AI sustainability – why “green” data centers aren’t enough
Ecosystem thinking will be critical to long-term AI sustainability
The ethical and sustainable use of AI are sure to be thoroughly debated topics amongst global leaders gathering in New York this week for the 80th session of the United Nations General Assembly. The resource intensity of AI is a shared concern across private and public sectors alike—a challenge the world continues to grapple with nearly 3 years after Gen AI first burst onto the global stage in late 2022.
We should take heart at the numerous innovations that have already emerged to make these behemoth workloads more efficient, but there is still much work to do. Understandably, the headlines scrutinizing AI’s sustainability are primarily focused on “green” data centers—facilities powered by renewable energy, optimized for cooling and designed to maximize efficiency. Yet, they are only one piece of a much larger puzzle. The reality is that, in order to build more sustainable AI ecosystems, we need to look beyond the infrastructure and embrace a broader strategy that seeks to maximize the use of low carbon energy sources while incorporating wider IT sustainability elements like data and software efficiency.
Beyond the Data Centre: Expanding the AI Sustainability Conversation
Undeniably, modern data centers are foundational to AI deployment. Innovations in liquid cooling, energy conversion, and rack-level optimization have made it possible to support the immense compute demands of AI workloads. It’s worth noting that these advancements are not only necessary, they are critical. Yet even the most advanced facilities face limitations. Just AI inferencing alone, the process of running trained models to generate predictions or responses, is projected to consume up to 20% of global energy by 2030. That’s not just a data centre problem; it’s a business risk.
The reality is that infrastructure improvements are vital but not sufficient. AI’s sustainability challenge is systemic, spanning the entire lifecycle of an AI model—from data collection and training to deployment and monitoring. Focusing solely on the data centre rather than the whole system is like improving your gym routine while maintaining an unhealthy diet—you may move the needle, but you won’t maximize your results.
Having a clear data strategy not only ensures better AI model output - it can also dramatically improve the energy efficiency of the model itself.
Efficiency Starts with Data
Good AI models rely on the input of good data. Having a clear data strategy not only ensures better AI model output - it can also dramatically improve the energy efficiency of the model itself. Yet, many organizations rush into AI deployments without a clear data strategy. The result? Models trained on irrelevant, unstructured, or low-quality data that not only underperform, but also waste valuable compute resources.
A 4-step framework for data efficiency—Collect, Curate, Clean, and Confirm—can help ensure the use of only relevant, high-quality data, in turn reducing the energy and time required for training and inferencing. For example, curating datasets to eliminate bias and redundancy has been shown to significantly improve model accuracy while lowering resource consumption.
Software: The Hidden Opportunity
Software efficiency is unfortunately often overlooked in AI sustainability conversations. Most developers are trained to prioritize functionality and speed, and assume unlimited resources. But resources are not unlimited, and inefficient code and bloated models dramatically increase energy consumption.
The good news is that a shift from “bigger is better” to “fit-for-purpose” is already underway. Many successful tactics are emerging in the developer community. Quantization, for example, lowers computation load by reducing model precision when extreme accuracy is not critical. Using guardrails is another useful technique that reroutes simple queries to lightweight models that use less energy. In addition, AI developers are making strides by using Small Language Models (SLMs) and lighter, domain-specific LLMs to achieve significant efficiency gains.
Equipment Efficiency: Matching Workload to Hardware
Just as software must be optimized for performance and sustainability, so too must the hardware it runs on. Equipment efficiency in AI deployment means using each piece of technology to do the most work possible with the least energy consumption and heat generation. AI workloads, therefore, need to be matched with hardware platforms specifically designed for those tasks. This requires enterprises to resist the existing tendency to significantly overprovision “just in case”, leading to lower utilization rates and unnecessary energy use. By aligning workloads with optimized hardware usage rates, organizations can unlock substantial efficiency gains.
A Holistic Framework for AI Efficiency
Business leaders must champion a new way of thinking that expands beyond device-level optimization: one that starts with big-picture thinking and deploys a comprehensive roadmap for AI deployment that considers optimization on a solution-level. This involves a significant cultural shift that may take time but cannot be ignored. Developing a holistic framework for AI efficiency necessitates upskilling teams, fostering cross-functional collaboration, and embedding sustainability into AI governance.
By broadening our field of vision to evaluate the overall AI ecosystem, rather than the underlying hardware alone, we stand a better chance of addressing AI’s sustainability challenges…
By broadening our field of vision to evaluate the overall AI ecosystem, rather than the underlying hardware alone, we stand a better chance of addressing AI’s sustainability challenges in time to make a difference to its environmental trajectory. By embracing data and software efficiency, optimizing equipment and resources, and aligning energy strategies with business goals, organizations can build powerful AI systems that take into consideration the energy constraints facing the world today, and position us to reap the benefits this new technology promises to deliver.
