers to large-scale cloud infrastructure. In 2010, the researchers estimated that 79% of data center computing was performed in smaller, traditional facilities—largely owned and operated by non-technology companies. By 2018, that picture had reversed: 89% of data center computing took place in larger, utility-style cloud data centers.
This migration matters because hyperscale cloud facilities operated by companies such as Google, Microsoft, Amazon, and Meta (then Facebook) represent the most energy-efficient computing environments on the planet. Their scale gives them both the economic incentive and the engineering resources to squeeze maximum performance from every kilowatt-hour spent.
How Hyperscalers Achieve Efficiency
Large cloud providers invest heavily across the full stack to minimize energy use per unit of computation:
- Custom silicon: Purpose-built chips optimized for specific workloads reduce wasted compute cycles and idle power draw.
- High-density storage: Modern NVMe and tiered storage architectures deliver more IOPS per watt than legacy spinning-disk setups.
- Virtualization and containerization: Virtual machines and containers raise server utilization rates dramatically compared to single-tenant bare-metal deployments.
- Ultrafast networking: Low-latency, high-bandwidth fabrics reduce the time—and therefore the energy—needed to move data between compute nodes.
- Precision cooling: Customized airflow management, hot-aisle/cold-aisle containment, free-air cooling in cooler climates, and proximity to large hydroelectric sources all reduce the energy overhead of keeping hardware within thermal limits.
Google stated that its data centers on average generated seven times more computing power in 2018 than they did five years prior using no additional electricity—a claim attributed to Urs Hölzle, a senior vice president overseeing Google’s data center technology at the time of the study’s publication.
A Brief History of Data Center Energy Forecasting
The study places its findings in historical context. From 2000 to 2005, data center energy use doubled. In 2007, the U.S. Environmental Protection Agency forecast another doubling between 2005 and 2010. In reality, researcher Jonathan Koomey estimated the actual global increase at around 56% for that period—already far below the EPA projection—and the new study with additional data revised that estimate downward further.
Koomey, a former Lawrence Berkeley scientist and independent researcher at the time of the Science publication, co-authored the new work. “We’re hopeful that this research will reset people’s intuitions about data centers and energy use,” he said.
Current Energy Footprint
As of 2018, data centers consumed approximately 1% of global electricity output—equivalent, by the researchers’ estimate, to roughly 17 million U.S. households. That is a meaningful figure, but the key point is that it has remained essentially flat even as the volume of computing has grown dramatically.
Looking Ahead: The Next Transition Phase
The researchers caution that the efficiency gains offsetting demand growth are expected to hold for only roughly three to four more years from the time of publication (i.e., through the early-to-mid 2020s). Beyond that window, the outlook is uncertain. Several emerging pressures could change the trajectory:
- AI and machine learning workloads: Training large models is computationally intensive and power-hungry at a scale not fully captured in 2018 data.
- Edge computing expansion: Distributing compute to thousands of smaller edge nodes may reduce per-task efficiency compared to centralized hyperscale facilities.
- Slowing hardware gains: As Moore’s Law matures, the cadence of transistor-level efficiency improvements is slowing.
To navigate this transition, the Science article recommends increased investment in energy-saving research, better measurement practices, and more transparent information sharing by data center operators worldwide. The authors describe the coming years as “a critical transition phase to ensure a low-carbon and energy-efficient future.”
Key Takeaways for DevOps and Cloud Engineers
- Migrating workloads from private on-premises data centers to public cloud infrastructure is not just a cost and agility decision—it generally reduces per-workload energy consumption.
- Right-sizing instances, improving utilization rates through containerization, and decommissioning idle resources all contribute directly to the efficiency gains the study highlights.
- Sustainability metrics such as Power Usage Effectiveness (PUE) and Carbon Usage Effectiveness (CUE) are increasingly published by major cloud providers and can inform vendor selection.
- The efficiency story may not hold indefinitely. Teams adopting GPU-heavy AI/ML pipelines should actively monitor and optimize energy-intensive training jobs.
Original reporting sourced from The Indian Express, February 2020, based on a study published in the journal Science. The study was led by Eric Masanet (Northwestern University / UC Santa Barbara) and included Jonathan Koomey and collaborators from Lawrence Berkeley National Laboratory.