A (2:25)

Thanks, Ed. Such an honor to be on this show. I've been a Longtime fan of the Energy Gang.

D (2:29)

Thank you. That's very kind, Very kind. So look, Faroon, you're. Well, what is the best way to describe you? I mean, you are the founder and CEO of Emerald AI and we'll come on to what that company does in a moment. But first, you've had a very interesting and varied career in energy already. Something we always like to do on this show when we get new people on is talk to them a bit about their careers in energy, how they first got interested in the subject, and how they got to the roles they hold now. So could you maybe do that a bit for us? What's your story?

A (3:00)

Thanks, Ed. You know, I like to think of myself kind of as Forrest Gump. I have had a front row seat to all of the important and interesting things that have happened in the energy and climate revolution and just feel lucky to be along for the ride. You know, I got my start doing my PhD in solar physics. And in our lab in Oxford was invented the Perovskite solar cell, a revolutionary technology that I still think will win the Nobel Prize. Ed, as you know, I'd go on to write a book called Taming the Sun on the future of the energy industry. In fact, you wrote a very helpful review. I was so grateful for it in the Financial Times. And my goal throughout my career has been to advance next generation energy technologies, whether as the lead US Clean energy diplomat in the administration to an executive at two publicly traded companies, Orsted and Renew Power, India's biggest clean energy company. But now I've kind of left it all to move into this AI space, because I think AI is maybe the most exciting thing I have ever seen happen to this energy sector. And I'm excited to play a role in it.

D (4:05)

Absolutely. And so tell us a bit about Emerald AI then. What is your company? What's your pitch?

A (4:09)

Well, it all starts with, again, I think AI is an exciting addition to our energy system. And in saying this, I'm trying to upend the typical narrative because a lot of people look at AI and they say, my goodness. And this is a scary addition to our energy systems. It's straining the grid, it's raising electricity prices. We can't support this much AI And I look at AI and I say, wait a minute. Actually, I think AI far from undermining the grid, could actually save it. It's an extraordinarily potent and powerful potential ally if we can just enable AI to help the grid out during periods of stress. That's what Emerald AI does. I founded this company to make artificial intelligence data centers, or what Nvidia calls AI factories, more grid aware and power flexible so that when the grid faces a moment of peak load or peak stress, an AI factory can ride to the rescue and provide the flexibility the grid needs so that we can have reliable and affordable and clean grids.

D (5:09)

Got it. And that's really what we're going to be talking about on this show, how that might work and what some of the opportunities and the challenges might be. Just before we get into that, though, it's very timely because I know Amy and Varun, you've both come just hot foot from this meeting at the Council on Foreign relations in Washington, D.C. where the energy Secretary, Chris Wright, has been speaking. And he's actually been talking about exactly this subject. I think he said something like, to me, this has been the subject I think about most. I think this is what I'm most interested in obsessed by at the moment is thinking about how to increase electricity supply and energy supply for AI specifically and for economic development in general. But, Amy, you were listening to him. What did he say exactly?

B (5:52)

Well, so when he was asked about what does he think is the state of the U.S. electricity grid in the first place, he said it was, quote, unquote, very concerning. And when somebody asked him about the pressure of AI, he said, we have to move fast. It's a Manhattan Project, number two.

D (6:12)

Right. And so this is the idea we've talked about quite a bit on this show in the past, that the United States is in an AI race with China today, just as it was in a race to develop nuclear weapons with Nazi Germany during World War II, as it was in the space race with the Soviet Union in the 1950s and 60s. And so it's kind of, you know, the stakes could not be higher. This is kind of the fate of the world is hanging on this. And therefore kind of everything has to be done to maximize electricity supplies, to be able to run those data centers, to deliver those advances in AI, right?

B (6:46)

That is correct. And he, he kind of went through the history of his own journey as a scientist like Varun and his studies as a scientist, and really laid out this as a national security challenge. Varun, I don't know if you want to add anything to kind of what he, what his perspective was, but, but definitely that. And there have been other speakers recently at the Council Foreign Relations who've made similar points, harkening back to the fact that our national security apparatus was set up after World War II and then focused on the Cold war, and again, that we need to innovate, we need to focus on cyber, we need to focus on AI. And we're having to basically think forwardly about reforming institutions and reforming practices to get us where we need to be, not only to be economically competitive, but also to really ensure our national security going forward.

A (7:43)

That's right, Amy. And let me just say also, I'm a senior Fellow at the Council on Foreign Relations, and this was a marquee flagship event where we hosted. Secretary Wright, in addition to everything you said, I heard the Secretary say that the single largest impediment to getting AI built and meeting those national security imperatives and competitiveness, as you said, editing, is getting electricity generation capacity online. He talked a lot about nuclear. He talked about other sources of energy that we're trying to bring online. And I think to that I'd say we absolutely need more energy. We need abundant energy. There's a range of priorities this administration wants, such as re industrializing America with manufacturing, advanced semiconductor fabrication facilities, et cetera. The caveat I'd add to what the Secretary said is if we can enable AI to have this flexibility, as AI becomes one of the largest and soon the largest consumer of American electricity, then we may be able to sidestep some of the infrastructure increases that it currently demands. And that could be good for ratepayers who are seeing sharply increasing bills. It could be good for grid reliability, and it certainly could be good for the speed of getting AI data centers built much more quickly.

