E10: Data centers and the race to power AI (part 1)

Madison Goldberg: Hey, welcome to Ask MIT Climate. I’m Madison Goldberg, and we’re back in your feed with a special episode on a topic that’s been in the news a lot lately: data centers.

And, before we get started, just a heads-up that this episode is actually the first in a two-parter. In this half, we’re going to dig into why this moment is such a challenging one for the energy system in the U.S. And in the next, we’ll go through some ideas for responding to it.

So…data centers.

These facilities are at the heart of the explosion in artificial intelligence. Companies are pouring money into data centers and their racks of equipment to train and run AI models. And with this massive buildout comes a surging appetite for electricity—much of it (though certainly not all) in the United States.

Costa Samaras: And what the electric grid is now faced with, both the people that work in it and the machines and the equipment that deal with it, is trying to accommodate very large new power demands very quickly, all at the same time.

MG: That’s Costa Samaras.

CS: I'm the director of the Scott Institute for Energy Innovation and a professor of civil and environmental engineering at Carnegie Mellon University.

MG: Professor Samaras takes a big-picture view of the energy system: how we can get to a world where we drive and heat our homes and do all the things we need energy for, without polluting the climate.

And he's been thinking a lot about the intersection between that energy system and the race to build data centers: specifically, why our choices matter for power grids and people right now—and for the long-term work of tackling climate change.

Because as he points out, if we’re going to stop warming the climate, the electric grid will need to handle a lot of new demand—no matter what happens with data centers.

CS: Moving away from burning fossil fuels to using electricity that's made from clean sources to get around and have warm and cool homes, that is at the center of how we deal with climate change.

MG: Plugging in all those electric vehicles and heat pumps will be a big lift for the grid, especially considering that the U.S. is kind of out of practice: For much of the 21st century, our need for electricity hardly grew at all.

CS: And data centers represent this early initial test of our society, our policy systems, our infrastructure systems, of: How do we add lots of new electricity demand in ways that are good for communities and good for the country and good for the climate? And right now, we're not doing so well with that test.

MG: Now, it’s worth noting right up front: how much data center demand we’ll be contending with—and how fast it shows up—is a big open question. Partly because the future of AI is uncertain, yes, but also because communities are calling for more of a say over how this buildout unfolds. In other words, we’re not just taking the test; we can also have a hand in writing it.

For data centers that do get built, communities across the country are working to set ground rules—including for how these facilities fit into the broader energy system. And Professor Samaras argues that these decisions will have ripple effects way beyond this moment.

CS: The way that I like to think about it is, what do we want to be looking back at, as society, at the end of the century, of what we did in this moment for the electricity grid? I would rather we get it more right than wrong now than we spend the rest of our careers trying to fix it.

MG: Our second guest today thinks about AI’s electricity demands a lot, both at work and in his own experiments with AI.

Vijay Gadepally: My son, a couple of years ago when he was learning to read, I made this little app on his iPad, and it was a story creator. And so he could go in, he could say, I want a story on, say, little blue trucks. And the system would use AI to generate not just a story, but also some imagery to go along with that story.

MG: Vijay Gadepally is a senior scientist at MIT Lincoln Laboratory and the co-founder of two AI infrastructure companies.

VG: So my research is at the intersection of AI, computing, and energy. And what me and my group are really interested in is understanding how AI is using computing and how to make this more efficient.

MG: Naturally, Dr. Gadepally was curious how much energy the app was using when his son asked it for a story.

VG: So it turns out that every time he was generating one of these stories and the images, he was actually using as much energy as that entire iPad, but it was occurring somewhere else, in a data center somewhere outside of my house. If every time he pressed “generate story,” the iPad's battery died, I'd probably have that immediate signal, like, oh boy, this is using a lot of energy.

MG: Now, the idea here isn’t to say, ‘This is how much energy AI uses.’ For one thing, Dr. Gadepally made this app a while ago, and AI’s energy efficiency has been improving fast. Plus, when you ask AI to do something, the energy demand depends a ton on the model and on the specific task—like answering a quick question versus generating a video.

In fact, we’re not going to focus too much on the energy footprint of individual AI use. But Dr. Gadepally’s story is a reminder that AI and energy are intertwined, even if we don’t always see it. It’s not like running an air conditioner or a microwave, where the electricity is used right in front of us—and shows up on our bills. Instead, the energy use is typically concentrated in a data center.

Of course, if you live near one of these data centers, all that electricity use can feel very up-close and personal.

VG: When you walk through a data center, you typically have rows of what we call racks. And within these racks are the computers. These are really beefy, beefy systems. And one of these data centers can have tens of thousands of these housed in them. Outside of that, there's a lot of cooling, a lot of electrical. So think of big switches, thick cables going from rack to rack. Lots and lots of network equipment, lots of flashing lights, if you're into seeing that.

MG: Data centers aren’t just for training and running AI models. They also hum along behind our movie streaming, social media scrolling, and plenty of other things we do in the digital world. But as tech companies rush to capitalize on the current boom, they’re building more infrastructure that caters to AI—packing specialized hardware into data centers that are ravenous for electricity.

VG: These are large, large projects, orders of hundreds of megawatts to even gigawatts of new builds that are coming online. Stargate's a good example, but in Tennessee with [SpaceXAI], in Louisiana, Meta has a huge project coming in.

MG: And just to put this into perspective…

CS: When we think about large hydroelectric facilities like the Hoover Dam, that has a capacity of two gigawatts, so two thousand megawatts of power.

