Everyone’s talking about the rise in electricity demand, especially from data centers. However, far too few are talking about what to do about it. In this episode of the Energy Capital Podcast, I sat down with Tyler Norris, a fellow at Duke University’s Nicholas Institute and co-author of a recent study Rethinking Load Growth.
What Tyler and his team found is extraordinary: by curtailing just 0.25% to 1% of a data center’s annual load—primarily during the most stressed hours—Texas could add up to 15 gigawatts of new data center load in ERCOT without adding any new capacity. That’s nearly the total demand of the entire city of Houston.
Even more striking: 90% of the time, at least half the load would still be retained. It’s not about shutting everything off, it’s about smarter operations during a handful of peak hours. This is exactly what we need as we try to integrate unprecedented load growth. Tyler’s research offers a better way forward.
We dove into the details:
The difference between legacy 24/7 data centers and newer AI-focused facilities, which offer greater flexibility.
How heating and cooling loads vary seasonally, and why winter peaks—not summer—are currently the biggest challenge in Texas.
What load factor really tells us about underutilized grid capacity and how ERCOT compares to other U.S. markets.
Tyler also explained how even small changes in behavior or technology, just 0.25% to 1% shifts in demand, can unlock massive headroom, especially during those critical winter mornings and summer evenings.
We talked policy, too. From ERCOT’s Controllable Load Resource program to proposed emergency response products with 24-hour notice, there’s momentum to better align incentives for large loads that can respond during grid emergencies. Tyler called this “speed-to-power” a key motivator, if data centers can get online faster by offering flexibility, that’s a win for them and the grid.
But it’s not just about data centers.
Tyler and I also explored how residential winter weatherization, efficient heating upgrades, and demand response on the household level could free up even more capacity for Texas’ growing electricity needs. This builds on ideas I explored in “Texas Needs a Vision for Customer-Side Solutions” where I argued that the most affordable path to reliability includes smarter demand-side management. We will need demand and supply side resources. We also talked about a novel idea in Texas: whether data center operators could pay into programs that reduce peak demand systemwide, benefiting all consumers.
Finally, we discussed ERCOT’s national leadership in interconnection performance, outpacing other U.S. grid operators by a wide margin, and why other states should take notice.
Tyler’s work is a big deal. It challenges some major assumptions in energy policy and offers a roadmap for smarter, more flexible load integration. As electricity demand soars, this kind of thinking is going to be critical not just for Texas, but across the country.
Read the report: Rethinking Load Growth – Nicholas Institute
As always, thanks for listening and please share this episode if it helps you understand this moment a little more clearly. We’re going to need every tool on the table to build a cleaner, more resilient grid.
Timestamps
00:00 – Introduction & why this matters
02:00 – Tyler’s background & origin of the study
04:00 – Takeaways from Rethinking Load Growth
07:30 – Exactly how much capacity could be freed up in ERCOT
09:00 – Does flexibility mean complete shut down or something else?
11:30 – Are data centers truly inflexible?
16:15 – The ERCOT grid is vastly underutilized in the vast majority of hours in the year
22:30 – Grid is more underutilized in winter, but peaks could be higher
25:00 – Residential demand flexibility could free up even more headroom, data centers that are inflexible could pay for residential demand reductions
30:00 – Solar creating more headroom, reducing loss of load expectation (LOLE) in summer; winter is bigger problem, particularly in the South
35:00 – Why demand response stalled in the 2010’s
36:30 – Policy solutions: speed-to-power / speed-to-interconnect for flexible loads and Controllable Load Resources (CLRs) in Texas
39:00 – “All-of-the-above” for demand side resources
42:00 – Differentiating types of large loads
46:00 – Can residential customers benefit from large load growth?
50:00 – ERCOT’s interconnection success story
55:00 – DeepSeek & the future of AI efficiency
58:00 – Final thoughts and what’s next for Tyler’s research
Show Notes
Research & Reports
Rethinking Load Growth (Tyler Norris, Duke University Nicholas Institute)
The foundational report explored in this episode. Quantifies how much flexible load (like AI data centers) could be integrated into U.S. power systems.
EPRI Data Center Load Flexibility Initiative - Ongoing research effort exploring demand-side solutions and flexible operation strategies in data centers.
Grid & Flexibility Context
Overview of Demand Response in ERCOT - summary presentation detailing ERCOT’s demand response programs, including the role of CLRs and their impact on the Texas grid.
Load Resource Participation in the ERCOT Markets - outlines how load resources, including Controllable Load Resources (CLRs), participate in the ERCOT markets, providing ancillary services and contributing to grid reliability
NERC Reliability Risk Priorities Report 2023 - Cited in the episode as identifying controllable loads as a key reliability solution. Reinforces ERCOT’s perspective that CLR participation is an optimal reliability pathway.
Legislation & Policy
Texas Senate Bill 6 (2025) - Referenced in the conversation as a bill addressing data center interconnection and demand-side solutions, including a 24-hour emergency response service.