D (9:00)

Yeah, so let's talk about that issue of flexibility then. As you all know, if you've been listening to this show for a while, it's an issue we've talked about quite a lot this year. If you heard our episode with Tyler Norris of Duke University earlier in the year, he talked about how adding new large loads for data centers will be a whole lot easier if those loads.

E (9:20)

Can be flexible, if you're able to avoid overloads on those system elements during these periods of grid stress, then you may be able to get online and operate again if you're willing to provide the flexibility. And the other thing to say about it.

A (9:35)

Right.

E (9:36)

Is that could be entirely an interim arrangement. So what we may see, and I think this is most likely a sort of tip of the spear, is you'll see some of the data centers implement flexibility to get online and interconnected more quickly until the upgrades are done, or what we'd call the firm upgrades are done to. To the transmission, distribution and for generation capacity, such that they can be treated as fully firm loads. You know, if the data center's already invested in, you know, whatever flexibility capability they're tapping, maybe it's worth it to them to avoid those costs. And so I think that that's how the sort of market will evolve from there.

D (10:14)

In other words, it really helps if the grid can say to customers that are coming on with new data centers, hey, if we're under strain, if we can see strain coming, we need to be able to dial you down, send you less power for a while. And if that can be done, say Tyler Norris has these calculations, it will be possible to add a lot more new loads to the grid a lot more quickly. But in my experience, when you talk to datacenter operators very often, they talk about always being available. You'll talk about Five9's reliability, in other words, being available 99.999% of the time. And there does seem to be that real tension here between the grid needing flexibility and data center operators not wanting to provide it. And we seem to hear a lot of talk about flexibility, not so much actually being implemented. So Varun, what do you make of that and how do you think it actually can be turned from, say, a kind of hypothetical thing that people would like to have and that would be great if they could have it, into something that actually is a feature of these large loads on the grid?

A (11:23)

Absolutely. And I can tell you I was talking to Tyler earlier today. Tyler is an advisor to Emerald AI and his paper made waves because it said, as you said, Ed, 100 gigawatts of available stranded capacity on today's electric power systems could be unlocked by just modestly flexible data centers that might reduce their consumption for up to 200 hours during the year, for just two hours at a time, and reducing their consumption by 25% at a time. But your point, Ed, is absolutely correct. The way I think about this is the immovable object meeting the unstoppable force. The immovable object is the incumbent electric power industry, a more than century old industry that has as its very sincere service obligation, the need to always provide power availability and firm service. And then you have the unstoppable force, which is the AI and technology industries where the industry convention for data centers, as you say, Ed, is we're going to provide our computing customers almost perfect 100% uptime. Now, what flexibility demands on both of these sides is just a slight, ever so slight change to their traditional operating approaches for the electric power grid. To say, look, I might not need to make a Data center wait 10 years while I strengthen the transmission grid and build more generation, but rather I can connect them in six months, but for 100 hours a year. Let's say I might have to request that they reduce their consumption by 25%. And from a data center's point of view, if you're a hyperscaler offering cloud services, for example, or an emerging AI neo cloud, you might need to say to your customers, look, I have the ability to access vast amounts of new power almost immediately and be paid for my trouble, but I will need to cap the amount of power that's being used in my facility every so often less than 1% of the year. These minor operating tweaks make all the difference, and we have to overcome this inertia. And I think the prize is just massive. The prize is $4 trillion of new investments in AI data centers. If we can get this right on both sides.

D (13:28)

The energy industry is changing faster than ever. So to make sure we're helping you stay on top of every development, we'll be adding some extra shows to the schedule over the coming weeks. First up, we'll be at New York Climate Week, speaking with leaders in clean energy, finance and policy, and partnering with our good friends at NYU for a very special panel discussion with some big names from companies including Nvidia and Amazon. That episode will be out on Friday, September 27th. So mark your diaries now. And we'll also have lots of extra shows throughout the week of September 22, so make sure you're following the show wherever you get your podcasts. So, Shahnu, what's your take on this then? When you think about flexibility, how widespread do you think it really is becoming already? And how widespread could it become?

