Artificial intelligence is powering breakthroughs in everything from health care to climate science, but each new discovery comes with a cost: significant energy. Now, researchers at the University of Wisconsin-Madison are asking a new question — how can this powerful technology be more sustainable?
Across campus, experts in data science and AI literacy are working together to ensure innovation is environmentally responsible.
“AI projects at UW-Madison range from foundational large-scale deep learning models to domain-specific ones that address humanity’s most pressing challenges,” Song Gao, director of the Geospatial Data Science Lab at UW-Madison and an expert in Microsoft’s AI for Earth, told The Daily Cardinal via email. “These include life-saving disaster response, sustainability solutions, digital agriculture and health and medical challenges.”
Gao’s work sits at the intersection of AI and Earth science, using “GeoAI” to analyze environmental data, predict natural disasters and track climate change. But those breakthroughs come with high energy costs — training large models can consume thousands of hours of computing time, significant electricity and graphics processing units (GPU) and powerful processors that handle intensive generative AI workloads.
To handle this demand, UW-Madison has expanded its research computing infrastructure with help from the Wisconsin Alumni Research Foundation. The Data Science Institute (DSI), the Center for High Throughput Computing, DoIT Research Cyberinfrastructure, the School of Medicine and Public Health and the Data Science Hub all received upgrades designed to support large-scale model training and storage.
A new RISE-AI Collaboration HQ initiative has also been introduced, which plans to bring in new investments, hire 35 new AI experts and hold events and seminars on AI and AI literacy.
“The RISE-AI Collaboration HQ will connect both new and established campus researchers to this increased capacity for processing and data storage,” Gao said.
At the DSI, researchers have access to some of the most powerful GPUs — Nvidia H100 and A100 chips — enabling high-speed model training and data analysis. Gao’s team is also working to mitigate AI’s environmental toll.
“At the Geospatial Data Science Lab, researchers are developing benchmarks and metrics used to evaluate the geographic biases and environmental impact of deep learning models and AI foundation models for Earth science,” he said. “When evaluating the environmental cost-benefit trade-offs, we need to consider the geographic regions that benefit from the AI model and those that bear the environmental costs.”
Bonnie Shucha, associate dean for Library and Information Services and a specialist in AI literacy, said sustainability also depends on how people use AI tools.
“Every interaction with GenAI requires computational resources and energy,” Shucha said. “According to Google, a typical text Gemini prompt consumes 0.24 watt-hours of electricity — equivalent to running a standard microwave for about one second. While that doesn’t sound like a lot, the impact can multiply quickly.”
Shucha said vague or poorly-structured prompts can dramatically increase computational load.
“If someone asks, ‘explain copyright infringement,’ and then spends the next 20 minutes clarifying that they actually want to know whether using a song clip in their student film project falls under fair use, that’s significantly more computational load than taking 30 seconds upfront to craft a clear, well-structured prompt,” she said.
Shucha argues that efficient prompting isn’t just a productivity skill, it’s a sustainability practice.
“When I talk with students and others about GenAI, I emphasize that thoughtful prompting is both more effective and more environmentally friendly,” Shucha said. “The same skills that make you a better AI user — critical thinking, preparation and asking focused questions — also reduce your energy use.”
She also said text generation is more energy efficient than an image or video. For Shucha, sustainability should be embedded in how universities teach AI.
“Students and professionals need to understand that GenAI use has numerous ethical implications, including its environmental impact,” she said. “Including sustainability in AI ethics education helps users become more intentional about when and how they use AI rather than treating it as an unlimited resource.”
Both Gao and Shucha see UW-Madison as uniquely positioned to lead by example. With cutting-edge research facilities and a strong emphasis on public responsibility, the university can model what sustainable innovation looks like.
“Universities have a unique opportunity in teaching students to use AI ethically and effectively,” Shucha said. “When we teach students to use AI tools, we can integrate sustainable practices into that training — not as a separate topic, but woven into how we teach AI literacy.”
