Overbuild

Free daily briefing on global business news.

Global business news for a smarter world

Nobody’s AI has gone down for lack of electricity. So why are we rushing to build more?

Episode 6 of Power Play, Quartz's six-part investigation into AI and the US power grid.

Missed episode 5? Check it out here.

Listen on: Apple Podcasts | Spotify

Nobody’s AI has gone down for lack of electricity. So why are we rushing to build more?

The chips keep getting more efficient, and the plans being announced may run well ahead of what gets built. If the demand doesn’t show up, households pay twice: first on their power bills, then again in their retirement savings. Jackie Snow makes the case for slowing down.

Power Play is a production by Quartz and The Capitalist Hippie.

Reported by Jackie Snow, Catherine Baab, and Joseph Zeballos-Roig.

Hosted by Marco Pimentel.

Produced by Simon Litt.

Edited by Georgina Kemp Walker.

This transcript was generated from the episode audio and lightly edited for clarity.

Marco Pimentel: Are we screwed?

Jackie Snow: I don't think we are screwed. I think what we're seeing right now is a lot of talk. And some of this talk will actually become action that will become data centers. But there's a lot of one-upmanship right now of making or announcing plans that aren't going to go anywhere. And I think if we build everything that has been announced, yes, I think we'd potentially be very screwed. But a lot of this is just not going to be built. And as long as people and our government officials and these tech companies don't go completely haywire, we are building out infrastructure that we will need, maybe not next year. I think some of the projections of how much we're going to be using AI is very optimistic. But these are data centers that will be able to process and create data that we will be at some point in the future needing to do.

Marco Pimentel: I'm Marco Pimentel, and this is Power Play. Over six episodes, we follow the race to rebuild the American grid for AI and the fight over who ends up paying for it. This is episode six. AI reporter Jackie Snow has been researching this with me since episode one. And after six episodes, we go through all that we have learned. Yes, we need more energy, but at the same time, we're not factoring in the curve of advancements of how much energy we need to produce a thousand or a million tokens.

Jackie Snow: I think it's kind of commonly understood by people not in the major tech companies, the people who don't have to go in front of VCs and say crazy things, that these chips and these models are going to get more energy efficient. So we are kind of building with the model we have today in mind for the future, if you look at all our past technologies, if you look at refrigerators, if you look at cars, these first models weren't very efficient. We didn't know what we were doing. And then as you iterate, as you learn, cars get much further mile per gallon now. Our fridges use a fraction of the energy that they did in the 1950s, 60s, 70s. So we're going to see that with these tools. As much pressure as there is to do that, it's going to speed those things up. So I don't know if I buy worrying about, oh, we have such a huge demand coming down the pike. Because one, that is not proven yet. That is thinking, oh, these tools are going to become widely adopted and even more useful and everyone will use them for their work, which is not promised. And two, that we're not going to see any iterations that get us better efficiencies from this. We could really be in a future where we've overbuilt the data centers because we've improved these technologies so much.

Marco Pimentel: Let's head back to Brendan Pierpont at Energy Innovation.

Brendan Pierpont: Each six months that elapse, there's rapid changes in what AI tools are capable of, and also rapid changes in the efficiency of GPUs and the efficiency of these models and how they use that compute resource. And so I think there's a lot of opportunities to do a ton with a more limited set of what gets built.

Marco Pimentel: The AI companies talk about how much energy they intend to use. They talk about how much energy that they need, the data centers that they want to build. But so far, it doesn't seem like any of the AI companies, at least the hyperscalers, have any articles or any research to suggest that they have currently been blocked by usage of energy, that their lights have been turned off or they can't do what they can do.

Jackie Snow: Marco, do you use any AI products?

Marco Pimentel: Well, it depends on which month, but I use a lot of them, yes.

Jackie Snow: Have you ever had it not work for you because it was down?

Marco Pimentel: Never. No, it's fast.

Jackie Snow: Yeah, same here. For you and I, I have not experienced any, oh, the servers are overloaded. It's down. Right now, at least from a consumer point of view, it doesn't seem like there's interruptions in services. There's no Salesforce or Microsoft saying, oh, these tools went down. This is so disruptive. So it's currently holding, which is another reason that I'm always a little suspicious of this build out.

