AI Data Centres Won’t Break the Grid (If We Force Them to Behave)

[00:00:00] Speaker 1: Electricity grid operators are used to planning for when power stations suddenly go offline. What they're not used to is dealing with customers doing that. But in July 2024, in Northern Virginia, that's exactly what happened. Within seconds, about 1.5 gigawatts of demand vanished, more than a large nuclear reactor. Dozens of data centers had sensor wobble on the grid and switched themselves over to back up power, all at once. The grid was able to handle that without blackouts. But now imagine multiplying this event by the increase in data centers that we're expecting due to the AI boom. Are data centers going to break the grid? They could. But they could also be one of the best things that's ever happened to it. For more than a century, we've balanced the grid by adjusting generation to match demand. Data centers could be the first big customers that flex for the grid instead of the other way around. The catch is that so far, they've shown pretty clearly that they won't do that on their own. I'm Rosie Barnes, an engineer with 20 years experience developing clean energy technologies and founder of Pilot Consulting, where we advise clean tech startups and investors. Welcome to Engineering with Rosie. So when it comes to data centers, how much energy are we talking about? Depending on who you ask, AI is either about to consume all the electricity in the world, or the issue has been wildly overblown. As usual, the truth probably lies somewhere in the middle. Let's start with a base case from the International Energy Agency. The IEA estimates that data centers consumed around 460 terawatt hours of electricity in 2024. That's roughly the same as France. And in the IEA's base case scenario, they predict that data center energy consumption will grow to more than 1,000 terawatt hours by 2030 and 1,300 by 2035. Even with this eye-widening prediction, data centers only account for roughly one-tenth of global electricity demand growth over the next decade. Other sectors, including industrial loaders, air conditioning, and electric vehicles, are also placing enormous new demands on the grid. AI is a major contributor, but it isn't the main reason that electricity demand is rising fast. Some of these data center predictions are based on pipelines that are hard to take seriously. Take Texas, for example. As of June this year, ERCOT, which runs most of the Texas grid, had about 475 gigawatts of large loads asking to connect. And about 90% of that was data centers. Texas' all-time record for peak demand set in July this year is about 91 gigawatts. So the queue is more than five times the size of the entire grid at its busiest moment ever. Add that on top of today's demand, and Texas would need a grid six times bigger than anything it has ever run. And yeah, that obviously would break it in some pretty fundamental ways. But is it remotely realistic? When you hear headlines about proposed AI data centers, it's easy to assume that all announced projects will ultimately be built. But that's not how it's going to work out. One of the biggest risks facing any new data center project isn't financing or construction, it's securing enough electricity and time. Without power, there is no data center. Now imagine you're a developer. You know you absolutely must secure a new data center and get it powered as fast as possible. Are you going to put all of your eggs in one basket and plan a single site? Or are you going to investigate multiple locations and submit multiple grid connection applications? I think it's obvious that the majority of projects in the data center pipeline are not going to be built. To consider by how much, we can look to how energy project pipelines have panned out in recent history. Lawrence Berkeley National Laboratory's queued up report shows that for solar, wind, batteries and gas plant projects, only 19% of projects seeking interconnection between 2000 and 2018 ultimately reached operation. On a capacity basis, only 14% of the requested megawatts became operational. We don't currently have similar global studies on the number of data centers that become operational. However, if you compare the frost of the data center hype today to the somewhat less effervescent energy hype over the past 20 years, I think we can extrapolate that this trend is likely to be even greater this time around. And ballpark guess, I'm going to say we'll see like 10% of these data centers we expect today to become reality. This doesn't mean data center developers are being careless though. For a data center, speed is way more important than getting the pipeline a little bit long, paying for projects that they don't ultimately develop. Hold on to that thought about speed because it matters a lot later. So what's the implication if only a fraction of the projects in the pipeline get built? The answer is we'll still be facing the largest increase in electricity demand in decades. And one thing that the global increases hide is that location really matters here. Even a manageable global increase of a few percentage points can still create severe pressure locally where several hundred megawatts or gigawatt scale proposals cluster behind the same substations and transmission pathways. So what does that mean for everybody else on the grid near those clusters? So far, we've been talking about AI data centers and electricity demand, but unless you're planning to build your own data center, why should you care? The first reason is simple: somebody has to pay for all the new infrastructure. Supplying energy to a large AI data center isn't as easy as running a really long extension cord to the nearest substation. Utilities often need to build new substations, reinforce transmission lines, upgrade local distribution networks, and sometimes construct brand new power stations. Those infrastructure investments cost billions of dollars. So who pays? In some electricity markets, large customers are expected to cover most of the cost of connecting their facilities. In others, a lot of the costs are shared