Aligned Incentives’ Chief Strategy Officer, Dr. Sarah Boyd, spoke at the National Academies workshop “Implications of Artificial Intelligence-Related Data Center Electricity Use and Emissions” on November 13, 2024. The workshop brought together leading experts from the computing and electricity industries, research institutions, and the regulatory community to examine data centers’ energy and environmental impacts and explore effective mitigation strategies.
Sarah presented in the “Sustainability Analysis of Data Centers” session, joining experts from Google, IEA, Clean Virginia, and the Piedmont Environmental Council. She highlights the carbon footprint of semiconductor manufacturing and upstream processes preceding data center energy use, and explores potential technical and policy solutions to address these challenges. This blog summarizes Sarah’s talk. Watch the session replay here.
As AI adoption accelerates, so does its environmental impact. While operational energy use in data centers gets much attention, the upstream environmental footprint, particularly the greenhouse gas (GHG) emissions from the production of data center infrastructure—known as embodied emissions—is becoming a critical factor.
Recent studies highlight that server compute, including CPUs/GPUs, memory, and storage, constitute most of a data center’s embodied emissions.
Using a scalable product footprinting solution like AITrack, you can analyze all tiers of a server’s supply chain and pinpoint the most significant sources of emissions. For hyperscalers such as Google, Amazon, and Microsoft, semiconductor fabrication in the upstream supply chain can account for the majority of datacenter GHG emissions for standard datacenter compute hardware.
AI increases the density of compute, and for a good reason: a system with tightly integrated CPUs and memory is more energy efficient, as less energy is spent powering thicker, longer signal lines between cores and memory. AI architectures integrate many AI accelerators, with densely packed CPU cores and memory, into close proximity to reduce energy losses at the interconnect level.
Driven by the exponential growth of AI, the demand for compute is skyrocketing. The growth in demand for semiconductor chips even before recent AI scaling already outpaced the scaling achieved by Moore’s Law, and AI has only accelerated demand.
By 2030, the carbon impact of manufacturing the advanced chips needed for AI training and inference could reach 20 to 50 million metric tons of CO₂e annually, equivalent to the emissions of a medium-sized country.
Currently, global wafer production stands at approximately 400 million wafers per year, but much of this is less advanced nodes from older manufacturing lines. AI compute requires advanced nodes, which require larger fabrication facilities and more energy – an advanced manufacturing fab typically consuming as much power as a small city.
In addition, advanced processing involves more process steps and, in some cases, higher flows of etch process gases, meaning higher flow rates of fluorinated greenhouse gas (f-GHG) emissions. To reduce potent GHG emissions from semiconductor manufacturing, it is critical that f-GHG abatement systems are installed and operating effectively on all relevant process equipment.
Addressing the environmental impact of AI compute requires solutions that become available over multiple timeframes.
Immediate solutions
Medium-term Solutions
Long-term Solutions
The need for the development of these medium- and long-term solutions highlights the need for committed research in these areas, as well as the integration of carbon analysis into semiconductor process research.
The imec consortium, a close collaborator with NIST, provides a leading example of pioneering research to reduce the carbon impact of semiconductor manufacturing by incorporating environmental considerations at the earliest stages of process development. Through collaboration with industry leaders, imec bridges research and application, driving significant progress toward low-carbon, next-generation semiconductors and system-level compute architectures.
Given the complexity and rapid evolution of the semiconductor and AI industries, smart policy is essential to driving action across global supply chains. Two prominent policy examples from the European Union are:
These policy frameworks advance transparency in reporting, allowing the right incentives to be set across complex value chains.
The cradle-to-gate environmental impact of AI compute presents a growing challenge, with semiconductor manufacturing at its core. By fostering collaboration, accelerating technical innovation, and implementing forward-looking policies, we can embed environmental principles into the fabric of AI compute, ensuring that technological progress aligns with sustainability goals.
Aligned Incentives provides an AI-powered sustainability planning solution that empowers cloud computing companies to:
Speak with our product experts to learn more 🡲
Sarah has spent her career using life cycle assessment (LCA) to drive decision-making in the corporate world. Sarah is Chief Strategy Officer of Aligned Incentives (now a part of Bureau Veritas) where she uses LCA at scale to identify granular carbon mitigation actions at the product and corporate level. Prior to joining Aligned Incentives, Sarah led product carbon footprinting for new product introductions at Apple. In her previous work in consulting, Sarah led over 100 product and corporate sustainability projects with clients across ICT, electrical utilities, aerospace, construction and consumer products sectors, and led data updates for the GaBi US LCA database for Sphera. Sarah holds a B.S. from Stanford University and M.S. and Ph.D. in mechanical engineering from the University of California, Berkeley. Her book “Life-cycle Assessment of Semiconductors”, published by Springer, has found broad use with sustainability practitioners in ICT looking to drive LCA-based decisions themselves.
