The Exchanges

Q not new. Um, transformers had been invented by that point, but like had not been, you know, broadly applied the way that they are today. So what, what is it that like this new wave of AI unlocked by things like by, by transformers and convolutional neural networks and so on? Like, what does that Enable above and beyond what you would have been able to do five years ago?

A Yeah, that's a good question. So the way I look at it is, so transformers are very similar to graph neural networks. Um, they, both of them are, so actually, let's, let's, let's take this back. So we used to use often convolutional neural networks, and the idea here is it learns a little function that's sort of local in an image. And then it sort of applies that same function everywhere, and then you stack up sequences of these layers, and that eventually lets you, like, one, the information on one side of an image communicate with the information on the other side of the image, because it's a hierarchy. A transformer architecture allows you to make a direct connection between the information on one side of the image and the other side of the image, the same way the graph neural network does. And I like to think about it like graph neural networks, because what it's like saying is, Well, in a graph, you have nodes, and you have edges or connections between the nodes, and a longer connection between nodes is for nodes that are farther away, and shorter connections are for nodes that are closer. So if you use the graph neural network analogy to describe the older convolutional networks, it's like the graphs are all small. They're all kind of, everything's kind of close. It's like nearby in an image. Graph neural networks allow you to choose how far away you want information to in…

AI assessment note: “A transformer architecture allows you to make a direct connection between the information”

Q There's a good comparison there. What do you think of as being, I guess, you just described something that is similar about what you can do in weather forecasting, thanks to a transformer architecture, to what you can do with large language models, which is what most people are going to be most familiar with in the, in the new wave of AI. What's different?

A So, that's a great question. I think what's, what's interesting is that the way that, so in, in language, the text is understood to be, is treated as a sequence. It's like, you know, token, token, token. We are also modeling sequences in weather, but we're not allowing our models to look too far back in time. So, because weather, weather is actually different from text in a fundamental way. Um, in fact, most physical processes are. They are what's called Markov, in that the most recent state of the system determines the subsequent state. So, like I said, in text, that's not the case. Uh, you know, right now, I'll just pause You didn't know what word I was going to say next, right? It kind of depends on the context, a bunch of words behind it or earlier. With weather forecasting in principle, if you know exactly what's happening right now, you can fully predict what's going to happen next. You don't need to look further back in the past. So we actually use transformers not to model the spatial, the interactions in weather over time, like the sequence of text, but in space. So in text, you actually don't have a sense of spatial structure, right? You just have one sequence of text. It's just word, word, word. When you read, you just see word, word, word. In weather, you have spatial structure. You have weather all over the Earth at the same time, and it's all, you know, especially…

AI assessment note: “weather is actually different from text in a fundamental way.”

Q big centralized cloud data centers, or will some or much of it potentially shift either to one of the other two categories you described, sort of edge, uh, localized or fully localized on device. So let, let's talk about the edge version first, which is essentially Smaller data centers, still data centers, but smaller and more local. What's the argument for why that might happen? And what are the limitations?

A So, so the argument in favor of edge computing is mainly the proximity to the end user, right? So when you, so we have, we have been conditioned in an era before generative AI that when we access internet-based services like a search engine, We expect the answer to come back on the order of a hundred milliseconds. That, that is the order of magnitude that we're, we're talking about. And, and as a result, to get those hundred millisecond latencies, oftentimes you require computation closer to the user. So you don't have to travel across the internet. You don't have to travel from the west coast out to the east coast and back again, uh, the data, I mean, um, and, um, and get that answer back, uh, in a timely way. What is interesting with generative AI is that we are being reconditioned to tolerate much longer delays. So if you use something like GPT or you use something like Claude or your favorite chatbot, oftentimes it's just sitting there thinking for seconds and seconds, maybe tens of seconds before it gets you the first token. So, so the question there is to what extent we care about that latency and need that really fast responsive access to the answer.

AI assessment note: “the argument in favor of edge computing is mainly the proximity to the end user”

Q not like we're waiting on workloads to show up that could accommodate this. And yet, if you look at everybody Most everybody building data centers, certainly the hyperscalers, and I think the colos and, and folks as well, you know, the focus continues to be on, we gotta find big sites for big data centers. Why don't we see more development of this small, smaller scale edge AI inference world?

A I think it really depends on, on the workload and the application, and we don't know, I, I view AI as a more fundamental basic technology, and we don't necessarily know what application or capability will be layered on, on, on top of it. I, I'd say that we've been talking about edge data centers a lot. There are other words for this, uh, type of data center. A content distribution network is one of those examples, a CDN, um, or a point of presence, uh, a BOP that, Uh, the facilities are sometimes called, and they exist in fairly significant numbers. Content distribution networks ensure that when you want to access, for example, newyorktimes.com or wsj.com, Your web page is not being served from the other end of the country. Those web pages are sitting close to you because the content distribution network took those updated web pages and moved them to facilities near you, data centers near you. Likewise, companies like Meta, um, when they have Instagram or when they have, uh, these social media applications, they also have these points of presence that supply data from, um, Local points of presence rather than retrieving content for your feed from across the country. So we already see that, but these are application level, uh, performance requirements, whether they be for social media or for other sort of news content. Once it becomes clear what applications of AI really drive f…

AI assessment note: “Once it becomes clear what applications of AI really drive further inference, uh, deployments”

Q up local, that, that's a significant shift and has, has pretty profound implications for the energy picture as well. Are you saying that 80%, just to Just to pin you down even a little bit more, is that local in the sense of being at the edge, or is that local in the sense of being on device, or like, what do you think the split ends up being there?