across the wider customer base through electricity tariffs. In those cases, households and businesses may end up paying for at least part of the infrastructure costs for a new data center. Alternatively, a large new electricity customer could, in theory, improve the economics of an existing grid. Most of the grid's cost is fixed. The poles, wires, and substations cost about the same whether they're busy or idle. So more demand on the same network should mean that each unit of electricity carries less of that cost. It's like mobile phones. Calls were really expensive when hardly anyone had one, and they got cheaper as the networks filled up. And this is where we come back to data centers one real priority: speed. A data center stuck for years in a connection queue is a very expensive missed opportunity. A faster connection is something they really would change their behavior for. So some utilities are starting to ask: what if data centers that agree to be flexible, turning their demand down when the grid is under pressure, got to connect sooner? Now, not all AI is a giant, power-hungry training run. Some of it is small and useful, like today's sponsor, Plod. Plod. This is the Plod No Pro, an AI-powered recording device that captures meetings and conversations with just one press. It has four microphones, can capture audio from up to five meters away, and offers up to 30 hours of continuous recording. It can automatically switch between recording phone calls and in-person meetings, and you can press a button to highlight important movements as they happen. There's also Plod Desktop, which captures online meetings, which is handy for people like me who work a lot remotely. But the interesting part is what happens after the meeting. The Plod app uses AI to transcribe conversations in 112 languages, identify different speakers, and generate structured summaries. You can even use Ask Plod to find specific information from a recording, like a technical specification mentioned during an interview, rather than searching through the entire thing. Now, a lot of work goes into researching Engineering with Rosie videos, and being able to go back and check one technical detail without re-listening to a whole interview is the bit that I find most useful. You can find out more about the Plod Note Pro using the link in the description below. Not all data centers are created equal. Traditional data centers have been some of the electricity grid's favorite customers for decades. The digital services most of us rely on every day – cloud storage, email, websites, YouTube, etc. – typically operate around the clock, with electricity demand that changes gradually and predictably. From a utilities perspective, they're almost ideal. They're large, steady consumers of electricity that are easy to forecast and plan around. But AI data centers are different. The issue is not simply that they're larger. It's that some of the workloads they run can change electricity demand much more quickly than traditional computing. This is not at all like the challenges associated with variable renewables like wind and solar, if that's what you were thinking. Renewables change their supply of electricity, and their variation depends on the weather, which can be forecast days ahead of time. AI data centers, on the other hand, can change their demand for electricity, well, on demand. When electricity demand changes faster than the grid can respond, operators have a big problem. And these are no longer just theoretical concerns. Utilities are already encountering situations where large data center loads require new planning rules and new ways of operating the grid. A situation in Ireland provides one of the clearest examples. Because Ireland's electricity grid is relatively small, and Ireland is also a booming technology hub, an unusually large share of the country's electricity demand comes from the rapid growth of data centers. Iagrid and Sony, Northern Ireland's electricity transmission system operator, modeled credible fault scenarios where data center demand could suddenly fall by more than a gigawatt, at the same time as electricity exports through a major interconnector were lost. The lesson it learned was not to ban data centers, rather that fault ride-through capability, connection standards, and real-time coordination become essential when one type of electricity customer becomes such a large stakeholder of the grid. The United States highlights a different challenge. NERC reported multiple customer-initiated data center load reduction events exceeding 100 megawatts during 2025, including two above 1000 megawatts. And then there's the Virginia event from the start of this video. So what happened there exactly? On July 10, 2024, in Northern Virginia, a transmission line component failed, triggering a series of electrical faults and brief drops in voltage. Data centers in the area detected the disturbances and automatically transferred their computing equipment to backup power systems to protect it. And as electricity demand suddenly fell, generation briefly exceeded demand. That imbalance caused grid frequency to rise until operators reduced generator output and restored balance. Notice what those data centers were doing. They weren't helping the grid, they were protecting themselves. And that's completely reasonable from their point of view. But it's exactly the behavior that makes them hard for the grid to live with. And you can't fix that just by building more power stations. Our new wind farm doesn't solve for a congested substation. More solar panels don't eliminate a 10-year transmission bottleneck. And neither of them helps much when a large electricity user suddenly changes its demand in a matter of seconds. That's why the next chapter of this story isn't just about generating cleaner electricity. It's about making the electricity system itself more flexible. Traditionally, we've treated electricity demand as something the grid simply had to live with. Demand was largely fixed and supply did all the work. Now, the ability to ask large data centers to temporarily increase or reduce