Tyler Norris says regulators have been getting two different stories. On one side, they’ve been hearing that data centers are largely inflexible loads. On the other, last year the U.S. Department of Energy recommended flexible data centers, and EPRI launched its DCFlex initiative to demonstrate the same.
So he and a few other researchers wanted to know, What’s the potential for data center flexibility? And what benefits could it have system-wide?
In this episode, Shayle talks to Tyler, a PhD candidate at Duke University’s Nicholas School of the Environment and former vice president of development at Cypress Creek Renewables. In a recent study, Tyler and his co-authors found there’s enough spare capacity in the existing U.S. grid to accommodate up to 98 gigawatts of new industrial load (enough for multiple Project Stargates), if that load can curtail 0.5% of annual load to avoid adding to system peaks. Shayle and Tyler unpack the study’s findings, including:
- How much data centers would have to curtail and how often
- Options for shaving peaks, like colocating or leasing generation, spatial flexibility, and deferring or front loading training runs
- Speeding up interconnection if the data center is able to curtail load
- How bridge power could transition to peak shaving backup generation
Recommended resources
- Nicholas Institute for Energy, Environment & Sustainability, Duke University: Rethinking Load Growth: Assessing the Potential for Integration of Large Flexible Loads in US Power Systems
- Latitude Media: EPRI takes its data center flexibility project global
- Latitude Media: Who’s really paying to power Big Tech’s AI ambitions?
Credits: Hosted by Shayle Kann. Produced and edited by Daniel Woldorff. Original music and engineering by Sean Marquand. Stephen Lacey is executive editor.
Catalyst is brought to you by EnergyHub. EnergyHub helps utilities build next-generation virtual power plants that unlock reliable flexibility at every level of the grid. See how EnergyHub helps unlock the power of flexibility at scale, and deliver more value through cross-DER dispatch with their leading Edge DERMS platform, by visiting energyhub.com.Catalyst is brought to you by Antenna Group, the public relations and strategic marketing agency of choice for climate and energy leaders. If you’re a startup, investor, or global corporation that’s looking to tell your climate story, demonstrate your impact, or accelerate your growth, Antenna Group’s team of industry insiders is ready to help. Learn more at antennagroup.com.
Transcript
Stephen Lacey: Hey, it’s executive editor Stephen Lacey. Many of you were loyal listeners to a show I created and co-hosted for eight years called The Energy Gang. Well, we’re back Jigar Shah, Katherine Hamilton and I are excited to announce our latest project, a weekly news round table called Open Circuit. Every week we’ll break down how major projects come together, how deals and policies get structured, and what it takes to build critical infrastructure at scale, all through the lens of current events. You can subscribe to open circuit wherever you listen to podcasts or follow it at latitudemedia.com,
Tag: Latitude Media: podcasts at the Frontier of climate technology,
Shayle Kann: I’m Shayle Kann and this is Catalyst.
Tyler Norris: Well, we’re hearing right, the tip of the spear and what would really present a value proposition for the hyperscalers and the large co-location data centers would be if you can actually interconnect them to the grid more quickly if they’re able to offer this flexibility
Shayle Kann: Coming up: Could load flexibility add something like 90 gigawatts of additional data center capacity to the US grid?
I’m Shayle Kann. I lead the frontier strategy at Energy Impact Partners. Welcome. All right, so data centers generally like to operate all the time, not necessarily 24-7 from a grid perspective, but 24-7 from an operations perspective. And there are a bunch of reasons for that and it varies a little bit based on the type and the use case, but as a rule, that is true. However, there is a good case to be made that if they operated from a grid perspective very close to 24-7 but not quite all the way there, or if they had sufficient onsite generation to avoid pulling from the grid for just a little bit of time each year, you might be able to build a lot more of them much faster and more plus faster is basically the name of the game in data centers at the moment.