C (14:15)

Yeah, absolutely. I think it might be helpful just to set the stage a little bit for why the tech companies may be at least perceived to be inflexible or why they're a little bit more cautious about that. And I think it just ultimately goes down to the quantum of capital being invested into this space. So I think everyone's seen those big charts about hyperscaler Capex, but just to put some numbers to it, right? I mean, we'll spend on the order of close to half a trillion dollars this year just from the big hyperscalers. And half of that is chips and it and other half of that is like physical land and things of that nature. And that grew at 50% this year. It'll slow down in the future. Right, but again, we're talking about hundreds of billions of dollars, approaching trillion dollars, and so when you think about the useful life of some of these IT equipment, it's typically like two to four years sometimes for the chips. So you can imagine for a tech company that's spending, you know, let's call it 30 billion to $40 billion for a single gigawatt training center, being told that you can't run that for every maximum hour is a really negative NPV type of calculation. And that's why they're so hesitant in doing so. But I think one of the things that, you know, to Varun's point is flexibility is coming into the conversation naturally just from the evolution of the technology cycle. I mean, one of the things I was playing around this morning with, I was writing about was just kind of the evolution of the increasing computational demand that we've witnessed. So originally it was like, let's maximize the individual GPU performance and then it became unified rack performance and now it's data center wide performance. And now increasingly, if you look at Nvidia and Broadcom's calls from the last week or so, we're moving into a multi data center distributed performance type of parameters where you're trying to scale across data centers. And that's happening just naturally because of these constraints around power, around real estate, for around building these really, really high density type applications. And so I think to Varun's point, we're talking about orchestration, we're talking about coordination of workflows. That's going to happen naturally just given the constraints that exist in building massive amounts of infrastructure. And that's going to be true whether it's training load, whether it's an inference load, etc. So I think a part of this is just kind of, you know, inevitable. And I think people that are kind of framing this as like this is a yes, no, and we're never going to go to flexible because it doesn't make economic sense. Just kind of misses the point a little bit, like it's going to naturally happen just given the constraints that are there. But ultimately, you know, I think what the hyperscales are solving for is just they want to ensure that they can get, you know, positive ROI on any type of investment they're going to be making. So, you know, if flexibility is a part of the equation, what did they get in return for that? And I know we're going to get into the PGM proposal part.

B (16:36)

Let's just go a little bit further down that rabbit hole. Shanu, because you raised some great points, really building on Varun's insights. So if I'm doing that, but I'm not willing to, like, have multiple workloads contributing to something at different times, or I'm not wanting to have a cluster that's on different substation loads. I mean, the risk is that I'm gonna get shut down accidentally. Either the weather's gonna shut me down, some kind of congestion's gonna shut me down. You know, everybody's gonna put their hair dryer on at the same time. So I think a precedent was set, ironically, in ercot, because even though ERCOT has all this talk about firm dispatchable power, the truth is they really actually put in a lot of batteries and they are very concerned, as they should be, about residential users. And so they were the first state to say, we're passing a law that gives us the right to turn a data center off completely if it means making sure there's no blackout for our residential citizens. And so, you know, do we think that's going to happen in more states? Varun, do you think that changes the conversation you're having with the hyperscalers? Because if that starts to be the regulatory framework across the country, they're going to be forced to think about flexibility.

A (17:59)

I also want to come back, Shanu, to some of the exciting things you mentioned about multi data center fabrics, but let's come back to that because this is a critical point, Amy, as you say, grid operators have the service obligation, and the last thing you want to do is have to shut off a residential community or shed rolling blackouts as a result of not being able to meet peak demand. And it's especially bad if the data center has caused this problem and yet they still get served. And so therefore, what can you do? Amy, you made the point that in Texas there's now a law on the books that in theory could enable the power system operator ERCOT to simply shut off data centers. Again. This is why I think hyperscalers, data center operators, tech companies, have an incentive and a reason to proactively come up with more graceful solutions than simply getting shut off. You know, I'm really excited to share. This just became public that Nvidia and Emerald AI are collaborating to enable a reference architecture for AI factories or AI data centers to have some of these capabilities like flexibility. And here's why that's important. If a data center has this capability, it can provide that flexibility, that demand reduction during a peak moment, for example, in a graceful way that protects the AI customers but also meets the grid's needs. And you don't actually have to wait until the warning lights start flashing red. One of the best conversations I've had. Early on, as I did market research to create Emerald AI, I spoke to the CEO of ercot, really a visionary thinker, Pablo Vegas, who shared that his goal would be to never get to the emergency condition to begin with. You never want the warning lights to flash red. Maybe they flash yellow. And that's the point at which you deploy the available resources you have. You know, Amy, as you mentioned today, they're batteries. But data centers could be that flexible resource that as soon as you're in a moment of even remote danger, the data centers are able to help the grid return to a normal operating condition. Now back to Shanu's point. There's trillions of dollars going into these data centers, and they are doing some of the most economically valuable work in our economy. I don't care if you think that generating cat images or images that look like Studio Ghibli is economically valuable or not. The numbers don't lie. The GDP benefit of US Data center and AI are staggering. And so therefore, those are the economic activities we do not want to curtail. At the same time, if we can do so gracefully in a way that the AI compute users still remain very satisfied with their quality of service, you have this rare win, win, win, win, win, right? I can go through the five wins, but I think you know what they are. The AI compute user is happy, the grid is happy. You have reliable, safe, clean and affordable electric power systems. I think I got to five wins, and that's rare in our field.