Tyler Norris Congressional Testimony (2025) - Tyler’s formal testimony on interconnection delays, praising ERCOT’s “connect and manage” model and calling for structural reforms nationally.
FERC Order No. 2023 - National interconnection reform discussed indirectly in the conversation, setting new timelines and process changes to address backlogs.
Also Mentioned…
Abundance by Ezra Klein and Derek Thompson - Referenced as part of the broader national conversation about energy abundance, demand growth, and large flexible loads.
SPEER Incremental Demand Response Report (May 2015
Toward a More Efficient Electric Market (June 2013)
Transcript
Doug Lewin (00:03.938):
Welcome to the Energy Capital Podcast. I'm your host, Doug Lewin. My guest this week is Tyler Norris, a researcher at Duke University, who, along with a team of researchers at Duke, produced an impactful report called Rethinking Load Growth. What they looked at here is what everybody in the energy world is talking about these days: how are we going to accommodate all of these data centers that are coming?
And Texas is likely going to get more than its share of data centers than a lot of other places for a variety of reasons. It's easier to interconnect here. There's a lot of low-cost renewables available. There’s a lot of things that Texas has going for it. But how can we accommodate all that load without causing grid reliability problems or raising costs?
Tyler and his team in Rethinking Load Growth talked about one of the ways to do that: if these data centers have even a small amount of load flexibility, it frees up large amounts of capacity. We talked about how even a quarter of 1% or half of 1% could free up many gigawatts of space. The same holds true when you get down to the residential side—if we could do some things there to reduce demand, we can also free up space for data centers.
This is an active conversation right now at the legislature Senate Bill 6, House Bill 3970. These are bills that include different concepts for how to integrate large loads. It was a very timely conversation. I hope you enjoy it. I learned a lot from Tyler and from the paper. We'll put a link to that in the show notes. And as always, please give us a five-star rating. It really does help people find the podcast.
This is a free episode, but it is not free to produce. We have both the Texas Energy and Power Newsletter and paid episodes of Energy Capital Podcast available to you if you subscribe at douglewin.com. And thank you so much for listening, and thanks for your support. Let's dive in.
Tyler Norris, welcome to the Energy Capital Podcast.
Tyler Norris (02:00.768):
It's great to be here thanks so much, Doug.
Doug Lewin:
Yeah, thanks for doing this. Excited to talk to you. This report that you’ve put out—Rethinking Load Growth: Assessing the Potential for Integration of Large Flexible Loads in U.S. Power Systems—obviously got a lot of attention, lots of folks talking about it all around the industry.
Before we jump into that, just very briefly, tell the audience a little bit about yourself and what your focus is there at Duke, at the Nicholas Institute.
Tyler Norris:
Yeah, sure thing. I'm a fellow and pursuing my PhD in electric power systems at Duke University's Nicholas School. I'm a somewhat non-traditional PhD student in that I came into the program with about 12 years of experience in the energy sector. I started at the Department of Energy working on tech commercialization programs, spent a couple of years at SMB Global doing their North American electricity market outlook, and then spent about five and a half years in large-scale solar and solar-plus-storage development at Cypress Creek Renewables.
There, I ended up doing a lot of work in front of public utility commissions, working with consultants on electricity simulation models, and trying to get better outcomes for resource planning and interconnection. That was the part of the job I loved the most. I figured if I could go deeper in a PhD program, I could be of greater service.
Just briefly—a lot of interest has been on grid interconnection, especially for large generators. But it turns out that a lot of the study methods for large generators are similar for large loads from an interconnection standpoint. So that's why we’ve gotten into the large load interconnection side of this too.
Doug Lewin (03:37.782):
Yeah, it’s really fascinating. For the longest time, nobody really talked about large load interconnection it wasn’t a thing that came up much unless you were way, way in the weeds in some technical forum. And now it is the topic du jour, which brings us to your study Rethinking Load Growth really getting into large flexible loads.
Why don’t we start at the highest level? What are some of the main takeaways? What do you want people to know about this report, some of the key findings?
Tyler Norris:
Sure. Maybe I’ll just give a quick backdrop. This emerged from conversations with state utility regulators and other state-level regulators last year through the Nicholas Institute, which is a think tank at Duke University. Everyone's talking about how to accommodate load growth. What the regulators were saying is that we’re hearing the new data centers are 100% inflexible, that we have to treat them as firm loads.
But at the same time, we were seeing EPRI launch its data center flexibility initiative, and other announcements—including from the Secretary of Energy’s advisory board—recommending progress on data center flexibility. So there was kind of this juxtaposition. That was the prompt for us: dig into this and figure out what the potential might be from a technical standpoint.