It’s for sure easier said than done though, and it’s not a panacea. It is really interesting though, and we’re starting to see glimmers of programs and tariffs that drive this across the US. Tyler Norris, who’s been on this podcast before, recently published a really good paper trying to quantify the amount of additional load you could add to the grid if that load were just a little bit flexible. He published it through Duke University and the numbers are pretty eye-watering, so I wanted to bring him on to talk about it because as always, the devil is in the details in these things. So here’s Tyler to talk through the details.
Tyler Norris: Tyler, welcome. Thanks for having me here. She
Shayle Kann: Excited to talk about large flexible loads with you. We’re talking about large flexible loads, but can we just say data centers? Data centers are most of what we mean here, right? I mean, I presume also this applies to manufacturing and other large loads, but really the purpose I assume is that we have an enormous spate of data centers that are soon to hit the grid. And your question had to do mostly with that. Is that an appropriate statement?
Tyler Norris: Yeah, I think what really opened our eyes were some of these forecasts suggesting that AI specialized data centers are going to be likely the single largest driver of US electricity load growth for the next five to seven years with some numbers suggesting about 44% of all US load growth will be from data centers.
Shayle Kann: Right. Okay. So we’ll talk about large loads, we’ll talk about data centers. We’ll kind of use ’em interchangeably even though they’re not exactly the same thing. But why don’t you start by just laying out the fundamental thesis of this paper and maybe the key findings as you see them, and I think we can dive into a bunch of the details and nuances from there.
Tyler Norris: Yeah, sure thing. Well, the origin of the paper was conversations with utility regulators in the southeast in the fall, and everyone’s trying to figure out how to deal with load growth. And what the regulators said to us is that we’ve heard that the data centers are 100% inflexible, and so we’re just assuming we have to plan for all of this as firm load. At the same time, the secretary of Energy advisory board had come out with the recommendations around data center flexibility, and then EPRI launched its data center flexibility initiative. And so there was this weird juxtaposition happening where they were being told data centers are a hundred percent inflexible, but others were suggesting they may be flexible. And so that was the prompt to us to dig into this and try to figure out what’s going on. We decided to do some modeling around it and initially we were just going to run this notion of curtailment at one to 5% of the max annual potential use of the data centers because that sort of aligns with the existing demand response program requirements for peak shaving sort of in that one to 2% range.
But–
Shayle Kann: Can I just clarify that what you’re saying is that’s a percentage of hours in the year, so one to 5% or one to 2% of total hours where effectively the data center from the perspective of the grid would’ve to shut off?
Tyler Norris: That’s right. Or yeah, one to 5% of their max potential energy use over the course of a year.
Shayle Kann: Oh, right. There is a distinction there, right? It’s not necessarily X number of hours where they’re at 0% capacity. It is relative to, I guess it’s a percentage of total load that they would pull if they’re operating on a hundred percent capacity, a hundred percent of the hours of the year. So there could be times when it’s 10 hours at 50% and other times when it’s five hours at 0%.
Tyler Norris: Exactly, exactly. So we assumed actually a hundred percent utilization, so constant load additions for the purposes of the modeling, but we were going to run it at one to 5% and that one to 2% being in line with existing demand response program requirements and 5% just being a high upper end scenario for calibration. And what we found that the numbers were so substantial in terms of how much new data center load you could add to the grid that we decided to run it at 0.5% and then again 0.25%. And to be honest, sha, we were very surprised by the amount of headroom that appears to be available on a large number of balancing authorities. And so we ran it for 22 of the largest US balancing authorities, which is about 95% of the country’s load. And what we found is that at 0.5% flexibility or curtailment of the new data center load, you could add up to 98 gigawatts of new data centers across the US and add 0.25% curtailment of the new load. You could add 76 gigawatts. And so what we’re talking about there is basically between three to five project Stargates, which is the mega data center initiative announced by President Trump and open AI in January.