MG: Meta says their data center in Louisiana will eventually host five gigawatts’ worth of computing capacity.

CS: To me, it was like, whoa. These are generational power investments that are being made at the same time, in many parts of the country.

MG: In 2024, data centers accounted for between four and five percent of all the electricity used in the United States. It’s super uncertain what the future will bring, but that number could grow a lot. Judging by recent forecasts, it’s not unreasonable to imagine data centers making up more than double that share—so, like, 10 or 11 or 12 percent—just a few years from now.

And to add another wrinkle, data centers often “cluster.” Like, in Virginia—which is home to the biggest data center market in the world—they already consume a quarter of all electricity.

VG: The other aspect of data centers, which is what's making it very challenging, is the peaky nature of AI workloads.

MG: Like, training an AI model can involve big swings in power demand, in a matter of minutes or seconds.

VG: Most grids don't have anything in place to be able to deal with movements of scale that quickly. I mean, I'm adding, you know, townships’ worth of power demand at once. And I don't fully understand what that's going to look like or when it's going to be utilized.

MG: Okay, all these factors—the scale, the speed, the density, the ups and downs—might feel like a lot to keep track of. And that’s kind of the moral of the story here. If you’re trying to operate a power grid, the data center boom might feel like you’re being handed a whole lot to juggle all at once.

And if something goes wrong, it’s not just a problem for the data center. Running an electric grid is a delicate balancing act between supply and demand: if they don’t match, you can get disruptions or blackouts. So, for instance, if a bunch of clustered data centers all ramp their power down at the same time, you could have a problem.

A lot of people are also wondering: What does the data center boom mean for our electric bills?

CS: It depends. In areas of the country where lots of new electricity demand from stuff like data centers gets added to the grid quickly, that may trigger new investments that are needed for power lines or transformers, or may change the fuel mix in that region.

MG: So, for instance, prices could go up if a utility makes expensive upgrades to the grid to accommodate its big new customers. Without safeguards in place, those extra costs can get passed on to everyone in the system.

CS: In other parts of the country where the existing infrastructure is capable, and data centers or other new electric loads are added, now all of a sudden you have the same amount of infrastructure spread out over more and more users, and costs may not go up, and they even could go down in those places.

MG: So, there’s a lot of complexity here, and we’ve barely scratched the surface. The impacts really depend on where and how far in the future you’re looking, and what rules are in place to protect other customers. But it’s something we need to be thinking about now. People are already dealing with rising electricity prices—and while a lot of that has come from pressures that aren’t related to data centers, they have played a role in some cases. Projected data center demand has recently helped push prices up in states like Maryland and New Jersey.

CS: The challenge for state regulators and elected officials right now is to ensure that we have an infrastructure system that can handle new electrification loads in ways that don't trigger existing ratepayers to pay for it.

MG: And we haven’t even talked yet about where all this electricity is going to come from. Some companies have ambitious plans to bring a lot of clean energy online, like from nuclear reactors and geothermal plants. But there’s also this:

CS: Because of the time that it takes to get connected to the grid if you're a large data center in many parts of the country, some data centers have opted to just make their own power onsite. Now, that's great if these are solar farms with lots of energy storage, but in many cases, it's natural gas—less efficient power plants that are built relatively quickly, because it's so advantageous for companies to get online as fast as possible.

MG: For those who are thinking about the climate impacts of the AI boom, this is a key concern: not only that companies will keep clamoring for more electricity, but that, in their scramble to get online, they’ll increasingly rely on polluting power sources—with implications for the climate and for people’s health.

CS: When a diesel generator is providing emergency power for a data center, or a natural gas-fired generator is running to power a data center, the impacts show up immediately, in the form of new air pollutants that are being added to that community, and those air pollutants can cause a wide range of health problems such as asthma, cardiac and lung problems. They're especially harmful to vulnerable folks, very young children, senior citizens, people with health challenges. And often these are added to an existing background air pollution in some communities that's already pretty high.

And data centers, like any industrial use, will have noise associated with it. And noise is a pollutant. Other communities are concerned about the water use of data centers. And this is an issue where the local context really matters. There may be parts of a county that is not under water stress and other parts are under water stress.

And so right now it's the wild west, with new data centers popping up everywhere. And it represents this reflection moment: What do we want the future to look like? How do we guide a future that has lots of new electric uses, that also enables electric affordability, electric reliability, getting to zero emissions, harmony with the community?

MG: If this all feels kind of daunting—first of all, same. But there are lots of people thinking about how we can respond to this moment much better than we’re doing right now.

And from the city level to the national level, policymakers are weighing how to approach the data center boom and the hunger for electricity that comes with it. So in our next episode, we're going to get into a few of the ideas on the table. Maybe some of them are being considered where you live.

Ask MIT Climate is the climate change podcast of the Massachusetts Institute of Technology. Aaron Krol is our executive producer. David Lishansky is our sound editor and producer. Michelle Harris fact-checks our episodes, and the music is by Blue Dot Sessions. And I’m your host and associate producer, Madison Goldberg. I also wrote this episode.

Thank you to Dr. Vijay Gadepally and Professor Costa Samaras for joining us. You can find more episodes of the show at climate.mit.edu. In fact, the second half of this two-parter is already up there and in your feed. You can also keep up with us on TikTok, Instagram, and Youtube @askmitclimate. Send us your climate questions at askmitclimate@mit.edu—and, as always, thank you for listening.