The relationship between energy and AI is evolving rapidly

[Music] Welcome back to the IAA's podcast Everything Energy. I'm Dan Hewitt. This week we're looking at what artificial intelligence means for the energy sector in 2026. Sometimes to understand the world around us, you have to follow the money. And right now, the money is heading towards AI. The sums are staggering. The capital expenditure of just five tech companies now exceeds global investment in oil and natural gas production. Last year that figures did at $400 million, and this year is expected to grow by 75%. So the money is clearly there for AI, but is the energy sector ready for what comes next? In April 2025, the IAA published a landmark analysis on the relationship between energy and AI. But the field is evolving so rapidly that new data has come to light, and new questions need answering. Here to help us understand the situation right now, are the lead authors of a new follow-up report, Thomas Spencer and Sid Art Singh. Okay, Sid Art Singh and Thomas Spencer. Thanks for joining us today. Now, this is our second podcast on the relationship between energy and AI. So, I think it's just worth reminding ourselves about the initial aims of the report, and also why we're doing another podcast, why you've done another report so soon after the last one. Yeah, great to be with you, Dan. It's a good starting question. You know, at the IAA, we have a report series that we do every year, like the World Energy Outlook. But typically, when we have thematic reports, they're more one-off. We don't revisit the topic so soon afterwards. But in the case of AI, things are moving so quickly that we felt it was an important role for the IAA. You know, we pride ourselves on being ahead of the curve. An important role for us to come back and look again at this issue and see what has changed since the last time that we picked up on this topic in April last year. I mean, just to give you a couple of examples, a year ago, a typical AI use case might have been someone writing a text query into a chatbot. Now people are using AI for coding. AI is running long-geration tasks in the background while people do something else. Work on something else. Come back two hours later and the task is finished. So the field itself is really moving incredibly quickly. And that, of course, has consequences for the energy sector, which we wanted to look at in this report. Okay, now let's start by unpacking some of the big themes, the big issues surrounding energy in AI and think about whether they've changed or not. So, Sid, I want to ask you about investment because one of the ways to measure the expansion of AI is to look at investment trends and there are some extraordinary sums in this report. So talk us through them. Sure, Dan. So the technology sector is witnessing a historical build out of physical infrastructure to cater to this boom of AI. So AI models, AI inference, which means the use of AI by people, the computations behind it takes place in data centers. And so that's what we track. And we find that the largest technology companies on earth. So what are called the hyperscalers and a couple of others are spending over 700 billion US dollars in 2026 on their capital expenditure. This of course includes data centers, but also includes a couple of other things, but data centers are the primary driver of it. It was as high as 400 billion dollars last year. So this number has been increasing rapidly and has been driven by data centers. In fact, we did a little back of the envelope calculation and the total amount that was spent on the Apollo space program. This was a 10 year long investment that was made by NASA that put humans on the moon. That program cost less than the amount that was spent last year on data centers. So that's the volume of investments that we're talking about here. All right. Now let's think about the efficiency of AI and how it's changed as opposed to this time last year. So how much energy does each AI query use today and has it progressed in the way that you thought it would? There's no easy answer to this to this question and this is because AI is not like apples or oranges. It's not a single good, right? There's different services that we call AI. I think the simplest answer to the question though is to say what does a simple text query? That's what people are usually doing today when they think about AI, they're asking Chatship a question. What is this cost in terms of energy consumption? And the answer is surprisingly little. We estimate that it costs about 0.3 watt hours to ask a model like Chatship a simple text generation query for people who don't immediately have an idea of what that means. It's about the same amount of energy as running a TV for the same amount of time that it takes Chatship to return your answer. That number has been improving extremely rapidly. So we say in the report that efficiency improvements in AI are moving faster than any other energy technology in history. And this is essentially happening in two ways. First of all, the hardware, so the chips that AI models are trained and run on are getting more and more efficient. We estimate by 50 to 60 percent per year, that's incredible, but also the models themselves are getting more efficient. Maybe by a factor of two to ten times per year. There's a wide range here because we don't have a precise data, but certainly the leaps are impressive. However, I said at the beginning, there's no single thing that is called the AI. It's a variety of services. And some AI services are much, much more energy intensive. For example, video generation can be an order of magnitude more energy intensive than text generation. And some of these long duration, what are called, "agentic AI services." So a model running in the background doing something for you, processing a data set of scraping the web, they can be maybe a hundred to a thousand times more energy intensive and simple text generation. So what we have seen is incredible efficiency improvements, but at the same time, the development of more and more energy intensive AI modalities that are increasingly being used by people in there every day. And that's why we project that energy demand for AI is going to continue to increase. Okay, so where does that