Shayle Kann: Or I guess another way to contextualize that is we have, I dunno exactly what the number is today, but we have sub 30 gigawatts of data centers currently operating on the grid. So you’re talking about with a half a percent curtailment, tripling that plus or minus, something like that, which is above most of the forecast for at least toward the end of the decade, the next few years. But I want to dig into what that actually means. I think it is an astounding number, but there’s more to it than meets the eye. It’s probably harder than it sounds, but So the basic thing that you’re saying is that you looked at these balancing authorities and you said, how much more flat load could you add if you shaved the peaks? Basically not the peaks of the load, but if you curtailed when there are system peaks, and as we know, I think probably most of the folks listening to this podcast know the way that the electricity system works is it’s basically designed for peak. And so because we have peaks and valleys in demand and load, now the systems are designed to make sure that you have enough power, enough generation that is deliverable at the pickiest peak of the year. And so this is basically saying if you added a lot of data centers but they did not contribute to those pickiest peaks, how much could you add? And that’s where you get these astoundingly large numbers. Do I basically have that right?
Tyler Norris: That’s right. Shale. And just to put some numbers to that utilization rate that you mentioned, so you said it exactly right. We build the entire power system around these occasional extreme peaks that are driven by either extreme heat or cold snaps, especially these polar vortex events. But just to quantify it, so in 90% of hours, more than 30% of the power system sits unused. And we actually found that on average across those 22 balancing authorities, that the average load factor, meaning the average consumption over the peak consumption is 53%. So what that basically means is that in any given hour on average, approximately half of all generation and transmission infrastructure is unused on average in the US. And it’s actually worse than that because you’re not counting for all the reserve margin that’s on top of those existing peaks.
Shayle Kann: And some of that of course, is that we have some generators that are not designed to run all the time peakers, the definition of a peaker is to solve peak. And so those are built to operate at low capacity factor. So in a world where you do all add all of this flat load minus the peaky events is one of the results because the concept you’re saying is that we could add all that load to the grid without building any new generating capacity. So presumably what would happen then is you would end up operating all of your existing assets at higher capacity factor, which means you’d be running your peakers 24-7 or closer to 24-7, which would be expensive and suboptimal, I think, given how the system is built right now. Is that a challenge to this model?
Tyler Norris: Well, you could do it that way, but that’s not actually the way we did this. So what we did is we took the existing realized peaks on each of the balancing authorities of the past nine years and calibrated to that. So we didn’t want the new load added on top of existing load to ever result in a realized peak above those existing peaks. And so we were actually discounting all the reserve margin that’s available on top of the existing realized peaks. And so if you actually accounted for all the reserve margin, the numbers here would be higher, but they’d also be shaved off by accounting for other constraints like transmission constraints and the intertemporal constraints on the generation on the load. But yeah, so I just want to be clear, what we’re talking about here is not tapping into all of that dirtier inefficient reserve margin.
Now you would still, I mean if you add new load to any given system without a change in the generation mix, you’re by definition going to be running some units at higher capacity factor. But that’s a given regardless. By having the flexible load, you actually don’t have to run the dirtiest, most inefficient units as often, right? Because you’re not having to tap into that reserve margin. And then you can get into the question of, okay, for any given investment to change the generation mix, what are the second order effects of that? But that’s sort of a different question.
Shayle Kann: I mean, you mentioned before that you were initially modeling this off of typical demand response program frequency. Is there anything different about what you’re proposing in terms of the curtailment, not the frequency necessarily, but the type of curtailment versus demand response programs? Is there a shorthand version of the takeaway from this paper, which is basically enroll every data center in a demand response program and you’ve solved a lot of problems?