We do a qualitative characterization of why there’s a thesis right now that AI-specialized data centers can be more flexible than the sort of legacy computational loads. We also did some modeling, which I think caught people’s attention primarily. We looked at how much new load you could add to any given electricity market or balancing authority if the load was able to be flexible—or more specifically, curtailable, for a limited number of hours in the year during the most stressed grid conditions.
Tyler Norris:
In traditional demand response programs for peak shaving, they’re typically in the range of 1 to 2% of the max uptime of the new load. We’ve seen that in a couple key demand response programs. So we ran it initially assuming the new loads—hypothetically—could be flexible for 1% of their maximum uptime in a given year.
The numbers were so substantial in terms of how much new load you could add to a variety of balancing authorities that we decided to run it at 0.5% of the max potential uptime, and then again at 0.25%. And it turns out we were quite surprised by the volume.
Just to put numbers on it at 0.5% curtailment, we looked at 95% of the country’s load across 22 balancing authorities. It aggregated to about 98 gigawatts of new load you could add. At the 0.25% level, it aggregated to 76 gigawatts.
As people are probably aware, that exceeds even the upper-end forecasts for data center load growth we expect over the next five to seven years.
We want to be clear—this is a first-order technical potential assessment. There are various limitations to it that we talk about in detail, including that we didn’t account for transmission constraints. That would require a much more extensive and longer-term study. And you couldn’t even do it for 22 balancing authorities because it would take so much time. But the headline is sort of… yeah, go ahead.
Doug Lewin:
Yeah, yeah, no, you're all good. This is the Energy Capital Podcast. We talk about Texas and ERCOT, and I know you've been on a bunch of podcasts talking about the national picture. We're going to really drill into Texas and ERCOT here.
So at that 0.25% level, 6.5 gigawatts of data centers could come online without any new capacity needed. At a 0.5% reduction, 10 gigawatts, and at 1%, 15 gigawatts.
Those are pretty extraordinary findings. Can you talk a little about what that 0.25%, 0.5%, and 1% actually mean? Like, 1% of the hours out of the year—there’s 8,760 hours so you’re talking about 87 hours.
But is that 87 hours of the data center being shut off completely? Or is it maybe the data center running at 50% for 100-and-something hours? What exactly does that look like?
And I also just want to make sure I’ve got this right—that basically, you wouldn’t need to add new capacity to bring on 6.5, 10, or 15 gigawatts at those curtailment levels. You're not saying we don’t need new generation, of course we do, but that with load flexibility, we can accommodate a significant amount of data center growth more efficiently. Is that the right way to look at it?
Tyler Norris: (9:57.078)
Yeah, thanks Doug. Great prompt there.
So first, yes it's a percent of the maximum possible uptime. One of the surprising findings was we assumed you’d have to shut off the new load entirely during a number of hours. But what we found is that we’re primarily talking about partial curtailment events. At the 0.25% level, that translates to about 85 hours on average in which some amount of curtailment would occur. But in about 90% of those hours, at least half of the new load is retained. And that 90% figure actually held pretty consistent across the different curtailment limits. In many cases, 75% of the load is retained even during curtailment events.
So it's rare that you're talking about going all the way to zero, or even anywhere close to zero. This is interesting because it opens the door to partial flexibility. For example, if you were using battery storage, instead of sizing it at 100% of the max draw of the load, even 20 to 25% could give you a lot of flexibility. And the duration of these events on average is relatively short, three to five hours. Of course, it varies from balancing authority to balancing authority. And yes, that’s an average—so extreme weather events might push it higher. But overall, that average is informative.
Doug Lewin:
Yeah, that’s super interesting. Batteries would be one way to do it—putting batteries onsite. You could also have other types of onsite generation.
What I found especially interesting in your paper was the discussion of different kinds of data center load. You started this by saying everyone’s been saying data centers are flat, 24/7, 365. I hear that all the time too “they can't be moved around.” But in your paper you get into the differences between inference and learning, and also the huge cooling loads, which I don’t think are talked about enough.
So yes, batteries. But there’s also some load onsite that could be shifted. Can you talk a bit about that?
Tyler Norris: (12:07.298)
Yeah, sure. Maybe we can take that in a couple parts.
First, I want to say that data is still limited. This is just a feature of the industry right now, and a constraint facing all independent researchers. Lawrence Berkeley National Lab recently did a congressionally mandated data center energy usage report it’s probably the best public analysis we have.
But even they prominently state that there are a lot of limitations because we just don’t have good hour-by-hour, season-by-season usage data for data centers.
That said, let’s start with cooling loads. Because in some ways, it’s fair to say these are 24/7, 365 constant loads, but even that starts to break down a little when you look at interseasonal variation.
It’s kind of intuitive put a data center in the South Pole versus the Sahara Desert, they’ll have different cooling needs. And this extends to other geographies.
You can go right now Google, to their credit, publishes the Power Usage Effectiveness (PUE) of all their data centers around the world each quarter. In their North Carolina facility, for example, PUE is around 1.06 in winter versus 1.25 in summer.