the electricity demand in a predictable and measurable way is on the horizon. Flexible AI data centers could become dispatchable loads. Active participants in balancing the grid, responding to conditions in ways that help maintain reliability while protecting the computing that matters the most. That means knowing what work can safely wait. In the same way that not all data centers are created equal, workloads come in a wide variety of shapes, sizes, and importance. When we talk about turning down a data center, it sounds as though a very large hand is reaching for like a volume switch. But that's not what's happening. Some tasks need an immediate response and have little to no tolerance for delays. But other jobs are far more flexible. For example, training the next generation of AI models or running routine background tasks can often be slowed down, paused, or moved to another time with little to no impact on the end user. The challenge is deciding in real time which workloads should wait and which ones shouldn't. Making those decisions fast enough for the electricity grid isn't something people can do manually. It needs orchestration software. That's like an air traffic controller for computing. The utility sends a request. Can you reduce your demand by 100 megawatts for the next 15 minutes? The software decides which work can safely wait. Then the data center adjusts only those workloads while keeping the important ones running. One company doing this is Emerald AI, which works with Nvidia. Their software tags each job by how much slowing down it can tolerate. When the grid asks for less power, it has a few levers. It can turn down the clock speed of the chips so they do the same work more slowly using less power. It can slow down or pause the most patient jobs, or it can shift work to another data center altogether. It also runs a digital twin, a simulation of the data center, to predict how much power each move will actually save before it makes it. In a demonstration with Oracle and Phoenix, Emerald says it cut a data center's power use by 25% for three hours while still meeting its workload performance targets. Instead of curtailing workloads, another option is simply to change where the electricity comes from. Traditionally, batteries inside data centers have one job: to keep the computers running if the grid suddenly fails. They're called uninterruptible power supplies or UPS systems, and they act as emergency backup while diesel generators or other backup systems start. But what if those same batteries could help the grid before there was an emergency? Instead of sitting idle waiting for a power outage, they could temporarily power the data center whenever the grid was under stress. To the servers, almost nothing changes. To the grid, a large electricity user has suddenly reduced its demand. Companies like Verus are designing data centers around this exact idea. Rather than treating batteries as insurance policies, they're treating them as balancers in the electrical system. Sometimes the work can be moved somewhere else entirely. If another data center has spare capacity, some computer tasks can be shifted there instead. Engineers are also finding flexibility in places most of us would never think to look, like cooling systems. A building can be cooled a little earlier than usual, allowing pumps and chillers to reduce their electricity use when the grid is under pressure. None of these ideas are particularly revolutionary on their own, but together they could allow data centers to respond to the grid in ways that simply weren't possible a few years ago. The encouraging part is that these ideas are no longer confined to research papers. National Grid in the UK, working with Emerald AI and Nebius, showed that a cluster of 96 GPUs could cut its power by more than a third in under a minute. Emerald is now running a commercial scale pilot with Silicon Valley Power in Santa Clara at a data center running NVIDIA's AI workloads. Now 96 GPUs is a long way from a campus with 100,000 of them, but it does show that the control loop works. That doesn't mean the technology is ready to transform every data center tomorrow, but it does mean that the question has changed. It's no longer "can this work?" it's "will anyone actually make them do it?" Alright, so if these technologies are so promising, why aren't they already being used everywhere? The answer is that most data centers don't want to be flexible. Power is a really small part of their costs actually, so they'd rather pay more and keep the freedom to run however they like. Some are even building their own gas turbines on site to skip the connection queues, sometimes with very little environmental oversight. Flexibility is a hassle that they'll only take on if it gets them something they really want, like a faster connection, or if we force them to. Utilities need confidence that flexible data centers will respond as promised, quickly, repeatedly, and consistently. So far, they're saying they don't trust that data centers will actually follow through. For more than a century, electricity grids have been engineered around one simple assumption: generation follows demand. AI data centers could become the first major industrial loads capable of reversing that relationship. Instead of forcing the grid to adapt to them, they could adapt to the grid. They could also help reduce costs for everyone on the network by allowing us to use those networks more efficiently. But so far, they haven't volunteered, and I'm not holding my breath that they ever will. If we want this win-win, regulators, governments, and communities will have to make it the price of plugging in. If we can get that right, one of the grid's biggest new challenges could also become one of its most valuable assets. A big thank you to Plod for sponsoring today's video, and also to my Patreon supporters and YouTube channel members. You're the reason that I can take the time to dig into topics like this properly. If you'd like to support the channel too, the links are in the description. Thanks for watching and I'll see you in the next one!