Masters of Risk | Season 3 Ep. 6: AI Data Centers: How Energy Keeps Them Running - Part 2

View Full Transcript

Stewart Webster   00:00:01

Welcome to Masters of Risk Season 3 Episode 6. In part of our conversation with Wish, we explored the foundations of AI data centers and the massive energy demands that come from powering the digital economy. Today, in part 2, we're diving deeper into the risk, trade-offs and global shifts shaping the future of energy and AI infrastructure. In this episode, we'll explore the risk of spiking energy cost, the opportunities and complexities of real estate in the data center boom and the geopolitical race for power, chips and AI dominance. So let's get into it, Part 2 of our conversation with Wish Bosi. Wish, thanks again for joining us.

Wish Bakshi   00:01:32

It's a pleasure to be here. Thanks for having me again. I'm surprised you have me again. I thought I would have scared everybody off the last podcast.

Stewart Webster   00:01:39

No, our first conversation was fantastic, and I'm really eager to do a part 2 here. We originally met in July, and now we're in September doing a part 2. And as always, there seems like there's a lot of changes going on day-to-day regarding AI, and there's new pertinent news that comes out every single day. And in preparation for our discussion today, it seems like every day I had new ideas or there was news that was changing what I thought previously. And actually, yesterday alone, the Oracle Corporation came out and just said that there's a significant amount of earnings that's coming from cloud data and AI, and their stock actually jumped 40% in 1 day. which is that's a 5-year move, right? That's not a 1-day move. So what's happening there, it's investors realizing that there's a fundamental change in the intrinsic value of this company, and it's all coming from the technology behind AI. So that's one of the things I saw just in the last 24 hours. But just turning it over to you, since we last spoke a few months ago, what have been the biggest developments in the artificial intelligence space and also the energy spaces?

Wish Bakshi   00:02:45

I mean so much to cover there, right? Like the cycle -- new cycle for this is 24 hours, like you said, and we're talking like 2 months apart. I think just from the macro perspective, just think about like this is molecules, right, to electrons to 0. And so if you really want to start is like the macro picture is still the same, LNG exports, power generation, natural gas infrastructure and AI data centers. That's kind of like the -- those are the pillars of where everything is kind of going immense amount of capital falling into. So like tell the story is like we know we don't have enough turbines. That's kind of like everybody knows, you can't procure them enough. There's a 7-year lead time, can procure transformers, right, substation equipment, everything that you need for the grid. But there is about supposedly approximately 100 gigawatts of new gas power gen that's coming online, right? And so the problem now is that's roughly about 100 nuclear reactors worth of megawatts output, ignoring capacity factors, et cetera, for a minute. So let's just say like that -- and there's about 16 gigawatts right now under construction. So out of the 100 gigawatts that are in the pipeline for new power generation powered by natural gas, 16 gigawatts are under. So that's 16 nuclear reactors worth of power gen that's coming up. So now let's just play a simulation. Let's just say like about half of them are phantom gen, meaning that people have plans to build them, but they probably won't get built, okay? So let's just shave it off. Now all that gen needs gas. So out of the 100 gigawatts, we're saying about 50 gigawatts, and we're saying 16 gigawatts are being built. Now all that gen needs gas. And guess what, we have plenty of gas, but it needs to get from point A to point B. And at the same time, what's happening right now is insane is 26 LNG vessels depart from the U.S. Gulf Coast, carrying about 100 Bcf of gas from the U.S. So just to give you some context, every 7 days, we export the equivalent to fueling America for a single day. Just think about that number, right? Like that's how much gas we're exporting every 7 days that powers America for 1 day. Now on top of that, building nat gas infrastructure, it takes years, if not decades, to depending on the capacity, right? So if you want to build any of this natural gas infrastructure, you're not -- even if you have the capital, it's not going to get built in a year or 2. It takes a long time to do this. Plus the gas generation takes about 7 years today because of procurement issues. So power gas gen, then there's natural gas infrastructure. And when we have lots of natural gas, we can't get it from point A to point B, where all the generation is. So we're talking about like colocation right now. So now we have a shortage of power generation equipment. We also don't have enough natural gas infrastructure. But one thing we have is plenty of natural gas. And now everybody is trying to figure out like, well, do I just put all my gas -- my power generation next to the natural gas fields. So that's kind of where they're trying to figure out is like now they got to time AI data centers perfectly so that all this equipment comes online, all the gas infrastructure comes online, all the equipment comes online. But then now they're figuring out where to site these things. So like you're talking about not just hundreds of billions, you're talking trillion dollars.

