Addressing AI Energy Consumption: Why the EU Must Embrace Ecodesign for Software

Zuzanna Warso is the Director of Research at Open Future.

Digital information and communication technologies (ICTs) account for 3.5–4 percent of the world's carbon footprint and 6–7 percent of its electricity footprint. Both the energy consumption of ICT products and their longevity are inextricably linked to the software products that run on them.

In the upcoming EU ecodesign regulation, software is addressed in a couple of contexts. For instance, there is a requirement that software updates shall not worsen product performance – however, because "products" are defined as "physical goods," it is not treated as a separate product group. As a result of this wording, software products remain outside the scope of the proposal, falling through the gaps of the ecodesign requirements.

Given the impact of software on overall ICT energy consumption, the lack of legal requirements addressing sustainable software usage represents a considerable blindspot.

This issue is becoming increasingly urgent in light of growing concerns about the energy consumption of AI-enabled products. As AI continues to permeate our lives and the products we use, the lack of regulations governing sustainable software is poised to deliver a significant blow in the not-too-distant future.

The energy efficiency of ICT and the “Jevons' paradox”

The energy efficiency of ICT systems has increased significantly over the last few decades. Nonetheless, the ICT industry's overall environmental footprint has been steadily growing.

Several factors, such as the proliferation of devices and the growing supply of data-driven services (e.g., video streaming), have contributed to this.

These, in fact, are all symptoms of what has been known as "Jevons' paradox." The paradox occurs when an increase in the efficiency of resource use leads to an overall increase in the consumption of that resource rather than a decrease.

While on the individual level, the energy use caused by ICT products and services may decrease, the aggregate use increases.

This is also due to the fact that products are becoming less durable, often as a result of planned obsolescence, including software-induced and software-related hardware obsolescence or the creation of artificial demands and needs through advertising. These processes lead to an increase in electronic waste.

As a result, ICT's total energy consumption and environmental impact, despite the energy efficiency gains, are increasing. And, with the integration in ICT services of AI components that consume significant amounts of energy (e.g. generative AI), this process will only accelerate unless there is a policy intervention.

How much energy does software use?

In essence, AI can be thought of as a type of software. Because AI systems rely on software components, algorithms, and data, they are part of the larger software ecosystem. As a result, when discussing AI energy efficiency regulation in the EU, it is important to consider it in the context of other measures aimed at ICTs and ecodesign, as they are interconnected and contribute to the overall energy consumption and environmental impact of digital technologies.

The simple answer to the question of how much energy software uses is that “it depends.” It is contingent on a number of factors, including how the software is designed and how it is used. Different additional factors add to this complexity. The software's energy consumption is tied to the underlying hardware it runs on. Modern computing environments encompass a range of devices, all of which have different architectures, power management mechanisms, and energy profiles.

And, software rarely operates in isolation; it interacts with users, external services, and varying workloads. Energy consumption can fluctuate based on factors such as user input, network activity, and data processing demands. Consequently, capturing the entire energy profile requires accounting for these dynamic elements. This variability makes energy measurement challenging as the same software might consume different amounts of energy in different runs. To understand how much energy software uses, one must look at all these different interactions and factors. Despite these challenges, methods for carrying out software measurements to record energy consumption have been developed. Still, more research and further developments in this field are required.

The myth of “green growth”

Discussions about energy efficiency for decades have been held against the dominant economic paradigm of continual economic growth (as measured by GDP). More recently, under the pressure of ecological collapse, the narrative of economic growth has morphed into “green growth.”

The concept of so-called "green growth" has had a significant impact on the EU's environmental policies. Green growth is the idea of an absolute decoupling of economic growth from resource consumption, including the use of fossil fuels. According to green growth theory, continued economic expansion is compatible with the ecology of our planet because technological change and substitution will allow us to completely decouple GDP growth from resource use and carbon emissions. However, ecological economists agree that combining GDP growth with greenhouse gas reduction is, at the global scale, not possible. In other words - “green growth” is a myth.

The AI Act

The issue of high energy consumption associated with AI has been receiving increased attention. As far as the EU regulation is concerned, the European Parliament’s position on the AI Act recognized that AI systems could have a large important environmental impact and high energy consumption during their lifecycle.

In terms of specific requirements, these are limited, which results from the intention to limit the AI Act’s scope only to certain types of AI systems. Initially, the AI Act aimed to oversee those AI systems deemed "high risk." However, with the broadening inclusion of “foundation models”, the proposal's scope has expanded. Nevertheless, it does not encompass all varieties of AI systems.