Caie Kelley: I think the answer comes down a little bit to timelines, which is to say over time, I think you will expect there to be more efficient chips. I think you will expect there to be more efficient transmission lines. What I think you're seeing today, though, is that the amount of demand for power and the amount of demand for new data centers and the amount of demand for new inference and compute in general is outpacing the rate at which you could install or optimize for particularly efficient solutions. And so, yes, I think you will see some modulation of pure energy demand through better solutions, more efficient buildings, all of the above. But I think today those timelines and those rates are slightly divergent.

Marco Pimentel: I keep coming back to opportunities as well from the business side. Speaking to Caie from Lowercarbon Capital, it was really interesting to hear how much investment they're funding in just the area of energy, all across the entire place, from fusion to better grid technology to the parts that power our power lines. There definitely seems like a lot of opportunity across the entirety of the grid.

Jackie Snow: It sounds like a lot of investments in what they call grid enhancement technologies, and those can be things as simple as sensors on lines that can better track how many electrons go through the power lines, which usually, pre-AI, pre-sensor technology, you had to just sort of do a number that was an average of what the lines could hold. And in reality, the lines can kind of do different things on different days. And that means you can get more through these lines, even things like some of the battery systems that we're seeing that can smooth out demand peaks are these things that really can enhance this grid, which the grid is run by a very conservative group of people because they should be conservative and safe because if power goes out, people die. But you can start to improve some of these technologies that are 50, 70, 100 years old with our newer technology. A lot of these things aren't even large language models. This is not a technology that kind of came into being three or four years ago. Some of this stuff are simpler machine learning tools, sensors that we have been working on for 10 plus years that can be added to the grid in a very safe way that could improve it for all sorts of different communities.

Marco Pimentel: And I just can't stop thinking about the fact that the grid is not being utilized to its fullest capacity and that there's so much stranded energy sitting there. It's just a matter of harnessing it, which I think ironically does seem like a bit of an AI solution because AI definitely sits in that area to help bring out some of that stranded energy that we have. Or maybe I'm just simplifying it. Because it just seems like a crazy number where how much it is actually being used. Why don't we just optimize everything there and we don't have a problem.

Jackie Snow: So the grid utilization rate does vary from area to area. But I think the number I saw is it's about 50 percent on average being used. So, yeah, there is a lot of electricity just sitting there and not being used because as an industry, utilities have gotten more conservative. They don't want brownouts. You don't want blackouts. Some way to make sure that doesn't happen is to not get too close to the line where then we start to see those things happen. But one of the bigger problems is this isn't such a technological problem as it is a business model problem. The business model doesn't reward using the grid to this 80 or 90 percent. Right now, so many of our utility companies get paid for building and not optimizing. And that's the incentive problem underneath everything.

Marco Pimentel: Here's Paul Denholm from the National Renewable Energy Laboratory.

Paul Denholm: The evidence pretty clearly points to the fact that optimization of the existing grid, we're really in the early stages of how we improve the efficiency of the overall grid. And I primarily mean the transmission assets. I mean, the generators themselves, they're there and we could increase the utilization of them. So it really is largely around what are the techniques on the transmission side, as well as on the demand side, demand flexibility, whether that comes from the data centers themselves or other sources. I think those are the big areas to explore. There's been for many decades, things like performance-based rate making, where if the utility does a really good job of controlling costs, they get compensated for that. So that could be a win-win situation.

Marco Pimentel: Let's head back to Ian Magruder at the Utilize Coalition.

Ian Magruder: Small but significant policy shifts that set the right incentives can meaningfully unlock a lot of this. And that's why we think that beginning by measuring and then incentivizing grid utilization is the right approach. And doing it in a tech-neutral way, it doesn't mandate any one specific technology, but can help shift the whole system in a meaningful direction that will lower rates for everyone.

Marco Pimentel: Is there a mismatch between how many data centers are being built today that already have the power that they need versus, okay, well, in order to build this data center, you actually need to add new energy to the system?