Stewart Webster   00:06:04

Wow, that's fascinating. And you talked about liquid natural gas vessels. You talked about the pipeline building. I wanted to get some insight from you about actually building and constructing these data centers, the AI data centers. And just so our audience knows, there are a couple of types of data centers, and you actually mentioned one of them. But the first one is the enterprise data center. So that's for one single organization for their IT needs, right? And that can be from a couple of hundred square feet to thousands of square feet. The second one, you actually mentioned this, the collocation data centers, and those are shared facilities amungst different companies, and those can be tens of thousands of square feet. But then we have the third category, which is the hyperscale data centers. And that choose from the Googles, the Amazons, the usual suspects, right? These hyperscale data centers can be millions of square feet. Like these are massive structures just so our audience knows. So when you think about these large AI data centers specifically, what are the most common technical and operational risk that operators need to prepare for?

Wish Bakshi   00:07:11

I think the part that everybody is trying to figure out is how do you model these things, right? So the ISOs, these are the regulators, right? So their mandate is one thing and one thing only, keep the power prices down, but make sure lights don't go up. So what's happening is like I think there's like 2 problems. There's load profiling. So these power flow models, super, super smart people, electrical engineers, et cetera, they're in these ISOs, and they're trying to figure out like how do these large loads behave, right? And so like we talked about washing machines as the analogy last time, right? And so think about like 1 million washing machines switching on and off, on and off constantly. You need these load profiles for studies and operations because if you have like 1 million washing machines constantly switching back on and off, on and off in a very, very constrained environment and a very constrained radius, like it stresses the grid because of the hard people. We'll talk about that in a bit. But AI trainers -- the other problem with AI training is that it's synchronous and it's bursty. So GPUs work in sync. So when you start training a model, they start talking to each other. There's a whole technical nuances around them, which is like the networking, the memory, et cetera, et cetera. The problem over here is that they ramp up and down too quickly. We're talking like milliseconds, right? So now if you have like 1 million washing machines, in this case, GPUs, they're all talking to each other, and they're going up and down, up and down, like -- let's just say like a washing machine is like 700 watts, which is like what the 100 is today. They basically kind of struggle with the voltage and the frequency. They put stress on the grid. We'll talk about that in a bit. But then on top of that, what else happens is like when the training kicks in, then they have like a steady draw for a time period. And then it suddenly drops when the training is over, right? And then they restart again. So it's like switching a Ferrari on and off repeatedly, except you don't know the start or the stop time. That's where like the challenge comes in. To do this, you first have to get -- you have to first get work with the substation feeder panels, the rack and the row level. These are like all the technical nuances that all the electrical engineers are looking at. And then they have to collect this thing called the GPU telemetry data, which is basically the data that is coming out of these GPUs to figure out like how these things behave. The problem today is that when you have these load profiles GPUs, there is no 1 gigawatt data center that exists today. So you're kind of like building an R&D phase and you're basically kind of modeling these things from a hypothetical standpoint because the design of these racks are fundamentally also changing. So every 2 years, a new rack comes in with a new spec. So like last time we spoke, we went from like 50-kilowatt racks to 130-kilowatt racks to 600-kilowatt racks, right? So you're talking about like 1.5 million GPUs kind of pushed into a data center, for example, washing machines. So now imagine like I gave you the analogy, like they're switching on and off. How are you actually going to model this when there is no good sample data that exists today or an actual AI infrastructure exists today that can say, "Hey, listen, if I put this over here, this is how 1 gigawatt, 1 AI data center is going to behave. Now you're talking about like 50 gigawatts. That's 50 nuclear reactors worth of washing machines coming on and off. That's a lot. Now finally, it's a long-winded answer, but like there's so many technical nuances around it. It's like, well, just the chips today, right? They cost about $50 billion, just the chips, not the gas turbines, not the shell. And their life cycles are maybe 3, maybe 4 years. So now imagine having to upgrade the facility every 3 to 5 years for new chips that also require special equipment. So it's like upgrading your nuclear power plant that was built for 30 years, but you're upgrading it every 3 to 4 years with new equipment, with new load profiles, right? Like -- so I mean the challenge is from an operational standpoint and from a modeling standpoint, you not only have to upgrade the chips, but you're also upgrading the infra. So now you're kind of holding the bag and trying to figure out how do you even model these things when we don't have a good replication. Like the biggest one that we have today is Colossus XAI, right, in Tennessee. I think it's Tennessee Louis, I always forget, I was confuse those. But essentially like it's about 500 megawatts by the time it's done it's probably going to be a gigawatt by the time Elon has done getting all the NVIDIA chips, but he's not going to give you access to that data, right? And so how do you model these? That's like the fundamental problem that we.