According to the text adopted by the European Parliament, the AI Act sets out requirements for so-called "high-risk AI systems." These systems must be designed and developed with logging capabilities that enable the recording of energy consumption, the measurement or calculation of resource use, and the environmental impact throughout the system's lifecycle. These requirements primarily focus on transparency, ensuring that stakeholders have access to data on energy consumption. However, it is important to note that, in this case, the AI Act does not compel measures to reduce the energy consumption of AI systems.

On the other hand, when it comes to "foundation models," the requirements are expected to be more robust. As per the Parliament's amendments, in addition to capabilities for measuring and logging energy and resource consumption and other environmental impacts where feasible, these models should be designed and developed in line with applicable standards aimed at reducing energy use, resource consumption, and waste. These standards are also intended to enhance the energy efficiency of the system. It is worth pointing out, however, that these obligations are set to take effect later when the “applicable standards” are developed.

In terms of putting these measures into practice and ensuring compliance, the Commission is tasked with developing guidelines. These guidelines are supposed to provide instructions on the methods for measuring and logging that enable the calculation and reporting of a system's environmental impact, including factors like carbon footprint and energy efficiency.

While the Parliament’s text recognizes the importance of transparency and environmental considerations in AI development, particularly for high-risk AI systems and foundation models, the implementation and enforcement of these requirements will depend on the development and availability of comprehensive guidelines and the specific timeline outlined in the legislation. Development of these guidelines will not be an easy task. A 2019 review of approaches to estimate energy consumption in machine learning revealed challenges to developing a common strategy for energy evaluation related, for example, to the rapid changes in neural network designs, implementations, and hardware.

Ecodesign regulations

The EU is currently reviewing its legislation on the ecodesign of products. In March 2022, the European Commission published a proposal for a new Ecodesign for Sustainable Products Regulation (ESPR). It dubbed it “the cornerstone of the Commission’s approach to more environmentally sustainable and circular products.”

The proposal for the ESPR establishes a framework to set ecodesign requirements for specific product groups to improve their circularity, energy performance, and other environmental sustainability aspects. The framework will allow for the setting of a wide range of requirements, including carbon and environmental footprints. The rules proposed under ESPR will apply to all products placed on the EU market, whether produced inside or outside the EU.

However, as highlighted in the introduction, software products largely fall beyond the purview of EU ecodesign policies. This gap will only see an incomplete rectification through the requirements set to be introduced by the AI Act, pending the adoption of the proposed amendments in the Parliament's text, due to its limited scope and the current lack of common standards.

Moving forward: recommendations

Given these ongoing and future developments and a growing recognition of an urgent need for “green digital policies,” two specific and one more general recommendation can be made.

First, the EU institutions should provide funding for additional research and the creation of methods of estimating software energy consumption and enhancing software sustainability while ensuring that this work is not done in isolation and brings the discussion on the sustainability and ecodesign of software and AI systems together.

Second, building on new and existing research, they must develop and enforce software-related energy efficiency requirements that incorporate recommendations made in the context of measures under the ecodesign regulations and the AI Act. This second recommendation elaborated the call to the Commission to prioritize ICT in the ecodesign work plans.

The call to establish requirements for software-driven energy efficiency, energy consumption reduction, and digital sustainability is not new, but with the rise of AI products and services, it has gained more importance. The preparatory studies for the Ecodesign Working Plan 2022-2024 recommended a feasibility study on the possibility of setting energy and resource efficiency measures on application software. The 2023 JRC study suggested that energy and resource efficiency measures on “application software” could be implemented in the form of minimum requirements, information requirements, and/or mandatory labeling. The study also recognized that research in this field that would lead to developing standardized approaches is necessary.

What has been missing from these analyses is an explicit recognition of the impact the proliferation of AI will have on the energy consumption of ICT, which puts more urgency on this call. The energy consumption of AI systems and the sustainability of software are both intertwined aspects of the larger effort to make digital technologies more environmentally friendly. Addressing both issues is critical for lowering the information technology sector's carbon footprint.

Improving energy efficiency alone will not be enough to offset the rapid expansion of ICT and the associated environmental challenges. Nonetheless, the EU should not give up on its efforts. The debate about reducing software systems’ impact on the environment must be, however, placed in a broader context, going beyond issues related to energy consumption. Because of the “Jevons' paradox” and the risk of the rebound effect, in order for energy efficiency measures to result in the reduction of total energy consumption, they cannot function in isolation, hence the third, more general recommendation.

Instead of relying on and propagating the myth of green growth, the EU's digital policies must be rooted in alternative narratives, such as digital sufficiency and digital sobriety, to navigate the challenges posed by the rapid expansion of ICT and its environmental and societal implications. These strategies emphasize a balanced and sustainable use of digital resources, providing a more realistic path to a livable future.