D (20:53)

So, Shani, what do you make of this? Does that sound too good to be true?

C (20:57)

I mean, I think there's a lot of steps between now and there, but I mean, ultimately I think the long term vision I'd agree with, right? And then point in terms of like, the different priorities, right? Like if I'm a utility, on the hottest possible day, everyone's operating at their peak, low capacity, AC is on, et cetera. I need to plan for that event. But the issue with the data center stuff, right, is that based on the operational load of that type of facility, whether it's a colo. Whether it's an infrastructure data center or a training data center, there's different utilizations that spike up and down. And as anyone that follows this stuff really closely knows, at a chip level, and then extrapolating to a data center level, these things are really transient. Meaning that the workloads spike. Like oftentimes the chips will be like, idle. Like operate at 30 to 40% of their max thermal capacity and then spike up to 100 to 110%. That's utilities worst nightmare. Right. But they're solving for assuming 100% all the time. And I think in reality what we see is when these things are done in practice, like, the percentage utilization is a lot lower. And that's why I think it's a lot. I've met more in terms of, like, it's up to the hyperscalers to figure this out eventually, because you're already seeing utilities start to cap some of the issues associated with the transients. There was a really good paper by a bunch of Microsoft researchers, power Stabilization for AI Training data Centers, published in August this year, and they talked about the fact that when you scale up with this many GPUs and you know they're moving up and down, it kind of creates bad frequencies for grid managers. And so they're already indicating, hey, we're going to cap your limits on how much you can swing up and down, because that makes it a lot harder for us to manage the frequencies on the back end. And so I think part of the conversation around flexibility is like, there's a different mismatch between the utilities solving for the worst possible scenario, but in reality the percentage might be a little bit lower. I mean, just today I was reading a report on ercot which we just were talking about. They did a study of data center lows from 2022 to 2024, and they actually only used 50% of the megawatts that initially requested. Right. And so I think that that introduces this mismatch and where there's an opportunity for flexibility here, where it's like if you requested a gigawatt and end up using 500 megawatts, that opens up quite a bit a lot. Right. For the flexibility side and the utility planner side, in terms of what's actually being managed according, or what's being planned for and then managed to.

D (22:58)

Yeah, absolutely. And I've seen other numbers like that circulating and I think we've got some wood McKenzie as well. As you say, saying, although the data center operators talk about Five9's availability, they're absolutely not used. They're not using 99.99% of the full capacity at all times or the full capacity at 99.999% of the time. So there does seem to be some headroom there. Again, just from a grid operator's perspective, that doesn't seem to be at all helpful though, because if you have no visibility into exactly what the variability of the demand is going to be. If you can't predict when that spare capacity is going to be available, and most importantly, it's not controlled by you, it's not kind of dispatchable effectively, it doesn't really help. So to your point, when you say grid operators always prepare for the worst eventuality, you can see why they do that, right? That's totally reasonable from their perspective. Because if you say, oh, well, this gigawatt data center is only using 500 megawatts, but then at the time of maximum stress on the grid, it happens to jump up to 1 gigawatt, then disaster ensues. So you have to be on the cautious side.

B (24:10)

Everybody help me here. I mean, I'm the least knowledgeable of the all of us having this conversation. I'm like the lay person. So let's just say, Shawn, a very interesting study and very interesting statistic from Texas. So how do we get to a moment where we had this massive blackout and somehow we didn't even have the data centers operating at max capacity? Is it just that so much generation capacity got wiped out? Could that same thing happen in California in a heat wave? I mean, what kind of percentages are we talking about where we can't control this fluctuation at times of grid stress from weather and then taking it the next step? Shawn, would just frighten me with the whole frequency thing thinking about Spain. So help me out here as like a layperson. How difficult is this?

A (25:07)

Sure, I'm happy to jump in here. Let me just break something down for our listeners, which is there are varying timescales at which AI can either create problems or provide solutions. So as Shanu was mentioning, there are very short timescales, millisecond transients, seconds, frequency issues, and there are other timescales, minutes and hours when the grid is facing a peak load moment, and we need to shave that peak across all of these timescales, different solutions will be needed and there isn't one single silver bullet for any of them. So, for example, when an AI training run starts up and then it causes a massive tens of megawatt increase in energy consumption. And then all of the GPUs do something called a synchronized checkpoint. It's kind of like saving your work on a Word document. Then the power falls by another several tens of megawatts. These are being addressed, as Shanu mentioned, through certain hardware fixes where Nvidia's chips help to smooth that curve. There are software fixes and there's energy storage, like batteries that can help you on the longer timescales. You also have a range of interventions. So I just first want to make the point that there's many different timescales. Now, Amy, the question that you were talking about is typically most visible at the several hours time scale when you have a mismatch between how much generation is available. Let's say a winter storm has knocked out some natural gas generation or a fire has knocked out power lines and your demand happens to be high. Let's say it's a hot day and there's a lot of air conditioning demand. Data centers may not be a silver bullet. You may not be able to get enough load reduction from the data centers, but if they are flexible, you may be able to get some load reduction and help a little bit. And then there will be other interventions that you'll have to take. Typically, as you know, Amy, it's a great idea for your grids to be larger and more interconnected and to have strengthened transmission such that if you have a larger area, it's harder for any one particular section to just to go down. But if you have an isolated grid, then that grid can in fact go down because of a supply demand imbalance more easily.