leave the energy footprint of AI right now? Is it where we thought it was going to be this time last year? Last year, about 500 teravatt hours of electricity was consumed by all data centers globally. That's about 1.5 percent of all electricity consumption globally. Now that might not sound like a lot, but in fact, data centers tend to cluster. That is, they tend to emerge around each other and they often tend to emerge around cities, which are already big consumers of electricity. And so in those particular regions, it can be as high as one third of the total electricity consumption of that particular city or that particular state, which is quite significant. But what's also interesting is what the subset of data centers that are actually dedicated to AI, what's happening in that space. So for example, off the new AI dedicated server racks. So think of a rack about the size of a refrigerator and it contains the computational equipment that helps train models and also use these models. This single server stack is on track to consume as much electricity as about 65 households as early as next year. So these are really small pieces of infrastructure that are consuming a lot of electricity very quickly. They also generate heat as heavy as an SUV truck. So it's the subset of data centers that deal with AI that is currently driving electricity consumption growth. And even into the future, that is where the big chunk of the growth comes from. Just remember, if that energy footprint is progressing in the way that we thought it was when we put out the report last year. Indeed it is. I think we are pretty proud of the fact that our updated numbers remain on track with what we had initially estimated in the first report. So we find that in 2025, electricity consumption from data centers as a whole grew at 17%, which is much faster than the 3% of the total global overall electricity consumption growth. So it's much faster than what everything else is growing. But also within it, it's the AI specific data centers that's driving that growth. In fact, these AI data centers saw their electricity consumption growth growing by 50% in this one year period. Okay, now let's think about the the bottlenecks surrounding the expansion and rollouts of AI. Are the bottlenecks roughly the ones that we thought we knew about or have there been some new ones that we've spotted? really to do with the scale of it and the amount of money that's been put into it right now. Bottlenecks that we saw coming last year have, if anything, gotten worse and new ones have a reason. So, last year we identified, for example, transformers and gas turbines as bottlenecks within the energy sector. Data centers require transformers to be connected to the grid. And many of them are powered by natural gas, which requires gas turbines if new power plants are to be built. So, last year, for example, we saw global orders for natural gas turbines increased by an astonishing 70%. That puts total orders in 2025 at the highest ever level since the year 2000. If you order a gas turbine now, you have to wait five to six years to receive the equipment. In terms of grid connection timelines and the time that it takes to build new transmission lines, for example, we're still talking five to six years in many of the places that data centers are being built. I think what is unique about the IA is that we look obviously at the energy sector, but we also look at the technology sector because trends in the technology sector will determine also how big the impact of data centers are on the energy sector. And here we've seen new bottlenecks emerge, which are going to limit in the near term how many data centers can be built. One of these, for example, is manufacturing supply shortage for high bandwidth memory, which is needed in these really cutting edge data centers that are used to train and run AI models. And where memory supplies are sold out for the next two to three years, where inventories have collapsed down to two or three weeks, and where prices have grown fourfold. And so there's just not enough manufacturing capacity in the IT sector to support more servers going into advanced AI data centers. Another last year or so, there have been some profound disruptions to supply chains. What can we say about the impact of that on the rollout of AI, or is it too early to draw the conclusions? Well, we looked at the potential implications of what is currently happening in the in the Middle East, which is disrupting energy supply chains. And is also having knock on effects to some of the key components that go into the manufacturing of the IT equipment for data centers. So one of these components, for example, is helium. So it's not as such an energy commodity, but it is a byproduct of natural gas production. And it's critical in the manufacturing of the advanced chips that are going into data centers. And before the conflict in the Middle East, about one third of global supplies of helium transited the Straits of Humors. So this is something that we need a bit more time to see how it will play out. Okay, now how has the energy sector itself responded to some of these challenges, some of these bottlenecks? What are some of the innovations we've seen from the sector? It's driving a whole range of innovations in terms of how data centers are connected to the grid and powered. So first of all, most data centers are connected to the grid, and most under development want to connect to the grid, because it's extremely reliable and provides relatively low cost electricity relative to the alternatives, which we can talk about. We estimate in the report that in the next few years renewables power purchase agreements will cover about half of data center electricity consumption. And in the United States, it's even more pronounced. The tech sector is responsible for 65% of corporate procurement of renewables in the United States. So really helping to drive forward that sector in the United States. But the innovations extend also beyond renewables. So the tech sector is driving the nuclear sector. So we saw since the last report, power purchase agreements for roughly seven gigawatts of existing nuclear capacity be signed between utilities. And the tech sector. That's helping to extend the lifetime of these