Jackie Snow: I think everything that's getting built that is connected to the grid right now has a promise of having the energy that they need. What is happening right now is some of the data centers that are getting approval are being told, yes, you can connect to the grid, you can use this power. However, on a couple of these hottest days, coldest days that we have, you have to be flexible, which was never something that they were told before. A lot of them did it just to be good neighbors, but now it is in their agreements, we are going to power you off. We have to keep the hospitals and homes on. And there's not pushback to this. I think these companies understand it. And a lot of them are building on-site batteries. So if that is to happen, they don't even have to shut down. They're off the grid. And I had a talk with Google where they're even looking into, can we be an energy source on those days? If it really was the worst day of the year for Texas, not only do we shut off, but do we actually put some of what we have in our batteries onto the grid? I don't think they've done that yet, but there are these possibilities where they can be good neighbors as opposed to just being subtractive from the grid. There's some data centers that are trying to build off-grid because they don't want to wait for these interconnections. From everything I've heard, it doesn't really pan out long term because then you have to become an energy provider. You're dealing with a lot of energy systems when all maybe you really want to do is just run these data centers, provide AI services to people. And all of a sudden you'd have to become a gas turbine maintenance company too. So there's definitely some that are doing this, but there might be some hard lessons learned about how, okay, yeah, you skip the queue, but guess what? Being an energy provider is really hard and that's not what you got into business to do.

Marco Pimentel: Here's energy economist Gernot Wagner.

Gernot Wagner: It's not as simple, not as quick to connect data centers to the grid as is needed to ensure the ramp up of those data centers. So it's off-grid natural gas, not a good climate story, yes, leading to some lock-in, possibly less so than building a dedicated natural gas power plant, which, yes, will be burning natural gas essentially forever.

Marco Pimentel: One of the things that was a bit scary was where all this CapEx ends up going. The cost of it also ends up going to just regular consumers. It has to be. And they're able to pass on those costs. And so there is a worry that in the future, if things go poorly or overspending goes and the need's not there, is that kind of your regular household consumer is going to be spending more money for this infrastructure that was built. At the same time, if this goes poorly, too, the stock market probably will be impacted, double whamming those same households.

Wagner: Depends on how much political power households have relative to the big corporates, who ends up holding the short end of the stick, who ends up paying, who benefits. And what we see so far is that it turns out that the political power as so often rests with those who have a lot of money, who are able to make arguments in sort of their court of public opinion. So they'll tell the rest of us that, no, no, no, AI is good for you. You ought to be willing to pay more for your household electricity bill because look what you're getting in return. While it's harder to make the case that the 100 million households are the ones who should be held home free.

Marco Pimentel: You do hear that, right, where they charge a dollar for a token, but it costs them $1.20. And then they also have the infrastructure that's depreciating because the chips don't last forever. And it all just goes back to NVIDIA, back to Anthropic, back in this whole circle. And you wonder where all of that kind of goes. But that is part of the story as well is that there's lots of uncertainty of how much of this is market-driven, how much of this is storytelling. And so when you are thinking about, again, big, big, big grid infrastructure projects and longevity, that definitely makes people probably a bit more uneasy on spending all this money up front.

Jackie Snow: Revenue is definitely up, but I think there's some questions about profits and some questions about people paying these $20 a month plans or even $200 a month plans, sometimes costing them more than what they're getting from them. That is sort of the way a lot of these VC products of the last 15 years have kind of worked where you lose all this money to acquire customers and then you have to raise rates eventually. It'll be interesting to see what people are willing to pay the $20 a month. Does everyone have to go to a $200 a month plan and the $200 a month people have to go to $2,000? We'll see if things start to break there. There's definitely some concerns I'm seeing about where the money is going to come from kind of here on out for this buildout. After the pandemic, a lot of these companies were sitting on giant balance sheets, had a lot of cash on hand, and that really funded a lot of the buildout since 2022. And now we're starting to see debt being issued. Companies don't want all of this build out necessarily on their financial records as data center build out. So it's kind of financial engineering to do debts and be able to kind of offload that and spread the risk. But a lot of that debt is going to go into places where all of a sudden a lot of Americans are going to have it in their pensions and represented in some of their 401ks. If these debts are bought out by different companies, all of a sudden the risk is being spread in another way. If this was to be a bubble and we see some pullback, maybe that lessens it because it's not just Microsoft, Amazon, OpenAI and NVIDIA doing the build out. Maybe the debt sort of being shared around will make it less bad, but it is changing right now. And there are people who are keeping an eye on this because of what this could mean for the economy in the future.