B (27:17)

So varun. I love that point. So explain to us again in this context the idea of having hyperscalers have a cluster of data centers that work in coordination and therefore we're less apt to have that problem. Because if we had, you know, interactions between different grids and we could move resources around better.

A (27:38)

Exactly. You know, we actually did a real world test that showcased what the potential might be to avoid blackouts. So as I mentioned, we went to Phoenix, Arizona, we had this Oracle data center, and we reduced consumption of the AI cluster by 25% for three hours at a moment of peak demand. We then went and re simulated a case where in California five years ago, in August 2020, there were rolling blackouts because a natural gas plant tripped offline. And we simulated what would the demand response curve have to have looked like. And it looks kind of like a stair step to avoid those rolling blackouts. And we showcased that the emerald conductor can create that stair step in real time, first providing a little bit of a demand reduction. And then when the grid says, oh my goodness, I need more support providing even more of a demand reduction while still protecting the performance of the AI workloads. If we were to deploy emerald conductor at commercial scale across a large network of data centers, it's even more powerful, Amy, because then if you have a problem in Arizona, it's the middle of the summer, your peak load is approaching what your generation can provide. Then you're able to move some workloads that are very latency or time sensitive away from Arizona, maybe to the Pacific Northwest, let's say, or over to Illinois where there is ample wind energy at that moment, let's say. And that's going to enable you to keep the AI workloads running in different parts of the country and those that could be curtailed or paused in the Phoenix data center, we just reduce the power consumption of those, let's say, fine tuning operations. Taken together the whole network of data centers completes the work that needs to get done for its AI users, which, as I mentioned, is the most economically valuable work there is. I'm not trying to put a moral judgment on it, but economically, GDP wise, it is the most economically valuable work you can do per unit energy. And at the same time, we have saved the Arizona region from any kind of rolling blackout. That's the vision I want to achieve.

D (29:30)

Right. But again, to your point, Varun, about what happens if there's a blackout and people lose power and communities go dark at the same time that data centers are able to keep running. Talk about this being the most economically valuable work that could be done in it's not really going to cut it. Right. People will be angry, regulators will lose their jobs, politicians will lose their jobs. These are the things again, when you think about why regulators and grid operators take the decisions they take, that at the end of the day is the ultimate sanction, that is their hard constraint that they've got. They have to keep the lights on and all those other considerations are going to be secondary.

B (30:11)

And that's what's happening in pjm. Why is PJM in the news? PJM is in the news because they had these capacity market auctions and the prices were stunning. And that made it a political thing not only for everybody involved in maintaining the system in PJM, but also for all the politicians in the different states that are part of that regulatory area.

D (30:37)

Well, indeed. And politicians say, oh well, the capacity market prices are too high, we shouldn't be having to pay for all this. You still need the capacity though, and if you don't incentivize it, probably it's not going to come. And so that's going to be creating problems for you down the road. So.

A (30:51)

But the point that Amy makes is dead on, which is the price of electricity, the household bill annually for a household in Columbus, Ohio on average went up $240 directly attributable to data centers. That is not a good look and that I think everybody has an incentive to avoid that happening. And that's why we need to be proactive in finding better solutions. Now, I do want to hit a positive part of the story, which is the tremendous pace of technological change. Earlier, Shanu mentioned that you heard from earnings calls from Broadcom, you heard from Nvidia a couple weeks ago, that the incredible technology progress, not just in the chips, but in the networking, is enabling multiple data centers to act in concert. In fact, Nvidia has a new technology called Spectrum X that enables such a high bandwidth set of network connections between data centers that it's as if data centers that are spatially separated can act as if they're all part of the same data center complex. And that's a very powerful tool. You know, I feel, again, bringing back this Forrest Gump analogy, I'm here watching all of these tremendous advancements in the tech industry. AI is forging ahead. And every time I hear someone like our colleagues at Oracle, they're our partners, talk about the tremendous fiber optic connections they'll build in between data centers, I get excited for a reason that has nothing to do with why they're excited. They're excited because they're going to serve their AI customers an amazing product. I get excited because I can piggyback on this incredible advance for electric grid purposes. If you have a very well networked set of data centers, it becomes even easier to shift workloads in between data centers, to pause workloads in one location and make sure they get picked up somewhere else to coordinate operations and even to do things like distributed training where you might have many autonomous workers operating in multiple different data centers. Google's DeepMind published a protocol, Deloco, that does this. And that's a very exciting opportunity because you might. Therefore, instead of having to take down a whole training run, you can just pause simple, decentralized autonomous workers one at a time in places where grids are particularly stressed without compromising the overall progress of the training run. So I think, again, just to bring this back to a 10,000 foot level, the pace of AI is moving and everybody who is advancing AI is focused on AI and its economic potential. I want to be focused on leveraging these advances for helping the grid so that we don't have the situation, Ed, that you mentioned, where the data centers keep getting power but the neighborhoods get blacked out. That's a real bad look. Or Columbus, Ohio has bills go up by $240. We've got to avoid those.