A Yeah, so I, I think Of the 80%, I would say most of that will be on the edge. Um, like it, I, like I suspect it is today. I, I think that, um, if you look at what we, what we talked about earlier, the content delivery networks, um, points of presence, they've probably identified 20% of the content that 80% of the people will be looking at most of the time, and they're putting it at the edge. I think maybe on the order of one percent, Ends up being put on your consumer electronics. Actually, even for today's compute, When we set aside AI, there is a trend towards, um, consumer electronics, hiding that flow of data back and forth between the, uh, the device and the edge for you. Right. So sometimes they'll like, if you use a, a cloud storage service like Dropbox, or if you're using a photo storage service, they will let you pretend that you have access to all of your videos or all of your photos and all of your documents. And they will transparently, behind the scenes, move things back and forth between the data center and your local device. So you may think you have all of it, but maybe you've only got a tiny sliver, less than one percent, on your local device.

AI assessment note: “Of the 80%, I would say most of that will be on the edge.”

Q Okay, so then on to inference. So you're saying inference does not contain that same challenge. So is there any, what is the downside to shifting inference workloads to the edge?

A Um, to my knowledge, there isn't much of a downside because the reason why inference, um, is amenable to edge computing is because when you send a prompt to, uh, for processing by a large language model, that prompt is probably handled by one GPU or maybe Eight GPUs inside a single machine. So, and, and the reason that is, is because the model sits in that machine, the data sits in that machine, and all of your prior conversations with that bot have, are sitting in that machine. And it's a very localized, uh, piece of compute that needs to be done. And you don't need tens or hundreds of GPUs to be coordinating to give you an answer back. You've got that one GPU or a tightly coupled GPUs giving you that answer back. And that is amenable. That is great for, for edge computing, and we can certainly supply that.

AI assessment note: “to my knowledge, there isn't much of a downside because”

Q Wait, how does that work? So, so that's a very, it's a good specific example. So I guess what you would do is Issue the purchase order for all of the material before you commit to a price to the customer? Like how do you avoid the terror?

A Yeah, exactly. It's timing it all out. So, um, when, uh, uh, just to compare this again to how it works today is you'd only, when you, when you say, Hey, we're the EPC, we're going to give you a performance guarantee or some sort of fixed firm price when you go to buy this or fund this project with like a notice to proceed or final investment decision. At that point, you just, you literally don't even know All your bulks, like, so your, all your commodity equipment, your, the steel that you need, the, the wire that you need, etc. Um, we will know all of that at this point in time because we're not 30% definition, we're a hundred percent engineering definition. And because we have this entire equipment list at this point in time, um, we know either how to say, hey, let's purchase everything at the, at the same day that the project gets funded, let's go either purchase everything, or let's design hedges from a cost perspective so that we're not exposed to some sort of volatility or risk.

AI assessment note: “let's purchase everything at the, at the same day that the project gets funded”

Q Can you articulate a little bit, like, what is distinct about this from the, you know, that EPCs have been using engineering software that is super mature for, for a long time. Maybe that's the problem, but, like, what's distinct about what you just described from what you could just get off the shelf to design a project?

A Most of the software in this industry really hasn't Hasn't changed much in the last like maybe 20 years. Um, and I, again, kind of always come back to the incentive structure. There's no real incentive structure for an EPC to want better software. It's the same reason your law firm doesn't necessarily want like AI software to help them accelerate, you know, and reduce their billables. So the way these softwares work now is that there's a bunch of kind of distributed software solutions or, um, or maybe decentralized solutions. So you'll have a CAD tool that you use to actually like do the designs. You'll do almost all of your kind of rough engineering or hand calcs in Excel. Um, you'll use dedicated simulation software for some sort of fluid design or structural analysis or whatever that might be. And then, um, and then you use another, another tool to like review those, those PDFs, you know, uh, uh, with a tool to put red lines on them and go through design cycles and then push to drafters and, um, and then other tools to manage, you know, how do you reach out to your vendors? And anyways, there, there, there's all these different solutions that, uh, software solutions that you're using in order to, um, like kind of quote unquote, like do engineering. And, um, Um, what, what we built, uh, kind of from day one was just a modern version of the software. So, uh, we put all those d…

AI assessment note: “we put all those different disparate tools into one underlying platform”

Q on balance, basically everybody we've talked about so far, you know, sometimes it's nuanced, but pretty much everybody is incentivized to play up the power bottleneck. Um, and indeed, you know, they are. It is also real, but, but they are playing it up. Who do you think is on the other side of this equation? Like, who is incentivized to say that power is not such a big bottleneck?