Doug Lewin:
Which means that in the winter, they're getting much more efficient use of the energy they're using more of the energy per unit of computing power than in summer, right?
Tyler Norris: (14:05.026)
Yeah that’s right. There could be other contributing factors, but presumably the difference in cooling needs is a big one. So we can see some variation.
Then, with legacy computational loads—CPU-dominated, real-time cloud services, especially global data centers—you got constant flat loads. But with AI-specialized data centers that are GPU-heavy, that picture starts to change.
Some are used entirely for training workloads; some are a mix of training and inference; others might be all inference. It’s still a bit unclear. For training—like neural nets and foundational models—power usage can vary dramatically, even sub-minute. It can jump from max draw down to zero. With inference loads, they also vary, and demand is still uncertain. I don’t think even the hyperscalers fully know what global inference demand will look like.
There’s also cooling load, batchable workloads, and other flexible components. So bottom line: not all of these are 100% flat, around-the-clock loads.
Doug Lewin: (16:14.422)
Yeah. So this brings me to the next point here, which you make really well in the study—this key metric of load factor.
I think we’re at a moment right now where everyone’s talking about abundance—Ezra Klein’s book is out, I’m going to an event tonight with the Abundance Institute, which is a right-of-center think tank. There’s this question of: how do you do abundance if you have scarcity on your grid at some times?
But the fundamental point is—we actually do have an abundance of energy for the vast majority of the hours in the year. It’s really about managing the peak.
You put some great numbers and graphics on this in the report. Load factor is basically how much energy is used compared to the overall functional capacity of the grid—not nameplate, but usable. And you show the load factor for ERCOT at 53%, which turns out is right at the median or mean, can’t remember which.
So our grid is vastly underutilized. You write in the report:
"A system's potential to serve new electricity demand without capacity expansion is determined primarily by the system's load factor—a measure of the level of use of system capacity—and grows in proportion to the flexibility of such load, i.e., what percentage of its maximal potential annual consumption can be curtailed."
So basically, we are in a situation where ERCOT could take on a lot more data centers, using the slack that’s there—as long as they come in with some amount of flexibility.
Can you talk a little about ERCOT’s potential relative to this load factor metric?
Tyler Norris:
Yeah, sure. I appreciate you walking through that, because I think intuitively we all know the grid is built around peak, but when you sit with the numbers, it’s kind of extraordinary. So, to put it another way: more than 10% of the system is built to serve about 35 hours a year of extreme peak load. That means there is headroom—it’s just outside those rare extreme swings, like heat waves and cold snaps.
ERCOT’s load factor landed at 53%, and just to define it: it’s the average demand over the maximum peak demand, based on nine years of hourly load data. That 53% was both the average and median. ERCOT actually has a higher load factor in the summer—around 65% on average. But what’s interesting is ERCOT is one of the few systems where the winter load factor is actually just below that average, which means more headroom is available in the wintertime.
That’s partly because you do get those occasional polar vortex-style cold snaps—Winter Storm Uri being the most dramatic example—but outside of those, there’s a lot of unused capacity. So yeah, we could peel back more ERCOT-specific metrics. But the takeaway is: if a new load can be flexible for even a small amount of time, you can add a meaningful amount of demand to the system. Now again, as I mentioned earlier, the numbers we quote—6.5, 10, 15 gigawatts—those should be viewed as first-order technical potential estimates. When you layer in transmission constraints and intertemporal generation constraints, some of that potential gets eaten away.
But even if it’s only half of those numbers, that’s still substantial. Also worth noting: our headroom estimates do not include the reserve margin. So in some cases, if you did factor that in, the true headroom could be even higher.
Doug Lewin: (20:34.774)
Yeah the load duration curve you included in the paper is really amazing. It reminded me of when I worked at SPEER (South-central Partnership for Energy Efficiency as a Resource). We did reports on demand response potential and I remember being struck by those curves.
For listeners, I’ll include a screenshot and a link to the paper in the show notes. But it looks like a skateboard ramp—or a hockey stick, but flipped. On the left side, for a few percent of hours each year, you get these huge peaks. Then it quickly drops off, and for the vast majority of hours, there's massive headroom.
ERCOT has more room than most other systems. And like you said, utilization is even lower in the winter, which suggests something for future research: it would be useful to break this out by season, because I imagine that ramp gets even steeper in the winter.
That brings us back to the question of: how do you manage those peak hours? So if you want to say anything more about winter vs. summer, go for it. But I want to layer in another question for you to take after that…
Tyler Norris: (22:58.336)
Yeah, really quickly on the winter vs. summer comparison: we didn’t include this in the report, but we did look at the average duration of extreme peaks and of curtailment events.
We found that the average duration is substantially lower in the winter than in the summer. That tells us the nature of the peaks in winter is spikier—you get a cold snap for a few days, and especially in the mornings (6–9 a.m.) when everyone wakes up and turns on heating, you get this quick spike—but then it dissipates fast.