B (33:44)

So, varun, just to restate what you're saying with this autonomous worker so people understand we have these tasks and they're being done instead of being done all in one central data center. And now I can't let that data center go down. It's as if I had human beings, electronic human beings, in different locations. And since they're all working separately, and then eventually I'm going to bring all their work product together into one final product. I could tell one of them to go take lunch at a different time than a different one, but we would still meet our target, which is to have the whole job done, to upload in a centralized place at a particular time.

A (34:25)

You have nailed it, Amy, and I'm so glad that you gave that much more understandable analogy. Look, this is still a research direction. Today, a large training run is done in the same facility with synchronized GPUs. But you can imagine, and there's great published work on this, and companies like Prime Intellect that are advancing this paradigm. You can imagine what you just said, Amy. The group project gets done by a bunch of different students, and each of them gets to take lunch whenever they want, but they share their notes afterwards, and they make sure that the work gets done on a reasonable time scale. And it's really robust and resilient, both to other things that can go wrong, like GPUs failing or outages, but also to Emerald's approach to pausing or slowing workloads or when the grid is stressed to provide AI data center flexibility. So think of it as the group project paradigm. We're not there yet, but when we get there, it's a good example of an AI innovation that's going to make it even easier for AI to help.

B (35:22)

The grid, and not only for AI to help the grid. I mean, honestly, all the companies should be working on it because it's an important national security imperative that you have a cyber division that way. And you can't just have something get disrupted because someone can target one particular data center because they have intelligence that data centers doing something critical.

A (35:45)

Nailed it. That's right. Shawna, you were going to say something.

C (35:47)

I think I was just trying to connect the two points earlier where I think we were talking about these things at contrast, where, you know, the utility operator has to decide between keeping homes lit and heated versus the data center running. And I think that's what the tech companies see in the. In the periphery, right? They're looking down and seeing that, hey, like these things could be construed. As the and or decision. And you know, like that is increasingly going to become a political hot button issue, especially as shitty inflation accelerates. And so they're already working on these solutions as just a result of their own redundancy and their own uptime initiatives. Like they want to be decentralized to a point where they can be in control of their operations. Otherwise they will be forced into these operations where the utilities can say, I'll shut you off. There's an Epoch paper that outlined the power density of the compute being demanded. One of the things they talked about was distributed training and synchronization amongst data center facilities. And right now there's already a host of labs testing this out at 15 to 50 miles. Broadcom talked about 100 kilometers at their last call. So you're already seeing this stuff happen in the background. It's just not at the forefront today. And that's why I think when Varun pitches this grand future of kind of the data center flexibility and these different options, whether it be hardware or software or firmware, you know, that's where you can see that naturally we're already headed that way in general, just given the fact that a lot of these companies are already thinking about what the problems are going to come up in the next one, three, five year timelines.

B (37:04)

So Shawn, I have a follow up question on that. You know, you follow the investment community. You cited these giant numbers. My whole community here in New York, the private equity community, they are just raising money for infrastructure funds right and left. Everybody wants to fund a data center, everybody wants to fund SMRs. Could we wind up in a situation where it becomes analogous to when everybody wanted to fund shale back in the early 2010s and then suddenly we didn't need all that energy? Or you think that no, this is a real thing. And even if some of the technologies that Varun is talking about start to become mainstay, there's just still going to be a giant need.

A (37:48)

Well, Amy, it's a great point. And if you go back to the paper that Shanu mentioned, which I thought was terrific, a collaboration between EPRI and Epoch AI, the graph that stuck out to me was not the one that said that the power demand is growing for AI at a rate of more than 2x doubling every year, but rather that the rate of computations is growing at a rate of more than 4x per year. That's meaningful to me because it's important to think about what might distinguish this AI explosion and demand and compute and power needs from Previous explosions in the increasing compute needs, like the Internet revolution, the dot com revolution. In those revolutions, you saw power demand largely flatlined because even as the demand for computations went up, the energy efficiency of those operations improved as well. And you flatlined your energy consumption at about 4% of American electricity use today because of that difference between 4x computations every year and the increasing energy efficiency of data centers, which is happening much slower today than it used to happen. The Net result is a 2x or more increase in energy every year. That means we're going to go from today 4% of American energy to 12% by 2030 to maybe 25% sometime mid next decade, where AI becomes the biggest consumer of electricity. So to your question, will there be a slowdown in demand for AI? Quite possibly. But as we look at today's scaling laws, at the diminishing marginal returns of training models, at the increasing use of AI in our economy, it's hard to see why you would have less demand for computation and less demand for energy than even some of the conservative projections. Sorry then. That even some of the aggressive projections we're seeing for energy demand today.