Marco Pimentel: These questions that people keep asking is, is AI going to ruin the planet or is AI potentially going to solve or help solve and maybe not completely solve the issue, but the pressure of AI and solving it as a business does put things like solar to the grid that we needed and putting energy as a major focus of investment and fast tracking a lot of those sources. Because at the end of the day, price is going to matter for AI, and if price matters, then the big part of their cost is energy and the cheapest form of energy is clean energy.

Jackie Snow: Like anything in this life, it really comes down to execution. Could this be done in a way that is really detrimental to the environment, to climate change, to the planet? Yes. Are there ways that we can do it efficiently and that maybe can actually optimize some of our systems, our electrical systems, power grids, farming, shipping, all these things that if we have enough AI to run through some of this data that we are collecting, yeah, it could improve things. Maybe it'll end up being a mix where there are some uses that are kind of a waste and use too much electricity while there's other uses that reduce electrical use or find new ways to boost our solar numbers in terms of how much we can get off these solar panels. So it really comes down to us doing this correctly, which I do think usually means slowing down, thinking these things through, not rushing out the gate, which I think for a lot of people in the tech industry, this feels like an urgency that they needed these yesterdays. And if we just slow down, we could do this in a way that it would keep people near data centers happy. It could do less damage to the environment and be a net positive.

Wagner: Long term, the answer is clear. Solar, batteries, renewables. And yes, geothermal is part of that equation, for example, nuclear for that matter, they are low-carbon technologies, and especially solar PV, especially batteries, can only get cheaper and better over time. Their technologies, well, gas, oil, coal, commodities, they will fluctuate. They have always fluctuated for decades, centuries by now. They'll continue to do so. And I'm not so sure data centers change that picture all that much. Overall increase? Sure, in prices. But the fundamental trends to electrification, to decarbonization because of those climate technologies being technologies, that trend hasn't changed. Data centers aren't necessarily speeding it up, but they're also not putting a huge caveat here that says, oh, no, no, we need to go back to natural gas. That's the only solution.

Kelley: Texas has more solar today than it has coal. Wind is the largest source of energy in South Dakota. And the reason that this is true is truly, truly not because Texas has a state mandate for installing solar because they love the look of a panel. That state has a lot of sun. They have a lot of land. It is a cheap way to generate electricity. It doesn't require any pipelines to get it into the state. It doesn't rely on any world wars in order to ensure stable supply of it. And so we see that trend continue to play out. And that is a micro example, but overall have a lot of conviction that a lot of the net new load being added to the grid is in fact clean and that a lot of the demand there is clean. Right now we are in a supply constrained environment for energy in general. And so I do think that you end up seeing more demand for non-clean solutions too. I think that is naturally a part of the story.

Marco Pimentel: From a China perspective, they seem like they have kind of the opposite problem comparatively to America, where they have lots and lots of solar power and lots and lots of energy. But does that influence how U.S. grid builders are thinking about their investment? And does that make them be a little bit more hesitant as well?