In contrast, summer peaks are more sustained—they extend from mid-afternoon to the middle of the evening. So the loss of load expectation is increasingly shifting to winter mornings.
But the good news is that a little bit of flexibility in the wintertime gets you a long way.
Doug Lewin:
Yeah. What I’m wondering, Tyler—and I don’t know if you guys looked at this for the report—is whether there’s potential for load flexibility throughout the system, not just at the data center level. Could there be some kind of fund, for example, where data centers help reduce residential peak load, which would free up even more headroom for them?
And then there’s also the broader question of what AI itself could do—optimizing loads in ways that don’t reduce comfort, and might even improve it. Like pre-cooling, for instance. I mean, you could imagine a future where AI is kind of self-aware of its energy use and shifts its own cooling load. The potential is huge.
But the main question I want to ask out of all this is:
Isn’t there real potential for residential load flexibility to help free up space for data centers?
Tyler Norris:
Yeah, I love that question because you can start to think about flexibility in a variety of ways. Why are we focusing on the large loads right now? Partly because so much of the near- and medium-term load growth appears to be coming from these very large commercial customers.
Some of the numbers are eye-popping—up to 50% of U.S. electricity load growth over the next five to seven years could come from AI-specialized data centers, according to some forecasts. So this growth is coming fast, and we need to plan for it better. But the opportunity it presents is that these loads are so large-scale and under the control of sophisticated operators—who also have a high willingness to pay. They can incorporate new technologies and systems more easily. So from a bang-for-your-buck standpoint, if you can get one 200-megawatt data center to be flexible, that might be a better return than signing up thousands of residential customers for demand response.
That said, if data centers are unwilling or unable to offer any flexibility themselves, one option is for them to procure flexibility from others—like a virtual power plant or distributed capacity procurement.
So maybe they pay customers to install smart thermostats or smart water heaters, right? Those water heaters could pre-heat two hours earlier before that 6 a.m. winter peak instead of coming on when everyone wakes up. Even weatherizing homes would help—anything that puts downward pressure on peak demand creates system headroom. And again, we should emphasize—part of why AI-specialized data centers could be more flexible is because training workloads are significantly more batchable and deferrable than real-time inference or cloud services.
There’s solid literature on this—it’s proven that in a lot of cases, you can front-load or defer computational training loads by a few hours and reduce demand during peak.
Doug Lewin:
Yeah, yeah. That’s exactly the point I was getting at—training vs. inference vs. cooling, etc.
Before we move on, I just want to circle back to something you said about weatherization and resistance heating. I loved this study—it answered a lot of my questions but raised like 50 more. So for researchers listening—and Tyler, maybe for a future paper—I'm especially curious about ERCOT, where we’re adding huge amounts of solar and storage.
We’re at over 30 gigawatts now, on our way to 40 or 45 GW in the next 18 months or so. It seems like winter is now becoming the real problem, not summer. With all that solar/storage coming online, the summer risk is going down, but the winter spikes—especially early morning—are getting more prominent.
We’ve got some demand flexibility on the residential side, but it really comes down to heating loads. Can we replace inefficient resistance heat? Because if we can, we create more headroom. I’ve even started talking to data center folks about this. If they really want to interconnect in ERCOT, we need to be talking about winter nights and mornings.
Did this show up in your study? Or does this suggest that we may need to reassess how we’re looking at seasonality as this resource mix shifts?
Tyler Norris: (31:51.038)
Yeah, absolutely. This is a critical aspect—this shift in loss of load expectation from summer afternoons to winter mornings is now happening in a growing number of jurisdictions.
The Southeast has seen it too. It’s partly because solar and storage are victims of their own success—they’re reducing summer peak risk, which is great. But what’s left now is those rare, intense winter spikes. And like you said, those resistance heaters—there’s a big opportunity there, but they are inefficient. And when they all come on at once, it creates a challenge. We all know: solar doesn’t produce much from 6 to 9 a.m. in winter. Maybe if we changed time zones… but I doubt that’s catching on.
The good news is those morning peaks are short—in some systems, even 2-hour battery storage can make a meaningful dent. And definitely 4-hour storage can help. It’s not sexy, but just getting people off those strip heaters—honestly, I think people would be surprised how much progress we could make with that alone. Also, a quick example from here in the South: Winter Storm Elliott a couple years ago was the largest rolling blackout event Duke Energy has seen.
The primary cause? Thermal unit failures, mostly gas and coal—similar to Winter Storm Uri in Texas. But there was also a big morning spike because of that polar vortex. So, even though solar doesn’t produce during winter mornings, it could help charge batteries the day before—or even save fuel in the hours leading up to an event, which helps with reliability. We need to get better at analyzing the synergistic capacity value of different resources. Groups like E3 have done some great work on this, but we still don’t fully account for the interactions between solar, storage, thermal, demand flexibility, etc.
