Generative AI is surging forward, but its rapid growth comes at a cost: mounting energy demands and deeper carbon emissions. A recent MIT article flags how data centers, critical to training and running large AI models, are expected to more than double global electricity consumption by 2030, reaching about 945 terawatt-hours. Much of that increase may come from fossil-fuel sources, driving a projected 220 million tons of additional carbon emissions.
MIT researchers argue that addressing these climate impacts requires more than incremental tweaks. They frame emissions in two buckets. First is the operational carbon, i.e., the electricity AI systems consume during training and inference. Second is embodied carbon, i.e., the emissions tied to constructing data centers, mining materials, and manufacturing hardware.
On the operational front, there’s room for optimization. Lowering GPU power draw to about 30% of full usage, for example, can cut energy use substantially without hurting performance. Stopping model training early, rather than pressing for marginal gains in accuracy, also saves energy. Further, smarter algorithms, modular architectures, network pruning, and other innovations are compressing models and reducing wasted computation.
Flexibility in timing and location also plays a role. Researchers propose scheduling computation during periods when grids are greener, or placing data centers where renewable energy is abundant or local climates reduce cooling demands. Storage systems, workload pooling across clients, and balancing deployment versus training tasks can also shift more operations toward cleaner periods.
MIT’s researchers also introduced the Net Climate Impact Score, a tool to help decision-makers weigh AI’s environmental costs against its potential benefits. Ultimately, the article underscores that tackling AI’s climate impact demands cooperation, among engineers, companies, regulators, and academia, and faster adoption of sustainable practices. As Jennifer Turliuk, one of the authors cited, puts it: “Every day counts.”