We mapped every data centre on earth

I believe that this nation should commit itself to achieving the goal before this decade is out of landing a man on the moon and returning him safely to the earth. >> Ignition sequence start. Lift off. We have a lift off. >> NASA's Apollo space program, which landed [music] the first humans on the moon and ran for over 10 years, is estimated to have cost around $300 billion >> [music] >> from its inception to the very last mission home. >> That's one small step for man, one giant leap for mankind. >> Large technology companies spent more than that last year just on data centers. In 2025, some of the largest technology companies invested [music] over $400 billion in physical infrastructure. Most of that went to data centers, [music] and they're on track to spend 75% more this year, $700 billion. >> AI is often talked about as a software story, but the technology requires an enormous buildout of physical infrastructure. Not only the data centers [music] that train and run AI models, but also the power plants and grids needed to make them work. >> And that level of capital expenditure has major implications >> [music] >> for how the world's power system generate and distribute electricity, where it gets [music] consumed, and of course, who pays for it. >> To understand whether all of that investment is actually turning into infrastructure, we went looking [music] for it. We used satellite-based tracking and geospatial analysis to monitor the buildout of what [music] many call AI factories. These are large data centers purpose-built for AI, and we found that installed capacity has doubled in the last 12 months. >> Inside these [music] vast, windowless buildings are hundreds and even thousands of rows of server racks, >> [music] >> each about the size of a large fridge. In 2020, before the release of ChatGPT, one of these racks used about [music] 13 kilowatts of electricity. That's roughly the same as running five electric ovens simultaneously. Not insignificant, but manageable. The latest chip sets being released are much more advanced and [music] consume much more power. Over the next year or so, a single rack could consume as much as 600 kilowatts of electricity. That means the same fridge-sized box as a few years ago with different technology in it >> [music] >> would consume as much power as 65 households. And that's multiplied across hundreds or even thousands of these racks in a large data center. A typical hyperscale data center now consumes roughly the same amount of energy as 100,000 homes. The largest data center currently under development will use about as much power as 2 million [music] households. Globally, data centers are on track to consume nearly 1,000 terawatt hours of electricity by 2030 or roughly 3% [music] of the world's total demand. That would put their energy use on par with one of the world's largest economies consuming as much power as the entire country of Japan. >> Data centers are growing fast, but it's worth putting that into perspective. Globally, advanced manufacturing, electric vehicles, and air conditioners are all [music] even larger drivers of electricity demand growth to 2030. >> Nonetheless, in some regions, data centers [music] are set to be the main driver. That is certainly true in the United States, the world's largest data center market. In the US, almost half of the [music] increase in demand for electricity to 2030 will come from data centers. In fact, [music] by the end of the decade, the US economy is set to use more electricity to power data centers than all its electricity used for the production of aluminum, steel, cement, [music] and chemical products. This strong growth has led to questions on the impact of data centers on [music] other parts of the energy system. For example, do data centers raise electricity prices for households? >> The inflationary shocks brought on by the COVID pandemic, the 2022 global energy crisis, >> [music] >> and the more recent conflict in the Middle East have made the cost of electricity top of mind for many people. The rising number of data centers [music] being built, there are concerns that their energy demands could compete with the energy needs of other nearby [music] consumers and potentially lead to higher local electricity prices. If a power system around a new data center [music] is already close to capacity, this could require grid operators to make investments in new [music] and upgraded infrastructure. There are concerns that these costs to upgrade networks could be passed on to other businesses and households in the area. However, that does not always need to be the case. [music] The impact of electricity demand growth on electricity prices depends [music] on a combination of fundamental factors and policy choices. Think of a flight route between two [music] cities serviced by an airline. If the number of passengers was only half of the plane's total capacity, then [music] adding additional passengers could help reduce the airfare for everyone on board, since the fixed costs would be shared more [music] widely. On the other hand, if demand jumps above the number of seats, pushing [music] against the limits of capacity, prices can climb. In our analysis, [music] we found that the relationship between electricity demand growth and electricity prices is not always direct. That's because [music] demand growth alone does not determine prices. It also depends on existing system conditions, how markets are designed, [music] and policy choices around how the cost of new grid infrastructure and generation capacity are recovered. >> [music] >> That makes the relationship far more nuanced than it first seems. >> Having said that, it is important to note that data centers can create [music] special challenges electricity affordability. Data centers are built very quickly and in many cases they need new electricity generation capacity and infrastructure. The key is getting ahead of it. With proactive planning, better transparency from technology companies, and smarter management of data center electricity consumption, there's no reason why reliable electricity and affordability cannot go hand in hand. >> Yet another energy sector challenge has been the trend of clustering of data centers. The presence of a large number of data centers in one location creates special challenges [music] for the local grid step. >> The IEA conducted [music] a geospatial analysis of every data center in the world. What it revealed is that they don't behave like most energy intensive infrastructure. For example, steel smelters and aluminum plants are often built away from population centers, usually near large power sources. Data centers are different. >> [music] >> They cluster near cities, near the fiber networks, and the customers they depend on. That can put enormous [music] new electricity demand onto grids that are serving large urban populations. At the local level, data centers in these clusters [music] can easily use up to 30% of the entire local electricity supply. Our analysis also shows that this clustering [music] is becoming denser. New data centers are typically built near existing ones, compounding the pressure on local [music] grids over time, rather than spreading it. And when data centers are built in new areas, our analysis of the pipeline of project applications around them reveals that [music] new clusters are forming around them, too. That means the challenge is no longer confined to areas with a history of dealing with data centers on their grid. It's moving into new regions, onto