A Yeah, and I think this part's a little bit more interesting, too, because to your point, I feel like a lot of people benefit if the power constraint remains. And, you know, that's just like an okay thing to observe. It's just facts are facts. But on the other side, you know, I'd, I'd call it a few buckets. And, and the, the first that I think is independent power producers, and this might seem counterintuitive, but let me lay out my logic. Is if you're an independent power producer, that means that you have a fleet of existing assets today. Those generate money based on selling into electricity markets. And so, you know, the input to their revenue is electricity prices. If we're in scarce conditions and those continue to grow, they generate a really healthy incremental margin on that, right? Because they're not really building out new assets. You have slightly higher O and M perhaps by running your assets harder, but you're just flowing through a high electricity price. That is extremely rewarding. Profitable for them. And that's actually been the thesis for a lot of owning these assets. So these, these are like the constellations, the talents, the energies of the world. One thing I noticed, it's interesting. If you look at the, like the last two earnings calls, right, you have the folks like the CEO of constellation talking about the fact that, you know, energy prices actually…

AI assessment note: “the first that I think is independent power producers, and this might seem counterintuitive”

Q them, um, what we are hearing from these various parties, and then the incentives that they have to say what they are saying, or maybe to say what they're not saying. Um, But let's start by categorizing. Like, walk me through at the high level, like, there are a bunch of, there are a couple groupings of actors that have similar incentives. So how do you think about the groupings?

A Yeah, at a high level, I'll try to keep it simple, but, you know, again, there's various degrees of nuance required. But at a high level, you know, let's start with the hyperscalers who are, you know, the folks that are driving a lot of this capex and investment cycle. You know, these are, think of it as, like, the cloud service providers, the Amazons, the Googles, The, um, you know, um, Microsoft's of the world, as well as including Meta, which is another hyperscaler, even though they don't have a legacy cloud business, um, that drives, you know, investment spend on GPUs, and that's kind of the equipment, like you have the technology hardware, if you will, that includes, you know, GPUs, CPUs, custom silicon offerings. That moves upstream into like the supporting equipment, which includes like electrical and cooling equipment. Uh, and then if you kind of go outside the data center, then you start to get into You know, who supplies that power equipment, as well as the overall actual power of the facility. And moving upstream there is like, you know, who actually builds power, which is like utilities, um, as well as, you know, the, the labor and the EPCs and the engineering that goes into those types of facilities. So if you kind of just go from who's spending the capital and follow that down the, the stream, that's a general way to think about some of these bigger pockets. And s…

AI assessment note: “start with the hyperscalers who are, you know, the folks that are driving a lot”

Q your job is to provide services or hardware into the build of data centers. My first thought would be That everybody there is aligned in their incentive to promote the power scarcity narrative. It's just, it just furthers their, the argument that there is limitless demand for their services, and we're just going to sell out as much as we possibly can. Do you think it's more complex than that?

A I think that one is pretty straightforward to your point, is that, just going back to our prior point about long lead time items, and the most constrained, a lot of this is equipment and As well as the labor, those are the direct beneficiaries. What do they get? They get higher backlogs. They get longer duration backlogs. They get better pricing in typically industries where you don't get, you basically get minimal pricing or as much as the market will bear. Um, whereas now you're pricing power for the first time in, in generations, right? For certain equipment providers or certain labor producers. So this one is pretty straightforward and you see that across the earnings, right? Um, you can look at the earnings transcripts of companies like Quanta, like record backlog, uh, MosTech talking about, you know, record Activity in their pipeline business, as well as really a strong, even clean energy pipeline. Um, you, you can go to the equipment or you send them like somebody, let's see something like a vertive, which has a backlog that was up 30% organically year over year. You have eaten, which is up 20% backlog organically over year on, you know, like billion dollar books of business. And so I think there, that's what you'll see to your point, like a very limited narrative on talking down the potential market opportunity, because if you think about it from their point of view, ri…

AI assessment note: “I think that one is pretty straightforward to your point”

Q How do you think about the, there's an, there's another category, which is like the, I'll group together like real estate owners, landowners, and then the, the like powered land developers, the developers who are going off and trying to find sites and make them powered and then sell them to a colo or a hyperscaler or whoever. What's their incentive?

