With governments worldwide racing to build artificial intelligence (AI) infrastructure, sustainability concerns are rising alongside the expansion of data centres. Priya Donti is an assistant professor at MIT and co-founder of Climate Change AI, who was also named among the 100 most influential people in AI in 2025. She speaks to Devansh Mittal ahead of the India AI Impact Summit about AI’s energy footprint, grid reliability, and how targeted AI systems can support climate goals.
With the AI boom, especially data centre expansion, sustainability hasn’t been central to the debate. How do you view this?
When people hear “AI,” they often think of large language models or chatbots. But AI is a broad and heterogeneous set of technologies, from rule-based systems to machine learning systems, or maybe task-specific systems. Some models, like the large language models (LLMs) of the frontier labs, have billions of parameters, while many others that are actually used across research and society are much smaller.
The AI used to optimise power grids is very different from the AI used in chatbots. The choices we make about which kinds of systems to build determine which applications benefit. So in the context of data centre growth, we should ask two questions: What are the local impacts — on electricity prices, water use and materials? And what is this AI development actually intended to achieve?
There’s an assumption that scaling current AI paradigms will automatically solve climate and health challenges. That isn’t necessarily true. Using AI effectively for climate requires targeted systems, not just larger data centres.
Are our electricity grids prepared for the massive energy demand from AI data centres?
There are two dimensions to this. First is total load growth. In the US, electricity demand was relatively stable for many years, so utilities aren’t accustomed to rapid increases. That has created challenges. Countries like India and China have more experience managing load growth.
Story continues below this ad
But the critical issue is whether that growth is aligned with climate goals. In the US, unexpected AI-driven demand has led to increased orders for natural gas turbines, which lock in emissions for years.
Second, data centres create highly variable demand profiles. During model training, power use can rise significantly, and when training ends, demand can drop sharply. Grid operators must balance supply and demand every second. These fluctuations add complexity to maintaining reliability.
Why do grid managers need to balance supply and demand?
At every moment, the amount of electricity supplied to the grid must equal the amount consumed. If supply and demand don’t match, it leads to outages. Traditionally, grids relied on coal, gas and nuclear plants, which are controllable. Demand, though not directly controlled, is fairly predictable in aggregate. With the share of renewables like solar and wind going up, generation depends on the weather. We can’t control how much sun or wind we get, so we must predict it. That means we need better forecasting and better tools to coordinate batteries, conventional plants and flexible demand to maintain balance.
Story continues below this ad
What changes as renewables increase? What contributions does your research make to it?
The system becomes more complex. There’s less controllable generation, more distributed resources, and less time to respond when imbalances occur. Faster optimisation and AI tools can help manage that complexity.
My work focuses more on control than forecasting. Once you’ve predicted renewable generation and demand, you need to decide how to charge batteries, how much conventional generation to dispatch, and how to coordinate flexible loads. Grid operators traditionally use optimisation-based methods. AI can help make these calculations faster and more scalable as systems grow more complex.
For someone who equates AI with chatbots, what is a concrete climate application?
Story continues below this ad
Solar forecasting is a good example. To integrate renewable energy, we need accurate predictions of the extent of solar power that will be generated. By analysing historical production data alongside weather, cloud cover and dust conditions, AI models can improve these forecasts. Better predictions allow grids to use renewables more efficiently and reduce reliance on fossil fuels.
Why are frontier AI labs focused on large general-purpose models?
These labs are businesses, and large-scale consumer products generate revenue. There is also interest in developing general-purpose systems capable of many tasks. But that objective is distinct from building targeted systems to solve specific societal challenges like climate resilience.
What is Climate Change AI, and what has its impact been?
Story continues below this ad
Climate Change AI is a nonprofit founded in 2019 to support work at the intersection of AI and climate. We run educational programs, workshops and grant initiatives to build skills and foster collaboration. Our summer schools have had more than 16,000 registrants from over 175 countries.
For example, one funded project focused on climate-smart shrimp aquaculture. Shrimp farming can degrade mangroves, which are important for carbon storage and coastal protection. The project used satellite imagery and AI to identify aquaculture sites where farming practices could be adjusted to protect mangroves without reducing productivity.
Has political change in the US affected funding for climate-AI work?
The funding environment has cooled. Philanthropic and corporate priorities have shifted, and some initiatives have shut down. However, the US is not the entire ecosystem. Funding and energy are emerging in other regions. Additionally, solar and wind are often more cost-effective than fossil fuels, which continues to drive private-sector momentum.