C (39:46)

Yeah, just to build off that, I mean, I think the paper highlighted, I think 30 to 50% efficiency every year, which is a large, large number. And typically that would be a lot to dent a curve. But in reality, if your compute is growing 4x a year, it's just a drop in the bucket right now at this point in the cycle. So I think what a lot of the private infra players and everyone else that's raising money are doing right now is they see that directional curve and they want to get ahead of it, so they're investing in infrastructure. I think it ultimately comes down to like one, do we continue to see those trends play out over the next three, five, ten years? I mean, I think it's too early to tell right now. They're making the bet that they can. And then two, I think the other point that the power players are thinking about, or at least I hope they're thinking about, is the eventual repurposing of the technology. Like, yes, like data centers are driving low today, but as we all know, right, Space heating, cooling, electrification, electrification of transport, etc. Will actually be even larger. So, I mean, if you're owning energy assets that can sell into deregulated markets, and maybe you feel a little bit better about funding that infrastructure. But Amy, to your point, I mean, we're seeing, you know, massive, massive partnerships between private capital providers and energy developers. I mean, there's like 10 I can name off the top of my head. So it does feel like there's a flurry activity. And, you know, I'm not going to sit here and call a bubble. I'm not in the business of that or nor would I be good at it. But, you know, it does feel that, like, at least we're at a relative level of exuberance that, you know, does mirror at least prior cycles in terms of kind of just like crazy behavior out there and, and the insatiable thirst, let's call it, to finance these deals.

B (41:12)

So I think Varun makes this point, so it's really an interesting question, like, where are we in the exuberance field? Because, Varun, to your point, if you listen to people from the tech side, you know, I'm going to have these robots in my home and they're going to be able to do all these things for me. So I can just do this podcasting all day long and I'm not going to have to touch a dish and. And then my car is going to be automated and Uber's vehicles are going to be automated and everything's going to be automated, and that's going to all mean inference, inference, inference. And so it does seem like it could be pretty exponential. But on the flip side, you know, I of course look at this in my work and Shano made allusions to it, but I guess this paper took that all into account. If I am, everybody is using E commerce and I do have all these AVs, and I do have all this AI being applied to airlines and to marine bunker fuel and all this stuff, then I'm going to have a huge gain in energy efficiency. Maybe I would increase electricity needs at some level, but I'm going to decrease it in other places. I've been recently working with some professors at NYU on the building sector in New York City, which is very controversial because they've told everybody to electrify. And some of the PE firms are doing some pretty innovative buildings, some of which they're saying, oh, we're not going to really need much energy from the grid because the whole built. It's called a circular solution where the whole building generates its own electricity internally. And they don't really. The grid is kind of a marginal amount of where the electricity is coming from.

D (42:59)

Of course, you know, Amy, what's actually going to happen is the AI is going to do the podcasts and we're still going to have to do the dishes. I mean, that's the.

B (43:05)

Yeah, I'm afraid we are going to come to that.

D (43:08)

It's a really interesting point. I wanted to bring us back from that kind of future prospect though to just because we're heading towards the end of our time. I just wanted to get back to this kind of short term question or shorter term question because I mean as you say Amy, a lot of those issues, great deal of uncertainty on the longer term about how much the demand is actually going to be. As you were saying varun earlier, a lot of these kind of ideas about flexibility of the compute and where that can happen and so on are going to be implementable in the longer term. But in the short term there's an enormous amount going on right now. As you were saying Shanoop, you know, there's whatever it is, half a trillion dollars of capex being spent this year in 2025, there's enormous amount of steel going into the ground, you know, racks of chips being put in all over the country. And so everybody, regulators, policymakers, utilities, the industry in general is working out how to deal with this kind of in real time immediately. And the example of that I just wanted to get your thoughts on was really interesting. This month there's been a lot of pushback to the proposals from pjm. Pjm, the Big Power market regional transmission operator covering an area from Illinois to Tennessee to New Jersey to North Carolina. And they've basically come up with this idea and let me just check the terminology. There's a new acronym they've created, right? The ncbl, which is the non capacity backed load. And what they're basically saying as far as I can tell, is that if you've got one of these NCBLs then they want to reserve the capacity to shut you off when they think they need to because of strain on the grid. In other words, if you don't have backup generation, not necessarily kind of co located but somewhere available on the grid that you can put into the grid to cover your power demand, then they're going to insist that you're flexible. This plan seems to have gone down very badly. I was looking at the newsletter Heat Map that they said the country's biggest grid has a plan to manage data centers, power use. Everyone hates it. And it certainly does seem to be the case that if you look at sort of governors of some of the key states and some of the key companies and industry groups and tech industry groups and utility industry groups all seem to be kind of pushing back a bit. I couldn't quite see why they hated it so much because it felt to me, like at the very least, PJM is saying, look, there's a real issue here, we need to address it, we need to address it quickly. And perhaps they're being a bit aggressive in the way they're talking to data centers about it. But I mean we were Talking earlier about SP6, the Texas bill that basically has said to data centers large loads in Texas, if we have to shut you down, we're going to do that. And if you don't like that, that's tough. And essentially it seems to be PGM kind of thinking about a very similar model. I don't know who's been Shanu, have you been looking at this and what do you make of what they've been proposing and does it make sense to you or do you see why everyone's unhappy about it?