A Their, their incentive is to talk up the constraints as well, because it increases the value of the asset they're owning, which is the real estate or the powered land bank. If we think about, Just solving again, if there's a constraint, that means that anything that accelerates you through the constraint is, is valuable, right? Or increases in value as a constraint gets worse or it remains. And so if I have a powered land bank, that means that I can move a lot faster than someone that has just a plot of land that has no interconnection access or anything like that, right? So I can charge a more premium price to developers and new clouds, et cetera, anyone else that would want to be further along up the development cycle. And so they're generally universally talking up the opportunity that they have, um, and the ability to move faster. I'd say, you know, a lot of what we're seeing there is, or at least what we hear chatter wise, right? Is that like you have transaction values that are much higher than they've ever been, or at least in recent cycles, um, as folks all are chasing the similar opportunity. So pretty much any land banks, uh, that are powered that are near plentiful energy sources, something like, you know, in the Northeast with narrow gas fields or in West Texas, right? You're seeing construction activity be plentiful because people think that it's a much faster Path…

AI assessment note: “Their incentive is to talk up the constraints as well, because it increases the value”

Q And then how much is all this stuff tied up in the question of permitting reform? How much does that matter? Because it, it doesn't, it's not clear to me whether that's actually on the horizon. It may be, but it certainly hasn't happened yet. Do we need it?

A I think, I do think we need it. I mean, you can, You can build things here and there, um, under the current regime, which again, you know, was proven by 10 years ago we, we did it. Um, but there are real limitations, uh, with that. Um, and some of it is this big inter-regional opportunity. Like, we don't really have a system in place to build inter-regional transmission from region to region, MISO to SPP, or SPP to the interior west, or interior west to the coast, all of these neighboring regions, there's a massive untapped value of transmission there, and, you know, we don't have a regional transmission organization with a planning process to deal with that, and, and this whole More macro, looking at the whole industry on a more macro scale, you know, we're evolving out of a industry that started with 3000 utilities doing their local thing in their local fiefdom, uh, without much of a, you know, tie between them, and so we're crafting onto the top this more regional and then interregional approach, and it's a process, a multi-decadal process that we're still working through, and we're nowhere near the end game on that, so we're still working through on that. Permitting reform, at least as, um, uh, set up in the Manchin Barrasso Energy Permitting Reform Act, EPRA, from a couple years ago, it has a transmission title. That has a big focus on interregional transmission planning, …

AI assessment note: “I think, I do think we need it.”

Q I'm really interested to hear more about what is happening in fusion. You know, you want to make New Mexico the center of the fusion industry, right? What does that vision look like through the end of the decade, and what kind of activity are you seeing right now?

A Well, we expect, we, we want New Mexico to be the center of fusion, if not for the US, for the world. You know, we're very pleased to announce that Pacific Fusion was moving their R&D facility here. Um, we're in conversations with a variety of other fusion companies, and again, this late, this talks a little bit about how we're doing it. So, just a month ago, the State Investment Council Created a new fund of over three hundred million dollars, part with lower carbon energy, uh, to fund fusion and advanced energy projects in New Mexico. So, you know, we're bringing that equity to the table. Uh, we've got Sandino National Labs and Los Alamos National Labs. So, if you think about companies like Pacific Fusion, who use magnet-based fusion technologies based off of the Z machine, uh, which again was the first place to do Kind of fusion here on, on Earth. That alignment and that proximity to be, work with those scientists is a big draw for New Mexico. Los Alamos National Labs, on the other hand, is really key to the laser-based fusion approach. So, I think bringing those sort of aspects together gives us a very powerful, unique value proposition, and then you layer that on with access to capital, and New Mexico becomes very attractive.

AI assessment note: “Pacific Fusion was moving their R&D facility here. Um, we're in conversations with a variety”

Q And to what would you ascribe the bubble bursting? Like, fundamentally, why did it, I mean, all bubbles burst, if they're bubbles, definitionally, but like, what made this bubble burst?

A Yeah, so great question. So why would the bubble burst? Well, what's green hydrogen all about? It's all about avoiding emissions in very hard to decarbonize sectors of the economy. Green hydrogen Unfortunately, has been and continues to be an expensive solution to those problems. Now, there's no cheap solution to those problems. It's worth pointing out. Right? All solutions to those problems are expensive. But at the, the peak of the hype, there was a naive hope that society was willing to pay the difference for deep decarbonization because it had turned into a, decarbonization had turned into a global priority. Right? We had, kind of, Europe, the US, Asia, everybody had a hydro, every country had a hydrogen strategy. The world's changed dramatically since then. And I think that political climate is resulting in a retraction from commitment to retooling critical industries at some expense for the purpose of decarbonization. Green hydrogen, it is also true. Is expensive, and it's been too expensive, uh, over the last few years. Um, and it's been too expensive for a couple of reasons. If you look at the cost of making a kilogram of green hydrogen, um, today, let's say in southern Europe, it might be unsubsidized. It might be maybe six dollars U.S. a kilo. I think rough numbers, that might be about right. That can be broken down Roughly fifty-fifty into CapEx and OpEx. OpEx being …

AI assessment note: “political climate is resulting in a retraction from commitment to retooling critical industries”

Q is, like, why is the CapEx for, for electrolyzer, for electrolyzer systems? Why has it Been so high. Can you break it down? Like what is the, what is the cost stack of a traditional electrolyzer project? And then you can briefly talk about like what electric hydrogen is doing differently there. But, but first I want to start with like, what, three dollars a watt, like what happened here?