Doug Lewin: (34:17.206)
Yeah, completely. And I think one of the big takeaways for people listening is that even small, incremental changes—a half-percent here, 1% there—can free up significant headroom, as long as they are timed correctly.
It doesn’t matter what the measure is—energy efficiency, demand response, batteries if it matches the time of system need, it adds real value.
You wrote in the paper that demand response participation stalled in the mid-2010s. Can you talk about why that happened?
Tyler Norris:
Yeah. From an academic standpoint, you’d want to isolate all the variables. But I think the overarching trends were pretty clear.
First, we became long on capacity in a lot of markets. That drove capacity prices down. At the same time, we were in a low-growth environment. So when capacity market prices decline, there’s less incentive for demand response participation. There were also new restrictions on participation from aggregated DERs in some markets, which had an effect. But the biggest macro factor was just low growth and low prices. That’s all changing now, though. Just look at PJM—huge increase in capacity market prices in the past couple years. Even aside from price spikes or supply chain issues, we’d expect more participation. But more significantly, we’re seeing more interest in load flexibility because of the speed to power advantage—getting online faster.
So let’s talk about that. Is that kind of a core policy solution?
If you’re talking to a policymaker trying to figure this out—someone who wants these data centers because of the economic development, national security, and tax base benefits—but who’s also very worried about another Winter Storm Uri or Winter Storm Elliott, what do you tell them?
What are the policy mechanisms that can help us manage this responsibly? Is offering faster interconnection in exchange for some load flexibility one of them?
Tyler Norris:
I think so. That’s one of the conversations we’re hearing, and ERCOT, as is often the case, is out on the leading edge here in terms of creating service constructs.
So the Controllable Load Resource (CLR) product—that’s a service ERCOT created. They improved it a couple of years ago to more explicitly quantify and characterize the trade-off between flexibility and faster interconnection.
And I think in that case, they were trying to design a product that could allow new large loads to get online within two years.
That’s meaningful. I’m sure there are ways that product can be improved or optimized, and there are always challenges with real-time control. But maybe there are variations we could develop that still preserve the core idea: faster interconnection in exchange for measurable flexibility.
Just one more thing—back to the winter conversation. With polar vortex-style events, we can often forecast them days in advance. Sometimes even a week or more. Forecasting is getting better, especially with the help of AI.
So, for those events, we may not need fully real-time controllable loads. Even day-ahead or multi-day-ahead flexibility could add a lot of headroom.
Doug Lewin: (39:50.742)
Totally. And I think this is where we need to talk about all of the above on the demand side, just like we do on the supply side.
There are controllable loads—and those are great—but there are also other forms of demand flexibility. You actually quote ERCOT in the paper, this was from a NERC presentation I think. and they say:
“The optimal solution for grid reliability is more loads to participate in economic dispatch of Controllable Load Resources.”
That’s a strong statement: optimal solution for grid reliability. But of course, there are other approaches too. In Texas, we’ve got Senate Bill 6 on the table right now. (We’re recording this on April 3rd, and I might already have articles out by the time this episode airs.)
I think that bill is mainly focused on large loads and their integration. I don’t think it fully does what we need, but there’s one part I do like: it creates a new emergency response product with 24-hour notice. Today, Texas has about a gigawatt of emergency response service—about half of that is demand-based—and the current notice periods are 10 minutes and 30 minutes.
But with these winter storms, like we were just saying, you see them coming. A 24-hour notice product could open up participation from loads that aren’t suitable for real-time dispatch but are willing to cut usage in an extreme circumstance. That’s a form of flexibility, too. So I think the main point is that there are many different kinds of demand flexibility. They each have different value and should be valued accordingly. But we need all of it—or else we’re going to be very limited in the amount of load we can add.
Tyler Norris: (41:47.798)
Yeah, I think that’s really well said.
And to your point about the 24-hour notice product—look, we probably shouldn’t expect AI-specialized data centers to participate in demand response just based on economic incentives alone.
Unless prices go extremely high—which could happen in submarkets—or unless we redesign price signals, it’s unlikely that price alone will drive widespread participation.
That said, there’s some interesting research potential around that. If the addition of a large load increases the loss of load expectation for everyone else, you could argue for a price signal or fee on that new load, based on the burden it imposes on system reliability.
That would be a shift in how demand response pricing works, but it's worth exploring.
Still, I think the main driver of flexibility in these cases isn’t going to be price—it’s going to be faster interconnection, and participation in emergency programs to mitigate blackouts.
Doug Lewin:
Yeah. And this brings up another issue I wanted to ask you about.
As we talk about new loads trying to come onto the grid—whether just normally or through some expedited process—should we start distinguishing between different types of large loads?
Because not all large loads are created equal. Within data centers, there’s a huge difference between those doing cloud, inference, or training work.