Tyler Norris: That would be a simplistic way of looking at it. So I think there’s one really key distinction from existing demand response programs, and that is that the vast majority of the loads that participate in demand response, they were planned as firm loads. So when the transmission provider ran their interconnection study and the transmission plan for that jurisdiction and then the reserve margin planning, they plan for that lotus firm. And then at a later date, once that load was online, it decided to participate in demand response for economic reasons. But what we’re talking about in this case is pulling all this up into the planning realm and it’s sort of like how we’re trying to get a lot more sophisticated with transmission planning. It’s like that, but trying to get more sophisticated with our load planning. And so just to articulate, I mean I think what we’re hearing the tip of the spear and what would really present a value proposition for the hyperscalers and the large co-location data centers would be if you can actually interconnect them to the grid more quickly if they’re able to offer this flexibility,
Shayle Kann: Which we’ve started to hear a little bit about. Some utilities are starting to say, here’s a program, here’s what I need. It’s either generic program, here’s a version of demand response, or it’s specific to a given customer given location, and they’re saying, we can interconnect you in eight years or we can interconnect you in four years if you’re willing to curtail somewhat. It’s that kind of thing that you’re talking about.
Tyler Norris: And so you see that in a limited number of jurisdictions. So Ercot has this controllable load service, which does make that sort of trade off explicit. Pg e is doing it with their new flex connect program primarily at distribution scale for EV chargers, but the hope is to expand it. And then Southern California Edison has a similar program, but we are lacking in sort of established official service offerings to large loads that specifically quantify what is that trade off for faster interconnected exchange for flexibility. And that’s what I think we’re hoping to see promulgated across more jurisdictions.
Shayle Kann: I guess I want to get a little bit more into what that curtailment actually would look like. So as you said, you ended up modeling it out on very infrequent or at least very low level curtailment 0.5% or 0.25%. How much are those times predictable? How would this actually work in your mind? Is it like it’s traditional demand response, the deal is like the utility or ER or whoever, it’s says, okay, we need to shut you off or we need to close you down to some degree X number of hours per year and we’re going to give you one day notice when it’s going to happen and we have a cap on how much we could do that? Is it that kind of a thing?
Tyler Norris: I think the program structure is likely to look something like that. I mean obviously the more controllable the load is on a real-time basis, the more valuable it’s going to be to the system for flexibility. But especially in the context when we’re talking about a limited number of extreme weather systems that are really driving these extreme peaks. Our weather forecasting fortunately is getting a lot better. So we can see incoming polar vortex events up to two weeks out and certainly within a week and begin to sort of plan around that. But the other finding, and we wanted to look at what does the list look like in terms of how much of the new load is curtailed for a given number of hours? Because I think initially we would’ve assumed that a lot of times we are talking about a hundred percent curtailment of this new load, but when we ran the numbers, so at that 0.25% load curtailment rate on average, we found that the number of hours in which curtailment would be required, some amount of curtailment would be 85 hours.
But in 73 of those 85 hours, at least 50% of the new load is retained. And then in 50 of those 85 hours, at least 75% of the load are retained. So we’re really talking about partial curtailment here and the vast majority of cases. And in terms of what this would actually look like for the load, I mean presumably with the data centers, they want to maintain high uptime for the servers. And so that’s why there’s been so much focus on onsite power and storage and this whole discussion around co-location and even with existing power plants. So that’s of course one of the primary options. But as you know, there’s a lot of progress happening and the ability to defer computational loads, especially some of the training loads that are more deferrable and batchable, the average duration of these extreme peaks, we’re talking sort of in the range of two and a half to five hours.
So you’re either deferring or front loading that training run, and then this spatial flexibility, the ability to distribute those workloads to one or multiple data centers located in different markets. Google obviously has been the most public and the advertising that capability, a little unclear if the other hyperscalers are developing it, but you would imagine we’ll move in that direction. And then I’ll just mention, of course there’s this other category which is they could simply reduce operations temporarily. That’s unlikely to be as preferable for the data centers. Although let’s be clear, a substantial portion of the load for data centers is the cooling infrastructure. So you can imagine here in the southeast are our worst sort peaking events. Now where you have the highest loss of load expectation is on winter mornings during polar vortexes. So you’re probably not going to need to run the coin infrastructure at a hundred percent utilization on an extremely cold winter morning for the data center. So you can get some headroom out of that. But as we know, the computational loads that are most flexible are the crypto mining operations and they’re now some of the most flexible loads on the entire power system, and they can go from like max draw to zero and one minute or a handful of minutes.