A Yeah. What happened at three dollars a watt? Um, so if you look at the cost stack of, of a conventional electrolyzer, you're a project developer, let's say in In, in Spain and, uh, and you're trying to build a hundred and 150 megawatt electrolyzer someplace, right? Um, how are you going to go about doing it? You're going to contract with an EPC. It's an engineer, procure, construct company. You're going to select your technology. Maybe you're going to, I mentioned two names before, so I'll pick one. You're going to maybe select Siemens. Um, that's your technology provider. And you're going to go through what's called a feed study, which is front end engineering design. That's where the EPC takes all of the requirements from the equipment supplier, the technology supplier, and figures out how to build that thing on your site, right? The total installed cost is what drives levelized cost of hydrogen. It doesn't matter what the electrolyzer costs per se. It matters what the constructed cost of the plant is. And in a typical project like we're discussing, Roughly half of the total installed cost goes to the EPC. What are they doing? They're grading the plot. They're managing stormwater. They're building the building that the electrolyzer is going to go into. They're building the substation. They're sourcing and selecting all of the support equipment, whether it's chillers or air co…

AI assessment note: “Roughly half of the total installed cost goes to the EPC.”

Q you solve it? How do you, I mean, not just the EPC, but they're tied to each other. Like, what is the, What's the solution to the CapEx problem? We could talk about what's the solution to the overall system cost, too, but on the CapEx side, how do you get away from this EPC ballooning and drive down the cost of the stack? Because I think you need to.

A Yeah, and I think it is important to talk about the overall solution cost as well, because they're not necessarily decoupled, but, uh, just starting with the capital cost, the The secret to it, it's no secret, is to think at the system level holistically, right? So what are those, all of those costs? Maybe let me back up. When we think about, um, when we think about our product, and the scope of our product, and how we present it into the market, the, um, the guiding light for us, the North Star, is levelized cost of hydrogen, right? So we think about it from the perspective of our customer's project pro forma. And when you look at the problem that way, you very quickly have to, uh, accept that the EPC cost must be in scope for your engineering team to try to address, right? There's just a big chunk of cost that's being thrown to the wind, left to others to contend with. And you could throw your hands up and say, well, yeah, but there's nothing that can be done about that. Because construction costs what construction costs, but it turns out that's not really the case. Now, there's a deep technology component to the solution, which is actually, quite simply put, making the electrolyzer as dense as possible. That enables the balance of plant construction to also be quite dense and small, and hence amenable to what's called modularization, which is just the chemical industry's ter…

AI assessment note: “making the electrolyzer as dense as possible. That enables the balance of plant construction”

Q question everyone gets, you get all the time, I'm sure, is, um, are we just gonna get a flood of cheap electrolyzers shipped from China And one, is that gonna solve the problem, right? From a societal perspective, are we just gonna get cheap electrolyzers like we've gotten cheap solar panels from China? Um, and two, like, what does that mean if you're not a Chinese player in the market?

A Yeah, let's talk about that. It turns out a large industrial electrolyzer green hydrogen facility looks more like a, you know, if you just walk up to it, it looks more like a gas generation plant. Like a combined cycle plant, then it looks like a solar array. What do I mean by that? I mean, it's a complicated thing. It's got a lot of pipes and valves and, and stuff going on, um, which makes it much less amenable to low-cost manufacturing at scale, kind of putting a low-cost product in a box in China and shipping it to someplace and letting someone install it. So it is true that the global market is flooded with really cheap Chinese electrolyzers. These are like Two megawatt electrolyzers that are the size of a school bus and way more than a school bus, and they need to sit in a building and have this, like, chemical plant wrapped around them to support them and operate them. So the, um, the normal approach that China has used in other industries to drive cost out, um, really kind of only addresses the cost of the electrolyzer stack or stack and power conversion. It doesn't address the EPC component Of the cost buildup. And so it's a limited, there's a limited opportunity for overall total installed cost reduction, taking that approach. It's real, by the way, you know, we see, um, integrators building or promising to build using Chinese equipment systems in the European market a…

AI assessment note: “I don't think the same game that worked in solar works in green hydrogen.”

Q your view, okay, if you can sell at a thousand bucks-ish today and drive, and that's, you know, entering the market at a thousand bucks and a And presumably we'll be able to drive costs down from there if you're able to scale. What, what does demand look like? What does it rely upon in terms of policy support? What's your view? Like, what's your thesis of the market here?