facilities and to fund refurbishments to make sure they can run longer with necessary safety requirements. We're also seeing the data center sector push forward the small modular reactors, which are a new kind of a nuclear power plant design. And the pipeline is even bigger. Last year we tracked about 25 gigawatts of what we call conditional offtake agreements. So these are agreements that are conditional on the technology coming online because there's still an emerging technology. You know, latest report that 25 gigawatts has grown to 45 gigawatts of conditional offtake agreements. Where essentially the tech sector is saying, if you build it, we will take the power. And that's providing a lot of pull for innovation in this in this technology. Okay. Now AI offers huge opportunities for the energy sector. Now, Sid, do you think that right now the energy sector is making the most of that potential? So the energy sector, of course, is extremely capital intensive, which means that there are, you know, cost pressures all around. It is a sector where there's a lot of data that's generated. It's a sector that's extremely complex. And you apply a technology to it that is able to sift through that kind of data generation that's able to sift through that kind of complexity. And offer solutions that can help make outcomes a little bit more efficient, a little bit more resilient. Those are the kinds of outcomes that we are already seeing new startups and existing energy majors apply to their processes. So to give you a couple of examples, we have AI enabled weather forecasts that have improved the accuracy of forecasts. And therefore, it helps reduce curtailment of renewable electricity into the grid. Another example is the application of AI in helping make batteries run better. So they're able to provide more energy output over a longer period of time. And we find through our analysis that at the global level, AI has the potential to reduce energy consumption by as much as the total energy demand by the country of Indonesia by 2035. But of course, we're not really meeting that type of potential for a variety of reasons. On the one hand, we find that although there's a lot of data that's generated through energy processes, through production, consumption and supply, that data is not publicly available. It's not open source. And therefore, it's hard to train models when that data is not available in the first place. In addition to that, we conducted a survey of energy companies and found that the single biggest challenge that these technology companies and these energy companies face is a lack of AI skills. And of course, there are also regulatory challenges and policy challenges. This includes the lack of interoperability, which means that if your say washing machine is not talking to the grid is not talking to the solar generator in real time, it won't know when to start operations and optimize on the generation of solar electricity as an example. Okay, now there's a consumer angle to this as well. Some people might be wondering if they live near a new data sensor. What could that mean for their energy prices? So what does the report tell us? Do more data sensors necessarily mean higher energy prices? So then we conducted a lot of research on this and I wish I had a spicy answer for you, but I don't. Our analysis revealed to us that it really depends on the context and the conditions that we are tracking. So firstly, we looked at all the different countries and states where data centers have been growing. And we could not find a clear relationship between electricity consumption growth and the rise of electricity prices in those regions. So there was no clear relationship, but having said that it is not to say that data centers don't contribute to rising prices. And to maybe you know, explore why that is we can take the analogy of the airline sector. So let's assume two cities that is serviced by a specific airline and they have, you know, let's say multiple flights in a day. If on this particular route, we find that the current capacity that's being utilized is say 50% that is half the seats in any plane are filled with people. In this condition, adding more passengers actually leads to the reduction in the flight ticket prices. Right. You need more capacity utilization of something that is that expensive to purchase and that expensive to build in terms of the infrastructure for the airports and so on. However, if this same route was had a occupancy rate of close to 100%, then adding more and more demand and the inability to supply more flights to meet that demand can lead to prices going up. In the energy sector, similarly, it's extremely capital intensive. So adding more demand does not necessarily mean that electricity prices will rise as long as the existing capacity, the infrastructure is being used. utilized better, so cost-tubbing spread around, right? But because data centers grow very quickly and the electricity system is unable to respond very quickly at the same rates, it's not able to construct new grids and new electricity generation capacity at the same speeds, that mismatch can lead to pressures on prices. But having said that so far, the relationship has not been clear cut. And in the report, we offer a few directional measures that countries and regulators can take to help mitigate any future pressures on prices. OK, now I think this is a good time to ask you both. When you look back at the report this time last year, what were the things that you either missed or the things that you got wrong? And it's going to sound like I'm marking your homework, but maybe Thomas, you want to start us off. I think this is an important exercise for everyone to do when they're trying to analyze a topic that is so complex and that is moving so quickly. So I would say, broadly speaking, our report last year was very comprehensive, and it got most things right, and it covered most of the issues that we've seen emerge over the last 12 months. I would say there's one that we may be missed, which we look at in depth in the new report. And that's the rise of on-site power generation options for data centers. So this is essentially a powering option where the data center is not connected to the grid, but rather has built its own power plants on-site