With Illinois emerging as a national hub for data centers, boasting 222 facilities statewide and over 150 in the Chicago metro area, questions about their environmental impact are increasingly central to public policy. Amid the AI boom and growing computational demands, lawmakers are considering new regulations, including Senate Bill 2181, which would mandate that data centers report energy and water usage to ensure greater transparency.
We spoke with Lav Varshney, associate professor of electrical and computer engineering at Illinois, about his work in this space, including his contributions to Meta’s $1.2 billion DeKalb data center, and his thoughts on balancing economic development with sustainability through innovation in technology and policy.
Can you tell us a little bit about your involvement with the $1.2B data center Meta built in DeKalb, IL?
Varshney: Meta’s DeKalb data center was a significant project—around $1.2 billion in investment. We worked with them to reduce the embodied carbon in the data center’s physical structures. Concrete alone contributes about 8% of global carbon emissions, compared to 3% from aviation, so we used AI to develop low-carbon concrete formulations. Meta implemented our AI-designed concrete in their builds, cutting the carbon footprint in half. Our AI model itself was very small, so the “carbon ROI” was almost infinite. Now our approach is being adopted by all of the big computing companies through the Open Compute Project Foundation (including Amazon, Google, and Microsoft) as well as major cement/concrete manufacturers such as Holcim.
As both an AI researcher and engineer, how do you assess the trade-off between the growing computational demand of AI systems and the environmental footprint of data centers?
Varshney: The prevailing AI paradigm is based on “hyperscaling,” meaning bigger models and more data, which leads to enormous energy consumption. While training large models used to be seen as the primary energy cost, we’re finding that inference, or using the model to respond to prompts, is now consuming a greater and growing share of energy. This shift means we need new ways to evaluate efficiency, like comparing the energy used by an AI system to what a human would need to perform the same task. My research group is exploring how to optimize energy use between training and inference, which we hope can inform both industry practices and policy decisions.
Illinois is considering legislation (SB 2181) to require greater transparency in data centers’ energy and water usage. What kinds of metrics or standards would you recommend to ensure this data is meaningful and actionable?
Varshney: There is such little transparency, yet a comprehensive environmental profile is essential. Besides carbon dioxide and water usage, we should include metrics like sulfur dioxide emissions that contribute to acid rain, and overall global warming potential. Transparency through reporting is a light-touch regulatory approach already used in other industries like iron and steel. There is no reason data centers, which are essentially manufacturing “intelligence,” shouldn’t be held to the same standards. Importantly, policy should address not just energy, but water and other environmental dimensions. Further, to make the metrics more actionable, one may want decomposition into parts such as communication, computation, and information storage as a function of useful intelligent work, rather than a metric such as power usage effectiveness (PUE) that lumps everything together.
Given your work on responsible AI and information infrastructure, how can we design data centers and AI pipelines that are both efficient and sustainable, particularly in states like Illinois that are becoming major data hubs?
Varshney: The hyperscaling paradigm is just one path to AI. My group has developed an alternative called information lattice learning, which is 5,000x more energy-efficient than large language models like ChatGPT. It’s also more human-controllable, interpretable, and IP-friendly, meaning that it enables everyone to monetize their creativity rather having it scraped and stolen. Even within the hyperscaling approach, there are ongoing innovations in hardware and software efficiency. However, we must be cautious about the Jevons paradox – the idea that making a technology more efficient can actually increase total use. Technical efficiency must go hand-in-hand with smart policy.
Some developers argue that measuring individual tenant usage in multi-client data centers is overly burdensome. As someone who studies optimization and fairness in systems, is there a viable way to balance privacy and accountability?
Varshney: Technically, yes, this is a very solvable problem, and is very much like how payments are charged to different tenants. During my time at Salesforce, a company that uses multi-tenant cloud infrastructure, I saw firsthand how detailed instrumentation can enable usage decomposition. Colleagues at Illinois, like Ashlyn Stillwell, are already breaking down energy usage by household appliance. Similar methods can be applied to data centers to isolate the energy used by individual tenants without breaching privacy. On the policy side, this might be further linked to know-your-customer (KYC) requirements that may further be important for national security reasons.
Is there anything else you’d like to share about this topic?
Varshney: One emerging alternative to centralized data centers is decentralized AI computation. Projects like SETI@home showed that distributed computing via idle personal devices can be powerful. While centralized centers are more efficient right now, decentralized approaches might offer different trade-offs in terms of energy, security, and access. It’s an area worth exploring as AI continues to scale.