A Yeah, the current market is still policy supported, primarily in Europe. So Europe has, to get a little policy wonky here, Europe has the Renewable Energy Directive III, REDD III, so-called REDD III. RED-III has a component in it called RFMBO, which is the, um, the part of the RED-III law that stipulates the, um, the gradual conversion to, partial conversion to renewable molecules, e-molecules, and these, these range from, uh, hydrogen itself, um, to things like green methanol and green ammonia for various purposes. Um, That policy, I think, has been, geez, I don't know when, when red two and three were enacted, but it's been years. Um, but the way European policy works, uh, the EU law has to be, what's called transposed or translated into national level rules, which then drive project decisions. And those rules are being transposed as we speak. It's underway. The process is underway. Uh, I think Romania has transposed, uh, Red Three now. Uh, I think Netherlands are close. Germany's very close. Uh, Spain is about to, to do so. Uh, long list. Um, and so we are seeing, uh, in the short term a policy-driven market. What do I mean by policy-driven market? I mean, markets where the green product is more expensive than the gray product, but There is a compulsion to make a conversion and hence, ah, an absorption of that cost by society. That's, that's, that's the nature of the market …

AI assessment note: “the current market is still policy supported, primarily in Europe.”

Q but I came away with a pretty, I think like a sober view of the, the path there. Okay. So fast forward a year and you left Google DeepMind, started a company to do that amongst other things. So I guess the first question that I have for you is, What changed in the last 12 months to make you, to give you conviction that, like, now is the time?

A Great question. So, when we talked, I was doing research in the field of computational material science and machine learning. You know, specifically, we were using graph neural networks, we were using density functional theory, and we were trying to discover materials. One thing that changed since our discussion was, uh, the LLMs have Improved even further. Um, so at the time, I wasn't using LLMs much at all. Um, but I think right around when we were talking, the O-one came out, right? The reasoning models started showing up. And that was a huge update for me, because you might remember that one of my big concerns is machine learning works best on the training set distribution. But in science and technology, we almost only care about auto-domain generalization, right? So what O-one showed is if you spend test time compute, you can get better results. So that was very exciting to me because there was one way of investing resources that was beyond the training set.

AI assessment note: “One thing that changed since our discussion was, uh, the LLMs have Improved even further.”

Q two. I'm curious in practice, like how you imagine that, that feedback loop working. So is it a traditional, you develop a theory, you run an experiment, you You generate data from that experiment, but in this case, you feed the experiment back into your customized LLM as an additional set of training data, and then that's the way that the loop works, or is it more complicated than that?

A Yeah, exactly. I mean, it's pretty simple, I think, as you said. So the LLM can propose, for example, synthesis recipes, or it can propose simulations to run. And because the LLMs are pretty good at tool use, it can actually do it itself. And then you get some results back. So the results from experiment could be some characterization data. Results from the simulation can be some, you know, trace or some, uh, simulation you did. And now the LLM can go through it with the context of its previous training, maybe the context of relevant papers, textbooks, but also now the results that it just got that no one else have ever seen. And then now it can Kind of tweak the experiment, tweak the simulation for the next step.

AI assessment note: “Yeah, exactly. I mean, it's pretty simple, I think, as you said.”

Q maybe Tyler, you can comment on this because I've seen you say it publicly. It seems generally, maybe outside of FERC, but in the public domain, this has been pretty well received, broadly speaking, which, I mean, I think does run in contrast to, I remember that previous letter about coal, 90 days of storage, and so on. So your, your sense is that the vibes here are good, Tyler?

A So, so far, you know, let's not discount that There will likely be other perspectives that have not yet been represented, but no, look, I think it was very significant that, um, you know, Commissioner Rosner came out of the gates expressing, you know, an eagerness to, to work on the proposal. You know, on the other side of the aisle, you had, you know, Senator Mike Lee that came out, you know, strongly supporting it, and a variety of different stakeholder groups, at least I've seen, and, and companies that are involved in this space, um, seem to view it generally favorably. I think Just to parse out a key distinction, right? There's this jurisdictional question, and on that one, I think there's obviously going to be a variety of perspectives, and there will be concerns on the part of, especially of some state commissioners, officials, and certainly the investor-owned utilities. But with respect to the substance, that's where I've seen the most excitement, and I know we'll get into it, but in terms of the What, what this would actually do to sort of improve the interconnection process, that's where I think there's been the most positive reception.

AI assessment note: “Senator Mike Lee that came out, you know, strongly supporting it, and a variety”

Q along, which is the, the G Hitachi BWRX 300. That's the one that's going to get built in Ontario and maybe TVA territories. That'll be the first one. So the basic question is like, are we going to see a ton of new reactors deployed in the market? Or are we mostly just going to see the one or two that Sort of have already gotten mostly through the gauntlet.

A I think the short story is there already is a nuclear renaissance happening globally, hasn't quite caught on yet here in, in North America or in most of Europe, at least Western Europe. Um, and we can see the answer playing out, which is that there's just a few reactor designs that are getting traction, and basically it's the ones that you mentioned, especially the AP 1000 at this point. Um, Um, and, you know, China is, is very much driving that. And it's actually one area of technology in which China is still, uh, you know, buying a significant, buying a significant amount of technology from Western, a Western vendor. Um, and I think that that same pattern is going to play out in the nuclear Renaissance as much as it happens anywhere in the world. There, there just can't be a Cambrian explosion of new reactors. The, Industrial logic of the nuclear industry just doesn't lend itself to that. I think best case scenario, it's bad for the industry if you end up with, you know, four or five competing reactor designs that are relevant in any given region, because really what you need for nuclear to come down the cost curve is you need, you need, uh, economies of scale throughout the supply chain, and you need to really come down the learning curve, uh, Uh, when it comes to deployment. And I, and I would say that learning curve extends all the way from policymakers and regulators down…

AI assessment note: “There just can't be a Cambrian explosion of new reactors.”