But also, outside of data centers, we’ve got big new loads from industrial electrification—fracking operations electrifying, steel mills, semiconductor manufacturing, etc.
Do we need to start classifying loads differently in policy, rate design, or demand response program eligibility?
Because if the only criterion is flexibility, well… crypto miners are extremely flexible. But if it’s crypto mining vs. a Samsung chip plant, most people would probably choose the chip plant.
Did this come up in your work? Or maybe this is an idea for future research?
Tyler Norris:
Yeah, absolutely. We were focused here on AI-specialized data centers, because this is such a big, fast-moving trend that everyone’s trying to get their arms around.
But you’re right—flexibility could apply more broadly. We often shy away from the idea of industrial facilities offering flexibility because no one wants to interrupt a manufacturing line. which also gets really expensive. But look, if these large factories have their own onsite power options or onsite batteries, I mean, it's the exact same principle. So let's not discount that some of them might have an interest in doing that. may not. I think the other thing is, yes, of course they also want to interconnect quickly, but it's just not the same like breakneck speed to market and also this global like, you know competition that has major national security implications in terms of the AI race. So that's a little different, I suppose. But ideally, should, for any large load or small load for that matter, that could offer flexibility that creates a value for the broader grid, we should be making sure that there are service contracts that acknowledge and enable that.
Let's not discount it. There is something about just the scale of the day. actually this is a good question. So how many of the non data center loads are like over a hundred megawatts and I don't have a good data sheet in front of me. My sense is that even some of the largest factories we're talking about, like not a lot of them go beyond, you know, the hundred to 200 megawatt range. So like in Duke Energy's case, they've rolled out all these new sort of guardrails to require like more upfront security and guarantees from new large loads, but the threshold they established was over 100 megawatts. And there's probably always going to be some degree of an arbitrary threshold there. But you can also think about in terms of the number of upgrades to the system that are being triggered by it. So maybe you do it, and there's a dollar threshold cut off. But I think it is probably useful to think about this when we are talking about truly quote unquote hyperscale facilities. And you're talking about billions of dollars potentially of system upgrades and new generation resources that may be required.
Doug Lewin: (45:55.714)
Yeah. You write in the paper about the potential for lower ratepayer costs from load flexibility. And I think about this a lot—affordability has to be front and center. It needs to ride sidecar with reliability. I don’t care which one is A or B—they’re tied together.
But I’ll be honest: I’m worried that these big, sophisticated customers are going to work the system to lower their costs, and everyone else is going to end up paying more. Sure, they have a high willingness to pay now, but as competition ramps up and margins tighten, they’ll have more pressure to control their costs.
So is it really feasible that residential and small commercial customers could actually benefit from this growth? I want to believe the answer is yes—tell me it’s yes.
Tyler Norris:
I think the answer is yes—but it depends on smart policy and regulatory oversight.
We’ve already seen, in just the last 6 to 12 months, multiple jurisdictions—at least in the Eastern Interconnect—starting to put meaningful guardrails in place. It starts with more upfront security posting, so that if a large customer backs out, ratepayers aren’t left holding the bag.
Doug Lewin: (47:38.606)
Yeah—that’s also part of Senate Bill 6 here in Texas.
Tyler Norris:
Exactly. And that’s sort of the basic first step. I imagine we’ll see more jurisdictions follow suit.
But there are more ambitious ideas, too. In Indiana, for example, after seeing massive forecasts for data center load, there were proposals to fully firewall the costs—like, assign 100% of any triggered system upgrades (whether for generation or transmission) to the new load, not existing ratepayers.
We used to have a paradigm of “meet any new load at any cost.” That just doesn’t hold anymore when growth is this large and fast—and when the willingness to pay from these customers is so high.
If you start sending the full price signal—in this current environment—it seems like they’re willing to pay it. But obviously they’ll try to reduce costs over time, so now’s the moment to get the policy structure right.
Doug Lewin:
Yeah. I think the piece you put your finger on in this paper—load flexibility—is the key to that affordability piece. Because if we can actually control those peaks, that helps with reliability, and it helps with prices, because that’s when prices are highest.
So I think the potential is there. If we get the policymaking construct and the pricing construct right—if—then I think large loads could bring in new resources, offer flexibility, and maybe even pay into a fund to help reduce residential peaks and create more headroom.
I can see that scenario. But I don’t know if that’s the most likely one to play out. That “if” is doing a lot of work.
Tyler Norris:
Yeah, agreed. This is why we need regulators to take this moment seriously and get as smart as possible, as fast as possible.
That was one of our goals with the paper—to provide some technical framing. But all of us need to be helping regulators and stakeholders level up their understanding. We’re all learning this in real time.
Doug Lewin: (49:28.268)
The learning curve is almost as steep as your load duration curve.
Before we end, Tyler—this has been fantastic—I just want to ask you one more thing.