new grids, and those new locations are essentially starting from scratch. >> Meeting that demand requires the data centers to be connected to the electricity grid. But grid connections can be slow. And especially in the United States, developers are taking matters into their own hands. Many are looking for innovative solutions to power data centers, from generating electricity on the site of the data centers itself using natural gas turbines to large-scale batteries. Some are even considering more emerging solutions such as small modular [music] nuclear reactors. >> But providing reliable onsite power generation isn't always simple. Developers of onsite natural gas generation [music] have to oversize their turbine capacity compared to what they actually need to ensure that the power supply is reliable, which is critical for data centers. [music] In some cases, capacity is set to be up to 70% higher than what they would expect to use on a typical day. And getting your hands on a new gas turbine is not always easy. And many other crucial pieces of power infrastructure also undergoing major supply [music] crunches, too. >> Data centers where AI training and usage take place also experience large swings in power demand at the server level. In one example from 2025, energy use jumped by 400% in just a quarter of a second [music] during a training run. A gas turbine simply can't accelerate fast enough to service that rapid of a demand surge. So to handle these sudden spikes, developers are adding energy storage at data centers that can be tapped when needed. [music] In fact, the largest ever battery project in the world, which is four times larger than the previous record holder, was recently contracted by a data center developer. Nuclear power and next generation geothermal are also getting a boost from the AI boom. Some traditional nuclear plants are being brought back online and there's growing momentum behind small modular reactors, although the first of these won't be operational until at least 2030. There is a potential upside to all of this. Data centers with their own generation and storage don't just have to draw power from the grid. They could support it as well. Unused electricity from on-site gas turbines could be supplied to nearby homes and businesses. At the same time, large on-site batteries could store excess renewable energy from the grid and release it when it's needed [music] most. >> So, one way to gauge what this all means for the energy sector is to look at what financial markets are saying. [music] Broadly speaking, recent movements in company valuations don't suggest that AI has created a [music] broad-based uplift for the energy sector. It's important to point out that of the trillions of dollars expected to be invested in data [music] centers between now and 2030, only around 20% is actually directed towards energy-related assets. The remaining 80% that goes into computing or other data center infrastructure. That said, a slightly different picture emerges when you look more closely at the specific parts of the energy sector. [music] Manufacturers of gas turbines and electrical equipment, certain nuclear companies, and [music] a new generation of energy startups have seen their valuations become more intertwined with market sentiment around AI-driven demand. That creates clear opportunities, but also potentially risks. Companies that invest heavily to meet data center demand could be left exposed if the pace of the buildout slows, leaving them on the hook for the cost of new infrastructure or manufacturing capacity for demand that ultimately doesn't fully materialize. In response, [music] many large companies have opted for a more measured approach to new investment decisions, and at the same time governments can also provide support for the deployment of energy technologies such as small modular reactors so that startups in the space don't have to rely solely on data center demand to scale up. >> Generating more power is the obvious response to rising demand. But what's happening on the other side of the equation as AI itself becomes more efficient? >> When you look at how much energy takes for AI to complete a single task, like answering a question, we see that it's getting much more efficient. In fact, the energy efficiency of AI by this measure is improving at an extraordinary pace. It now takes much less electricity to generate an AI response than it did a year or two ago. A simple text query today typically consumes about as much electricity as running a television for a few seconds. If every traditional web search today was replaced with an AI query, that would add roughly 4 terawatt-hours to annual electricity demand. That's less than 1% of total power consumption from data centers today. AI is increasingly being used for much more energy-intensive workloads, including high-compute single tasks like video generation and longer-running multi-step workflows called [music] agentic AI, such as coding or research. Those kinds of agentic tasks can consume hundreds or thousands of times more energy than a simple text query. And as AI agents become more capable and more widely deployed, these could outweigh the efficiency gains we're seeing at the task level. >> But meeting energy demand for AI is only one half of the story. >> [music] >> AI also has huge potential to help the energy sector solve some of its toughest problems. This is especially valuable today when the pressures on energy security and sustainability are triggering the need for new and innovative solutions. >> We've looked far and wide to understand where AI could help the energy sector tackle the challenges at hand. Already we see that AI solutions are being deployed globally across a wide range of energy technologies. They help manage complex systems, [music] improve forecasting, strengthen resilience and security, and drive greater efficiency. [music] We estimate that AI applications already available today could unlock energy savings of [music] 13.5 exajoules by 2035 if deployed at scale. To put that into perspective, [music] that's roughly equivalent to the amount of energy Indonesia consumes today. >> To deliver on the potential benefits of AI, a few things need to be done in the energy sector. AI needs data to train on, a lot of it. However, data across the energy system tends to be fragmented and is often very hard to access. And in many markets, the incentives make it more financially attractive for energy companies to build new infrastructure [music] rather than adopt technology that could make existing infrastructure work better. And perhaps most importantly, energy firms themselves see the [music] lack of AI skills as the single largest barrier to the uptake of AI. >> This is where an important distinction [music] comes into play. The large, general-purpose AI systems that power [music] chatbots, generate images, and write code are generally the kind to be highly energy intensive. [music] They're the primary driver of the increased energy consumption we've been discussing. But much of the AI being deployed within the energy system looks [music] very different. These are smaller, more purpose-built models trained on specific [music] data for specific tasks. Their energy footprint is typically orders of magnitude lower than that of general-purpose models. In other words, the AI that is driving the surge in energy demand and the AI that is helping to manage and improve the energy system are not always the same thing. The label may be the same, but the technology, the scale, and the energy footprint are not.