Q my world is a pretty small portion of the world. And broadly, I think even in data center universe, it has not really been considered very much yet. Like mostly it is, we're going to build a data center. Can we get grid capacity? If not, can we build a big gas turbine? Um, and so my argument is, because of all that, it is still under-hyped as of today.

A Yeah, that's an interesting perspective. I would say in my world, which is a little bit broader because I talk to policy folks, and I talk to big businesses, and I talk to VCs, is that it actually used to be more hyped because it was one way in which everybody was justifying having all these big renewables goals being set up because, look, don't worry about it. I know it is variable, but we'll have these technologies that are just going to come Online within the next few years, and they'll make it easier to manage this variability in renewables. At that time, it was pretty hyped. Now, I would say it's underhyped, because it's not something people talk about that much, but those solutions are actually starting to bubble up. So we had the CEO of Octopus Energy, which is the largest utility here in the UK on the pod, and this year, they've launched a BYD lease program, where you can just Pay monthly to get a BYD electric car, and they would give you 12,000 miles of free range in a year, as long as you make sure that you put your car Into the charging port whenever you're home, because they'll use the car for a virtual power plant, essentially. Um, that sort of thing is now becoming very frequent. Like, I get text messages from Octopus Energy saying, hey, 12 PM to two PM today, free electricity, use all you want. Um, and so, I feel like now those technologies are here, and they can…

AI assessment note: “Now, I would say it's underhyped, because it's not something people talk about”

Q So that brings us to the next, uh, Topic, which is transformers. Um, transformers are this object that sit between generation, between a power plant and your home, because they have to make sure that the voltage at which power is delivered is just right for the devices that you're going to attach to it. And, uh, where do you land on whether transformers are overhyped, underhyped, just right, hyped?

A So again, if the question is, Uh, overhyped, underhyped, as a bottleneck to load growth, to meeting load growth. Then I think that they're overhyped. I wanna clarify, transformers are incredibly important, um, and there is a bottleneck, and the lead times are very long, and I've, I've witnessed it firsthand both from the utility side and the load side. It is a problem. I don't think it is the rate limiting factor even today with long lead times. It's an annoyance to get anything built. Things will still get built, and, and certainly those lead times for transformers are still shorter than they are for gas turbines. Um, But I, I do think we also will see a lot of new transformer manufacturing capacity come online in the next few years. It's, the, the lead times are gonna come down, and I think we'll see a wave of, of new technology. I'm an investor in a company called Heron Power, which is building solid state power electronics. You know, I, I think stuff like that is going to revolutionize that sector, but even in the absence of things like that, I don't think it is going to be the thing that stops The load growth from getting met. So I'm going to say overhyped on that one.

AI assessment note: “Then I think that they're overhyped.”

Q a watt in the U S cheaper in Australia and other places. Um, but anyway, give me the mechanics of like, how do you get a, just from a pure, you have, you know, you have customer acquisition costs and you have, um, Uh, labor costs and truck rolls and smaller batteries and all of that. So like, how do you overcome all of that to deliver a cheaper megawatt?

A Yeah. So this is kind of the heart of the issue here. And I think the shortest version of the answer is that vertical integration is the magic. And we'll talk about what that actually means tactically. And I can kind of start by saying, if you look at a utility scale battery deployment, you have to buy or lease the land that the battery sits on. You have to pay for the interconnection to the grid. You then wait in the interconnection queue, which adds additional cost. You then do what most firms consider project development, which is some level of construction to level the site and prepare it for the system. You have a big EPC firm come in and plug in all the hardware that you buy from an OEM that adds a bunch of margin on top of the cells that are reasonably commoditized. Um, and so there's a bunch of kind of line items in the model that add to a cost that we think is higher than where we can get by vertically integrating. So compared to our system, we do have CAC that they don't have. Um, they have install costs. We have install costs, too. We'll come back to that when we talk about design. Uh, but we don't buy or lease the land that the battery sits on. We don't pay for interconnection to the grid because it's already there. We don't wait in the interconnection queue, obviously, because this is behind the meter. On the hardware side, we're designing and manufacturing our own…

AI assessment note: “the shortest version of the answer is that vertical integration is the magic.”

Q you base are operating that battery all the time doing arbitrage or what else you're going to do in the meantime. So are you guaranteeing some level availability to the customer or how do you, how do you square the circle between like you're going to be charging and discharging the battery every day, but you want there to be some reserve available to the customer when there's an outage?