Shayle Kann: One thing that we’ve been seeing a fair bit that I’m curious whether you have been seeing as well, so lots of new data centers are, you mentioned this right, bringing their own generation, and that could be renewables and batteries, but it’s often natural gas generators and they’re doing that to increase time to power, I’m sorry, decrease time to power, increase speed to power I should say. And in some cases what they’re doing is operating off grid or semi off-grid. They’re doing bridge power, right? That’s the term people are using for it where they bring their own generation until they can get the full grid connection in place. And so one thing we started to hear about a little bit, which might be a solution to what you’re describing is if they’re going to bring bridge power anyway, they’ve already got generating assets sitting on site.
So maybe what you do is you operate those assets at full bore for as long as you need the bridge, but once the grid connection is there for you and all the equipment is there for you, then that same asset turns from a high utilization bridge power thing to a very low utilization curtailment thing and you just operate it half a percent of the hours of the year. And that would be difficult economically in any other condition except that you already basically amortized it because you were using it as primary power for some period of time. Do you see that as being a scalable solution here?
Tyler Norris: Absolutely. I think it’s critical to mention this could be entirely what we call provisional a temporary arrangement until all the firm upgrades are done. And so I think that’s likely what would be even perhaps more economical is if it was possible to do that on a leasing model where the data center would lease onsite power from a third party, that then once the firm upgrades are in place, that onsite power source could be taken to another location. And that’s what we’re hearing actually was just talking to someone in California a few weeks ago, is developing a business model just like that, but around lithium ion battery storage. So the idea is they truck in these batteries while the large load is waiting on the firm upgrades to get done, and then once they’re in place, they will take that to another customer. So that would be from a resource efficiency standpoint, that would be the better way to do it. Obviously you can’t do that with all onsite power infrastructure, so there’s going to be some variation there.
Shayle Kann: Okay. So then I think the other big question here is that to some extent the fundamental premise of the research that you’re doing was that the constraint on this load growth is generation, is the ability to serve enough power to meet the peak load that these new data centers and other large loads will introduce. That certainly is going to be true in some cases, but it is also certainly true that in many cases the main constraint, the rate limiter is not generation but deliverability, it’s transmission and distribution infrastructure. So do you have any sense of the 98 gigawatts that you could add from a generation perspective? How much of that could you actually add or how much would you end up just getting constrained by TD anyway?
Tyler Norris: There’s really no way to know without running it. And every jurisdiction is different. Every portion of the network is slightly different, so it’s hard to put a number on it. If I was forced to, I’d probably say something like 10% is going to be eroded, but you just won’t know.
Shayle Kann: Is that all really? You think it’s that?
Tyler Norris: That’s just my very rough back, yeah, finger first order estimate, just me talking and we want to be clear. Well, let me just say a couple of things. One is we want to be crystal clear that we don’t want to discourage investment and new generation and transmission infrastructure. We’re going to need it for a whole bunch of reasons, including all the other loads that are coming in as well as decarbonization priorities and just improvements in reliability. And we also acknowledge there are very often going to be local network constraints that are going to be going to be a barrier, but it’s critical, and you said the right word, deliverability. It’s a misnomer by the way, because it doesn’t actually imply that without deliverability, you can’t deliver electrons to load. What it means is that under the most extreme system conditions during both contingencies and the need to run all local capacity generators at the same time that you don’t have any bottlenecks. And it’s not to say that this conditions don’t or never occur, they’re just extremely rare. So it’s a very, very rigorous study criteria. And the bottom line is we can get a lot of load and generators online, by the way, without having to have upfront deliverability. And then you can work on those full deliverability upgrades in the background. And that’s exactly what Ercot does for generation. They don’t require upfront deliverability and it allows them to get generation online much more quickly and at much higher rates.