C (46:15)

Yeah, I mean if you look through the commentary and the original proposal, I mean one of the things I think is markedly absent from it was that they didn't really at least specify the benefits you'd get from participating. I mean, you wouldn't pay the capacity functions or premiums, but you don't get any speed for interconnect. And I think that's the whole trade off right here. At the end of the day it goes back to that economic incentive where it's, you know, do I get any type of benefit for voluntarily shutting my load? And in this case it wasn't very clear or spelled out in the original proposal. And so like what you see in the comments here is like a willingness for demand response if there's some associated benefit. Right. Because in this case then like the economic benefit of time to power is a lot greater than, you know, this voluntary thing that you can get around by, you know, either figuring out flexibility from a software or hardware or storage perspective. And I think that's what was missing. And largely when you look at the commentary from some of the big players, the data center coalitions and etc. Like Tyler Norris has covered this ad nauseum. They seem constructive. It's just a matter of like, hey, like what are we getting in return for this? Otherwise you're just discriminating against our larger load. Varun, I'm sure you have some opinions on that.

A (47:20)

Totally. And it's going to be a push and pull for a little while as we try and find the right compromise. But I'm just glad that folks are thinking proactively about this. Look, I gave a speech recently at the EPRI summer seminar EPRI has been a wonderful partner in and the convening body for the DC Flex initiative through which energy and technology industry folks are coming together to advance data center flexibility. And what I said was we need three key next steps. The first is for the tech and the energy industries to come together to do demonstrations. We at Emerald AI with Oracle, Nvidia and the Salt River Project utility in Arizona did the first one of these demonstrations through EPRI DC Flex. But now we need 10 and 20 and 100 more of these demonstrations at larger and larger scales. We're really excited at Emerald that our next few demonstrations, you know, one is kicking off on Monday in Chicago, also through EPRI DC Flex with Commonwealth Edison. Others are going to happen in the United Kingdom, all over the world. The second thing is that we need real progress on accelerated interconnection processes. Shanu, as you mentioned, this is the real incentive that's going to entice data center and cloud companies to come to the table and say we're willing to do flexibility if we can get faster time to power for a new data center or take an existing data center and get a larger power connection so that you can swap out the GPUs for the latest Nvidia generation, switch to liquid cooling, etc. And the third thing was we need to advance the technology frontier for how we achieve flexibility. We have so many tools already. We have infrastructural tools like batteries, and we're just at the cusp of commercializing these computational tools like spatial and temporal flexibility that Emerald does. So you do those three things and I think there's a real forward progress over the next 12 months to scale up data center flexibility to large commercial scale and then to multiple data centers to create, you know, one of the things you talk about, Amy, a virtual power plant of AI data centers. That I think is the holy grail solution. It's going to be the most powerful demand side solution we've ever seen on our power grid. Because data centers are massive. They're virtually controllable. They have millisecond response time and they're connected with each other at the speed of light to move work all around the world. I think the sky's the limit at that point.

D (49:37)

Yeah, that is a great point. And certainly I think I would have a certain amount of faith in the ability of the free market to deliver solutions here because the rewards are so massive, right? The incentives, if you can get this right, if you can work out, as you say, how to kind of reconcile the irresistible force and the immovable object, if you can work out how to get over these issues of flexibility in order to get more capacity connected and operational and on the grid sooner. That's going to be something that is going to be very richly rewarded. That's going to have a huge payoff. And as you say, a lot of people will be very strongly incentivized to make their work, I guess, including you, Varun.

A (50:21)

I'm super excited to push this work forward. I'm so grateful for the chance to speak with all of you. Congrats on a great show.

D (50:28)

Well, thanks very much indeed. Thank you very much for joining us. We do unfortunately have to leave it there, but thanks very much indeed. Amy.

B (50:35)

Thanks, Ed. Thanks, Varun. Thanks, Shanu. Always great to see you both.

D (50:38)

Yeah, many thanks, Shanu.

C (50:39)

Yeah, thank you everyone. It was a pleasure joining you all.

D (50:41)

Thank you again. Varun. All the best with your efforts. Will be very interesting to see. We should talk again perhaps next year and catch up, see how things are going for you. But certainly hope it all works. Thanks to our producer, Toby Biggins, Gilchrist and above all, of course, as ever, many thanks to all of you for listening. We really do value your feedback. Please do keep that coming. Leave us a review, send us an email, whatever you want to do, and we'll be back in two weeks with all the latest news and views on the energy transition. Until then, goodbye, Sam.