A So the short answer is yes, we guarantee 20% of the capacity of the battery to the customer no matter what. The reality is that The discharge window of the system is not very long. It's, you know, one to two hours a day. As you know, and listeners of the pod know, power prices are, are spiky and unpredictable, but like reasonably predictable in terms of kind of the, the pattern throughout most of the days, and it's different in the summers and the winters, et cetera. But, um, the windows in which you are discharging, um, they're, they're, they're reasonably predictable, and outages and high prices are, are actually not as correlated as one might think. So, uh, what I'm saying is that The likelihood that an outage happens at the bottom of the discharge window is statistically not improbable, but, but reasonably low probability given the fact that there are 24 hours in a day and only one or two of those 24 hours. The battery is kind of low state of charge. Um, so most of the day the battery spends its time at a higher state of charge. It depends on where the battery is and what the optimization function is. Um, but the reality is it is very unlikely that an outage happens at the bottom of the discharge window. Even if that does happen We do maintain 20% state of charge for that situation, and as a benefit of having really large systems, you know, our next generation product, whic…

AI assessment note: “we guarantee 20% of the capacity of the battery to the customer no matter what”

Q That too, but actually before that, who are the customers that participate? I mean, I think that the expansion from it's, you know, uh, Jill Schmo running a factory to, like, a pretty wide array today is interesting. So, like, if I look, if I were to look at the Voltus platform today and the customer set on the platform, like, how would you, how would that pie be split?

A It'd be, so to give you a sense, there'd be over 50 different verticals just within commercial and industrial, and then there'd also be residential. So going, uh, Maybe smallest to largest. You'd see everything from an electric vehicle in a home to a smart thermostat to, uh, mom and pop kind of retail shops, big box stores, any kind of commercial load, school districts, wastewater treatment plants, on up through larger industrials, maybe first larger real estate building, commercial real estate is huge in particular in areas like New York, on up through Industrials and on and on, steel manufacturing facilities, massive loads. Those are, as a sidebar, the original large loads, in my mind, were like the paper mills, paper mills and the steel mills, and then there was crypto, and now we're seeing the, there was always data centers of the quote-unquote traditional sort, and then cloud compute, and now we're seeing the AI data centers kind of at the top. So that's the, it really runs the gamut, and in many ways, that's the strength of the portfolio. You can take things that maybe have operating parameters and constraints and pair them with other things, or other customers that have similar but different constraints, and now you can respond to what the grid needs by Tetris-ing these things all together.

AI assessment note: “there'd be over 50 different verticals just within commercial and industrial, and then there'd also be residential.”

Q Can you talk a little bit about geographies today? Like, where, it's sort of national to an extent now, and thanks in part to those FERC orders that you described that happened a few years ago, but is there significant concentration in some regions over others? Is there a significant pricing differential in terms of the value capture in some regions over others, or is it all pretty uniform?

A Let me answer volume first, and we'll come back to pricing, because they're slightly different answers. Uh, we very intentionally Built the company thinking that we wanted to be in every wholesale market in every territory eligible across North America. So that is true today. We have some markets that have emerged as just larger than others, but it's, it's by and large owing to the size of those markets. So you'll have your PJMs and ERCOTs and New York and SPP being bigger than, for example, the Canadian markets. But that's because of percent penetration against the peak load, not because that we've Really seeing that one market is where we're going to concentrate. And that's a business choice for us. There's other aggregators you'd ask that say they're heavy in one market or another. Um, but because of the second part of your question, pricing, pricing can get a little crazy out there. Anyone who's watched the PJM auction for years knows that, well, one year you think it's happening, and then the next year it's this, and it goes back and forth, and you might find a trend line over time. But if you're trying to run a business, it's really, really, really hard. And so the best way is to adopt more of a portfolio. So we have a portfolio of portfolios, and that's how we built the business is approaching it through the lens of the actual risk management and the fact that we are exp…

AI assessment note: “Let me answer volume first, and we'll come back to pricing”

Q All right. Give me a tour of ag waste to start. Like, where is it? What is it? How much of it is there? Give me the, give me the The quick high level.

A Yeah, I think many people would maybe push back first on, on the word waste there. Uh, for the most part, when you have ag residues, it's mostly things that are unutilized or very underutilized. Uh, and so in some sense there's waste there, but, uh, really I would think of it as sort of extremely underutilized, uh, and has a lot of value maybe that is, is going to waste or getting lost. Um, but you know, there's, there's an enormous amount In, in the United States, which maybe is where we can focus for now, um, you know, just corn stover, for example, obviously concentrated in the mid, in the Midwest, and particularly the northern part of the Midwest. There's about 90, ninety-five million acres of corn grown in the US every year. There's order magnitude four dry tons of stover per year on each of those, uh, on each of those acres, and so, you know, you're, you're, Just on corn stover, you're looking at 400 megatons of, of ag residues that don't really get used. They just rot on the field and, and return to the atmosphere.

AI assessment note: “Just on corn stover, you're looking at 400 megatons of, of ag residues”