You’ve done a lot of work on interconnection, and you recently testified to Congress on this. I’m going to read a quote from your testimony earlier this year:
“While load flexibility can help, interconnection delays remain a major obstacle to deploying new generation. The volume of projects stuck in interconnection queues has quadrupled. However, one market—the Electric Reliability Council of Texas (ERCOT)—has taken a different approach and is achieving significantly greater interconnection performance than other U.S. markets. Between 2021 and 2023, ERCOT interconnected at least 70% more generation capacity than any other organized market, despite serving a load that is half the size of PJM.”
I’m often critical of ERCOT, but I also want to give credit where it’s due—and this seems like a major success story. Can you talk about what ERCOT does differently, why it works, and what the rest of the country might learn from it?
Tyler Norris: (51:23.042)
Yeah, I don’t think you can overstate the significance of what ERCOT has accomplished over the past five years.
It didn’t become clear just how different ERCOT’s performance was until the other queues melted down—and that’s only become obvious in the last three to four years.
It’s rare that you have one market that is so fundamentally different in performance—and where that performance is clearly attributable to a different set of rules.
ERCOT is an energy-only market. It doesn’t study and assign costs to new generators based on deliverability studies, which are essentially used to determine whether they can qualify as designated capacity resources.
Doug Lewin:
So it’s a more administrative process elsewhere—someone decides what’s “deliverable” and assigns cost accordingly. But in ERCOT, that step isn’t necessary?
Tyler Norris:
Exactly. In all the other organized markets, interconnection and capacity eligibility are tightly linked. So when a new generator comes in, they usually request Network Resource Interconnection Service—the type that allows them to participate in the capacity market.
That triggers a very rigorous set of study criteria, and any upgrades that are identified get assigned 100% of the cost to the contributing generator. This creates a huge bottleneck.
Now, there are other factors behind the queue problems, but the most fundamental one is this rigid linkage between interconnection and capacity eligibility. It creates a binary outcome—you’re either fully capacity-eligible or not.
But we should be clear: this is a policy decision, and what it’s done is create a high barrier to entry for new generation. And in a constrained environment, that leads to price spikes—which is exactly what we’re seeing in PJM.
Now, we probably can’t just copy-paste ERCOT’s model into the rest of the country, but we should be figuring out how to adapt its efficiencies. That’s going to be a major task going forward.
Doug Lewin:
Yeah. There’s no pure energy-only or capacity market. They all have pieces of both. It’s really a question of where you set the dial.
But I really appreciate you highlighting that. I think it’s underappreciated. There’s definitely a negativity bias in this space—we talk more about what’s going wrong than what’s going right.
And this is something ERCOT is doing right—Connect and Manage has worked. It’s put more generation on the grid than any other market. So thank you for pointing that out in your congressional testimony.
All right, I’ve got one final question, and then we’ll wrap.
You mentioned it earlier, but I want to circle back to it: DeepSeek.
How does DeepSeek change your outlook on load growth, if at all? Just for folks listening: DeepSeek is a Chinese AI model that surprised a lot of people. And part of the conversation was: are GPUs getting so efficient that maybe we won’t see as much load growth?
Has that shifted your thinking?
Tyler Norris:
Yeah, so some quick background:
Before DeepSeek, the consensus was that the U.S. was about three to five years ahead of China in developing large foundational models.
So when DeepSeek launched—and showed comparable performance to our most advanced public large language models—it was a shock. Especially given all the efforts to restrict China’s access to cutting-edge GPUs.
Now, it turns out DeepSeek wasn’t a new foundational model. They built on open-source U.S. models and used clever engineering to optimize efficiency.
We don’t know what they spent—it’s not publicly available—so it’s hard to judge. But it does highlight that efficiency gains are possible through optimization.
That said, the initial reaction created a bit of a lull in projections. But within a few weeks, things were back in full swing. The forecasts for data center load growth are still massive. Hyperscalers are continuing their investment. OpenAI is full steam ahead on Stargate.
So for now, it looks like we’re still on the high-growth trajectory.
But what DeepSeek did do was inject more uncertainty, especially around inference loads. Some of this work can now be done locally, which is what Apple has been pushing.
So it’s not just a question of “how much energy?” but where and when it’s being consumed.
Doug Lewin:
Yeah, so interesting. There’s so much more to explore.
Tyler, this has been great. Your report is a huge contribution to this conversation—Rethinking Load Growth. We’ll include a link in the show notes.
Before we close, tell listeners where they can find you online—and is there anything I didn’t ask that you wish I had?
Tyler Norris:
You can find me on LinkedIn and follow some of the work through Duke University’s Nicholas Institute. And no, this was terrific really enjoyed the conversation, Doug.
Doug Lewin (59:37.748):
Awesome. Tyler, thanks so much. Really appreciate it.
Thank you for listening to the Energy Capital Podcast. I hope you enjoyed the episode. If you did, please like, rate, and review wherever you listen to your podcasts.
Until next time, have a great day.
Share this post