AI Can Move Beyond Data Centers
Amid growing concerns about AI energy and resource use, three scientists from the Eric and Wendy Schmidt Center for Data Science & Environment (Schmidt DSE) at the University of California, Berkeley and the University of Colorado Boulder propose a practical solution to reduce reliance on data centers. In a commentary published today in Nature the scientists posit that open source AI models, which tend to be far more energy efficient than their commercial counterparts and run for free, offer an outside-the-box approach to meeting demands for use.
"Proposals to hurl data centers in space may appeal to those concerned about limited resources here on Earth," said Cassie Buhler, a postdoctoral fellow at the Cooperative Institute for Research in Environmental Sciences (CIRES) and the Environmental Data Science Innovation and Impact Lab (ESIIL) at the University of Colorado Boulder, and an affiliate of Schmidt DSE. "But data centers are an incredibly resource-intensive way to deliver AI. Instead, researchers can opt for alternatives that run on your laptop.”
Open source AI models can be downloaded onto a personal computer and run entirely on that same machine. By definition, “open source” tools are publicly available for anyone to view, modify, and distribute. These models are free for the user, allowing full control of their data. No data center or third-party company is required for access. According to the researchers, open source AI models are more often performing at pace with leading commercial models such as ChatGPT and Claude. Because open models run on a local machine, open source models may use 100 times less energy than commercial counterparts.
"For many tasks, open source AI models that can run on laptops or desktops can be just as effective as the most popular proprietary systems from a year ago. And, when paired with the right tools for the scientists' needs, they can be even more effective while maintaining the integrity and privacy of the research data," said Fernando Pérez, Faculty Co-Director of Schmidt DSE and Associate Professor of Statistics at UC Berkeley. "Scientists have led the adoption of open source tools, from the internet to programming languages like Python. Widespread use of open source AI models will require more investment from academic institutions but we must pursue this more sustainable and ethical path.”
Since its inception in 2022 Schmidt DSE has co-developed nine open source, decision-support tools for policymakers and scientists. The team is increasingly pivoting to developing local AI tools for stakeholders, including tools for the United Nations (UN) to help reduce global greenhouse gas pollution and establish new marine protected areas. The latter, called the Geospatial LLM-Enabled Navigator (GLEN) High Seas Explorer compiles and visualizes data on existing protected areas, ocean floor geology, biodiversity, and global fisheries activity and links this data to UN policy text. With the chatbot, users can ask questions like, “if I created a new marine protected area that was 100 kilometers wide in the high seas right next to my country’s border, how many undersea mountains and how many whale migration pathways would be included?” and will receive scientifically-rigorous answers in minutes.
"Many climate scientists have avoided AI altogether because of the environmental and privacy costs of relying on data centers, and we share those concerns," said Carl Boettiger, faculty advisor for Schmidt DSE and Associate Professor of Environmental Science, Policy & Management at UC Berkeley. "Yet when AI is used responsibly, we also recognize that it can be a powerful tool for solving problems. Let’s be smarter about how we use it and that should include leveraging local, open source models.”
Today the accelerated growth of data centers is not environmentally or economically sustainable. At the same time, the commentary notes a clear trend in data science: within a year or two, AI infrastructure that once required specialized computing capabilities will be increasingly accessible on consumer-grade hardware. In late July Microsoft’s open letter to the tech industry called for more use of AI models with “open weights,” which can help democratize access to these tools by giving users more control over them. From the commentary authors’ perspective, this statement is beneficial for scientific innovation, user privacy, and for increasing competition in the tech economy. Over 200 industry giants signed the letter including NVIDIA, Google, Amazon, and OpenAI.
“Scientific progress can, in fact, rapidly advance beyond data centers,” continued Boettiger. “Local, open source models are key to a more sustainable future for AI in science.”