Shayle Kann: I guess final question for you is you described at the beginning, there’s this interesting dichotomy in the market right now where there’s a premise, and I’ve heard this from many of the data center developers and hyperscalers as well, that we are not flexible load, we need to operate 24-7. On the other hand, there’s this constraint and everybody is starting to say, oh wait, we’re going to have to figure out how to square this circle. It’s been a few weeks since you published this report. Have you had reactions from, I guess either from side of this equation, from the grid operators or from the data center world on how much does this suggest we should just move quickly on demand response type programs? Or as you said, the type of program where there’s a deal in exchange for faster time to power? Is there momentum around this or do you feel like you’re fighting an uphill battle for some reason?
Tyler Norris: Yeah, great question. So I want to say one thing upfront. I don’t know where this idea came from of 24-7. They are not 24-7 loads, and it’s critical to clarify this. So the servers may be 24-7, right? The chips themselves, but there’s an enormous amount of cooling infrastructure and other infrastructure around those servers. So actually like Lawrence Berkeley lab and their recent data center and energy usage report in December, those congressionally mandated, they put out the number 50% utilization rate. EIA and E three have used a number closer to 85% utilization. But I just want to be clear, are not, you’re not running the data centers at the, were a hundred percent potential max draw at a hundred percent of the hours. So I’m hopeful that we can just further clarify. I think the regulators in particular have been very confused by this. So good to establish that.
But to your question, it’s a really interesting dynamic, right? I mean, we’re in such a hyper competitive environment where I think the hyperscalers and even some of these co-location data center developers, they don’t necessarily want to be a hundred percent forthcoming about the extent of their capabilities and to the extent they are going to be forthcoming about them, they want to be compensated for those capabilities wherever possible, but they certainly don’t want their competitors to know what they’re fully capable of. And the other thing is I think the whole market environment in terms of what type of contracts everyone is used to, whether it’s the owner operator or their financing counterparties, it’s like everyone is just much more used to firm service. And that’s just kind of the gold standard. And I think part of it is that we need to get more of the parties used to these sort of quasi firm arrangements and to actually see those banked so that more market participants can get comfortable with it.
I think the tip of the spear there is likely to be, as you said, faster speed to power and exchange for flexibility. And once we see some of those deals set up and executed and we see the capabilities are working, I think the hope is that that will then proliferate beyond necessarily being restricted just to speed to market. But it’s an interesting dance right now. I mean, I hear about it almost every week. These conversations are happening. There is happening primarily on a bilateral confidential basis between large loads and transmission providers. So there isn’t necessarily a big push, I think, from either of those parties to make these public established tariffs. But I think that is where we’re going. And we were quite surprised last week when Duke Energy said at a public event that they are now going to require all new hyperscale loads above a hundred megawatts that participate in demand response. And we’re hearing a lot of interest from other state regulators in this too. So I think hopefully we’ll get some of these tariffs set up over the next year or so, and then the market will kind of evolve from there.
Shayle Kann: All right, Tyler, super interesting. I think you sort of hit this all at the right moment for this conversation, so I’m glad you did and put some numbers to it. We will talk more about it as this dynamic world evolves. But thanks so much for the time.
Tyler Norris: Thanks, Shayle.
Shayle Kann: Tyler Norris is a PhD candidate at Duke University’s Nicholas School of the Environment. He’s also a former VP of development at Cypress Creek Renewables. The show is a production of Latitude Media. You can head over to latitude media.com for links to today’s topics. Latitude is supported by Prelude Ventures, prelude backs, visionaries, accelerating climate innovation that will reshape the global economy for the betterment of people and planet. Learn more@preludeventures.com. This episode was produced by Daniel Woldorff. Mixing and theme song by Sean Marquand. Stephen Lacey is our executive editor. I’m Shayle Kann, and this is Catalyst.