in order to supply its power needs largely through natural gas. And essentially, since our report in April 2025, we've seen a huge number of project proposals. Still mainly at the proposal stage, emerging the United States. So we wanted to look at this issue and to see whether this was a real viable trend for data centers. What we find is that data centers are being pushed to explore on-site power because grid connections are so slow. On-site power raises complexity, because data centers need very high reliability, and that means that you need to overbuild your generation facilities to build in the redundancy necessary to provide that very high reliability. It also raises costs quite substantially because you need to overbuild generation facilities. And it raises operational complexity as well, because data center loads can actually vary quite a bit, and that can put the equipment under stress. And so you see operators start to integrate things like battery storage plus natural gas turbines to match the reliability and meet this variability of demand. So what we find in the report is that this is an extremely rapid emerging trend really over the last 12 months. There are projects that are moving into the construction phase, so this is a real thing. But at the same time, you know, with the constraints in gas turbine order supply chains with these additional complexities of the on-site business model, it's not as silver-bullet. We expect some data centers to move forward with on-site power. We expect most data centers to still want to connect to the grid, and even data centers that do move forward with on-site power may be to look to connect to the grid at a later stage when it becomes available. And if I could add to that another one, we thought is related to robotics and what we call as broadly physical AI. So physical AI is anything that deals with the intersection of artificial intelligence and the embodiment of that AI. So in other words, anything that can physically interact with the world. And this includes robotics, but also includes things like drones, self-driving cars and so on. So we do have some new analysis on that. And we find that firstly, there has been a consistent and growing uptake of industrial robots, for example. Some of this also has implications on industrial productivity. We find that the heart of industrial competitiveness in the future may rely on how quickly they're able to use and deploy artificial intelligence at various layers, including robotics, to help to reduce costs as well as increase production. I would have went by looking ahead. So let's just think about what that relationship between AI and energy could look like by 2030. Maybe you could both talk us through what you think the big themes could be. So in 2030, we expect that the electricity consumption of data centers will have more than doubled from where it stands today. But it will still make up only about 3% of global electricity consumption. It will be much more in some regions and some specific markets. It might surprise the listeners to hear only a doubling. We hear so much about AI today. Why is it only a doubling? And here I come back to the bottlenecks. So we think that our projection for a doubling of the electricity consumption of data centers is really at the upper end of what today's supply chains can deliver, whether we talk about the IT equipment or whether we talk about the energy equipment and our grids and so on. But this does mean that in 2030, we may be standing in front of even faster growth of the electricity consumption of AI and data centers if things like agentic AI really take off. If we have our AI agents running in the background doing our shopping, doing our day-to-day digital tasks and so on. And that's really a question of how fast AI as a technology moves itself. On the other hand, we may find that AI has become so much more efficient that we're starting to run more and more on our laptops and phones. And really calling on data centers only for the really big models for the very sophisticated tasks and so on. And that means we may start to see a plateau in data center electricity consumption. Or we may see that the tremendous investments that Sid mentioned at the beginning have run a bit ahead of monetization. And so there's a bit of a pullback in data center construction because they're so capital intensive, monetization was not able to keep up. Or we might see that monetization has kept up. And so the outlook is for continued growth of electricity consumption from data centers. So all of this to say, we're faced with a really unique position in the energy sector today that the near term is fairly well constrained. We know roughly where things will be by 2030. The investments have been made, the projects have been announced and many are already under construction. And there's just not the capacity to go much faster. But by 2030 we could be standing in front of a wide range of uncertainty. A question for Sid, see what more the future look like. The energy demand from AI or from data center specifically has pretty much been baked in. This is infrastructure that is being constructed as we speak. This is electricity generation that is also being constructed to supply to this growing electricity demand. On the other hand, what's not baked in is the applications of AI in the energy sector. So what we think of as more productive uses of AI across the economy to make it efficient, to make it resilient, to optimize outcomes to make the energy system more innovative. We find that for example, digital meters are actually a small subset of the total electricity meters that exist today. Similarly, connected appliances are only a very small fraction of the total available appliances. So while energy demand from data centers and AI has been baked in, what's not baked in is this potential to offset that increase. We find an analysis that in fact, the energy reductions from AI can more than offset the total energy demand growth from now until 2035. But we really need to work towards overcoming those barriers and ensuring that we're able to make the best possible use of AI in the energy sector. Okay, Sid, I'm saying in Thomas Spencer, thanks so much for coming at a total twist today. Thank you very much. Thank you Dan.