Jackie Snow: So in a newsletter I wrote for Quartz last year, I pointed out that between January and May of 2025, China added enough solar and wind capacity to power a country the size of Indonesia. In April in 2025 alone, China installed more solar power than Australia had built in its entire history. The way that China is building out their grid and building out cleanly is unlike anything we've ever seen before. They're not going to run into power constraints the way we are. That's partially because they are really building and their grid was not really at the max when that was happening. And also because their political system, having the centralized authority, does kind of allow them to build even if maybe there are communities there that also aren't thrilled about a solar farm in their neighborhood or a data center. But their power concern, they don't have power concerns. They just have power, power, power. So there's been two DeepSeek moments that we've experienced. The first one I think was the beginning of 2025 where DeepSeek came out with a model that used quite a bit less energy and need a lot less training than our closed models being developed in the States. And that really shook the stock market because our closed models all of a sudden didn't look as competitive for these cheaper tools. And then we had another one earlier this year with a different Chinese AI company putting out another paper that just showing that their open source models are more efficient, cheaper to train, and there is some question about if they're being fully honest, but realistically, even at 50% of what they're saying, it is a pretty major step compared to what we've been able to do in terms of energy efficiency. So what China is doing is a threat to the American system and to our build out. If China's going to build the models that the world develops, why are we building so many data centers? Why are we building out this grid this way? Right now, it does still feel somewhat being talked about in financial circles as opposed to over the dinner table. And there are also now ETFs being created that strip out the AI companies. Some people who are paying attention are worried about what this could do to their investment strategy, their portfolio, and are taking these steps to reduce their exposure. But again, if you have something like a pension, if you're a teacher here in California, I think this debt is going to be rated very highly. These are companies that print money with their advertising apparatus. So that debt that they issue is going to be very highly rated. Companies or portfolios are always looking for a way to diversify and this is going to be a highly rated debt. So it's going to sneak into a lot of portfolios of people who are maybe concerned but not paying as close attention as some people who are Bloomberg terminal users.

John Reeves: AI has the added complication too, which the powers that be in Congress or among citizens don't understand a lot of the issues involved because it's very complex. So I think most of us don't have the technological knowledge. And so you often sort of just kind of raise your hands and, okay, this is just happening and there's nothing I can do. And I think they want us to feel that way.

Marco Pimentel: That's John Reeves, a historian whose next book is about the taking of the Black Hills.

Daniel Palken: If you were left asking Congress to respond and regulate, you would have serious buyer's remorse because Congress, I worked as a staffer in Congress for a number of members for almost five years. Congress simply would not be able to keep up with trends here. So that is the challenge we face. I wish I had the answers for you, but it's hard enough to just describe the problem.

Marco Pimentel: I started this series because I wanted to find out if we were screwing ourselves over with AI. It's undeniable that it's effective at what it's used for, but no one seemed to agree on what we were trading it for, how we could best prepare for the future. After listening to a TED talk, I went in thinking this was a power problem. It isn't. The grid runs at about half of what it could carry. There are cheaper, faster, smarter innovations being built right now, which means we don't just have to multiply what already exists. What we're dealing with is a decision problem. Every one of the fixes I learned about had hit the same wall. Bureaucracy, slow-moving governmental agencies whose operating rulebooks were written before any of this was dreamt of. John Reeves illustrated how people look at problems this complicated and just bury their heads in the sand. But he also reminded me that there are companies and operators who benefit from us feeling that way, who want the scale of all of this to make it feel inevitable. But it isn't. There are passionate, intelligent, hardworking people trying to figure this problem out. They're coming up with plans, adjusting those plans, and working to realize them. And sure, I may disagree with their methods at times, but isn't that the way things always work? The outcomes of today aren't following laws of nature, they're following the laws of people, of you and I. Decisions are being made constantly, even if almost none of them make the news. So, no, I don't think we're screwed. I think we're in the midst of a massive technological leap and changing the world as we know it. But I keep coming back to something Jackie said.

Jackie Snow: If we're not running into these problems, of course we need more, but do we need it this fast?

Marco Pimentel: That's Power Play. The series is a production by Quartz and The Capitalist Hippie. It was reported by Jackie Snow, Catherine Baab, and Joseph Zeballos-Roig. Our producer is Simon Litt, and our editor is Georgina Kemp Walker. Additional support from Olivia Blackwell, Dan Hirschhorn, Sarah Douglass, and Armando Roggio. I'm Marco Pimentel. Thanks to everyone who sat down with us to explain the grid. You can find the reporting behind every episode at qz.com slash powerplay.

Join 500,000+ readers who start their day with Quartz.

By subscribing, you agree to our Terms of Service and Privacy Policy.