Artificial intelligence (AI) continues to advance at incredible speeds and with unique consequences. One need only consider the attention DeepSeek has garnered for its ability to operate on approximately one-tenth the computational power required by comparable models. Regardless of all the challenges around the chips, just consider Jevon’s paradox. It posits as technology becomes more efficient, overall resource consumption may increase rather than decrease, since there’s a lower barrier to entry. By investing in compute infrastructure, we aim to create a more sustainable and prosperous future. But, what implications does an increasingly digital world have for the planet’s energy demands?
The surge in data center energy consumption
The magnitude of change in data center energy consumption has transformed over the past decade. It’s hard to fathom. From 2010 to 2020, electricity usage in data centers remained stable, even as data production surged exponentially. It seemed that enhancements in efficiency, such as improvements in information and communication technology (ICT) as well as power usage effectiveness (PUE), were successfully keeping pace with the growing digitalization demand.
However, 2020 marked the rapid run-up of energy consumption by data centers. It’s raising concerns about the electrical grid’s capacity to keep up. As the demand for artificial intelligence (AI) and cloud computing fuels the expansion of data centers, electricity usage is expected to skyrocket. Think about it this way…this growing demand taxes grid capacity and power generation, leading to concerns about reliability and sustainability. While the alarm is being raised everywhere, with hyperbolic headlines and lots of conversation, the challenge is more nuanced, and it’s regional.
We must create a better dialogue and smarter path to discussing solutions. We need collaboration to help data centers understand how to taper energy consumption and build further operational efficiency. We believe there is a path forward.
Correlating and analyzing data can provide us with industry insights
Jim Simonelli, SVP and CTO for Schneider Electric’s Data Center and Networks division, spearheaded a comprehensive analysis project in collaboration with the Data Center Research & Strategy group, where Victor Avelar and I sit. The goal? Examine potential paths to slow the growth. The outcome is our latest white paper, Bending the Energy Curve: Decoupling Digitalization Trends from Data Center Energy Growth.
So how does one “bend the curve?”
In this report, we analyze the evolving landscape of data center energy consumption from 2010 to 2030, highlighting this shift in energy consumption patterns, beginning in 2020. We quantify energy growth drivers. Then we establish a new metric for the information and communications technology consumption (ICT) – the ICT performance factor – for evaluating ICT efficiency improvements, and model scenarios that demonstrate how even modest efficiency gains can significantly impact future energy trajectories, by as much as 17% by 2030.
We break the analysis down into physical infrastructure and ICT energy, to show what it would take to reduce energy – to “bend the curve”, slowing down the growth in energy consumption. This approach expands the discussion to include infrastructure providers and ICT technology suppliers. It emphasizes the need for collaboration to address energy challenges.
Throughout the report, we make our work transparent, ensuring that even its complex aspects are simply described and understandable.
The results in a nutshell
We believe there is an opportunity for the industry to accelerate and amplify current plans to bend the curve earlier than current forecasts expect.
By targeting Power Usage Effectiveness (PUEs) of 1.2 instead of 1.3 for newly built data centers… and achieving PUEs of 1.3 instead of 1.44 for just 10% of existing data centers through modernization improvements to the power and cooling infrastructure… we can bring the weighted average PUE down from the current estimate of 1.43 to 1.38. This incremental reduction can reduce energy by 3.6% or 38 TWh in 2030. That’s a start… but PUE improvements alone can’t compensate.
ICT improvements, on the other hand, offer the opportunity for more dramatic reductions, making it the more influential lever. Here, the “ICT performance factor” depicts data center energy in TWh divided by the year’s data generation in zettabytes (a rational proxy for the world’s digitalization trend).
Based on industry-released data on growth and energy forecasts, the ICT performance factor is expected to improve from 0.92 TWh/ZB in 2020 to 0.45 TWh/ZB by 2030 (downward shifts reflect more efficient performance).
With a concerted effort across the industry, this could be further reduced to 0.34 TWh/ZB by 2030. Achieving this target is feasible by:
- the widespread adoption of accelerated computing platforms for artificial intelligence,
- employing software and dispatching algorithms for optimal resource utilization,
- continuing advancements in CPU, GPU, TPU, and memory performance.
With a concerted effort, we believe we can see a decrease in the forecasted data center energy consumption by 14.4%, or 153 TWh. Combined, this means we can potentially “bend the curve” by as much as 17%, and once again, de-couple the growing energy demand from data center and AI-driven growth.
Behind-the-meter generation reduces demand on utilities
While onsite power generation may not reduce overall energy consumption, it can further “bend” energy usage from the perspective of utility grid power. When data centers prioritize and implement “behind the meter” (BTM) power generation, they help reduce the strain on the electrical grid. Each megawatt of BTM generation installed results in a savings of 8.76 TWh. Here’s more on onsite battery energy storage!
There are now more sustainable options available for data centers, including solid oxide fuel cells powered by natural gas with carbon sequestration, solid oxide fuel cells utilizing green hydrogen, and gas turbines equipped with carbon capture technology. Furthermore, small modular reactors (SMRs) are anticipated to become a viable alternative in the near future.
Check out our proposal to “bend the curve” and pursue a new path forward. I also encourage you to check out Patrick Donovan’s report, The Looming Power Crunch, to dive deeper into grid constraints and options for alternative and prime power for data centers.
Add a comment