Stewart Webster   00:11:32

I'm glad you mentioned the GPUs, right? Thinking about the current technology behind AI data centers, it's mostly GPUs and chips. And from my understanding, those can be very energy intensive. They give off a ton of heat. And oftentimes, those chips are overclocked, right? So they're forced to work harder than they were designed to be. So -- and that shortens the life expectancy. And that's exactly what you mentioned there. So in my eyes, it seems like the output is fascinating, right? Where AI is the output that's giving us is fascinating, but the input could be quite inefficient, it seems like. So are there any emerging technologies like liquid cooling or modular data centers or new chip architectures that you could -- that you think could fundamentally change the economics and the risk profile of the data center?

Wish Bakshi   00:12:18

I mean, yes, they are, but I'm going to caveat this by saying nobody is about to disrupt NVIDIA's mode, okay? They are so far ahead. They have like 30,000 employees that are solving one problem or one problem only, which is GPU optimization. They're not going to get disrupted. However, now having caveated that, liquid cooling is coming back -- sorry, it's not coming back. It's coming in a big way. The new racks will mandate that they all have them. So every data center has to come with the liquid cooling capability because copper and HVAC air fans, for example, they simply can't do the job, right? But that doesn't really change the risk profile that much. Everybody trying this new thing called the -- it's not -- sorry, it's not new. ASIC, application-specific integrated circuit chip designs, right? So Cerebrus and Grock are 2 of those. Now fundamentally, what has happened is like compute and memory, they have always been -- it's called the voluman architecture. It's kind of been separated. So compute and memory don't talk to each other. I'm sorry, compute and memory has separated under the voluman architecture. So it's great for like conventional computing. So problem with AI is that it's called the memory wall, like that's what all the technical folks kind of call it, which is that moving data in and out of memory or between chips isn't keeping up. And that's what Cerebrus and Gro, et cetera, these companies are trying to solve for is to say, "Hey, listen, if you just build bigger chip sizes or wafer sizes, can we improve the latency? Yes, you can, and they have had immense improvement. The problem is that they still build for like the transformer architecture, which is the underlying technology to LLMs. And if a new algorithm comes out tomorrow or day after, right, the only one that can keep up with the new algorithm is NVIDIA because their GPUs, they call it graphical processing units, but I call it general processing units or at least the industry is -- because they can be retrofitted for anything very quickly and they can adapt, right? So they're like a one-size-fits-all kind of thing. So yes, like if you look at their memory, they haven't really made massive improvements in memory, but they have made massive improvements in compute. And then industry-wide, like compute has increased about 240x, okay? Memory has only improved between 2 to 8x. The memory is not catching up. So what happens is like the compute and the memory, like it's like -- if you have read like thinking fast and slow, you have your system 1 thinking and your system 2 thinking. Just think about it like that. You have your memory and your compute and then you have different types of memories. It's called DRAM and it's called like RAM. They all have to talk to each other. And the problem is that these GPUs are waiting for -- just like on standby for the memory to pass on all the information between each other. So now you have very, really expensive infrastructure that is sitting around there, waiting for data to be moved between different GPUs and between memory and compute and then updating the weights and processing it between different nodes. Now you asked like are there any emerging technologies I think NVIDIA is on the cutting edge, but rock and Cerebus are also trying to solve this problem. Liquid cooling is coming. There's a whole bunch of things happen, but I would say still NVIDIA is not going to be disrupted.

Stewart Webster   00:15:27

You talked a little bit about the RAM, right, random access memory. What type of RAM or what number would you need? Because you're right, when I look at some of the newest computers now, we're still at 32 gigs of RAM or 64. Maybe if you want to spend a ton of money as a person, it's 128 gigs. But at this point, I thought it would be like a terabyte of RAM. And I really don't see that in the retail space. What -- how much RAM will we need...

Wish Bakshi   00:15:54

Honestly, I mean, this is like literally a trillion dollar question that you're asking right now. So there's types of RAMs, which is like HBM, for example, DRAM. So DRAM is the one that sits on the GPU physically. And then there's a high-bandwidth memory RAM, which is what Hynix, for example, makes Samsung makes, like they have a monopoly around this pretty much or oligopoly. And so how much -- so there's 2 things. It's not just the amount of RAM, it's the speed of the RAM, right? And so how quickly can you speed up the information that flows between the compute and the memory layer. And so I think the size of the RAM is like -- is not going to be the bottleneck. It's like how quickly can you get data in and out of these RAMs. That's like the memory wall, right, because they're not able to kind of solve for that problem. Now keep in mind, like when you look at your computer, you're just looking at your RAM, which is 32 gigabytes, 64 gigabytes, 128. However, when you actually build these racks, right, and these clusters, it becomes one massive conglomerate of massive amount of RAM together. So it can go into terabytes. It can go to hundreds of terabytes. It all depends on how big the model too is, right? So for example, like H100s, you'll need -- like I think you need about 8 H100s to store deep-Seq R1 model. I think that's the ratio, right? So you need -- it's not going to fit in one GPU. You need multiple GPUs, sometimes even clusters. And that's like assuming that you have like 3 or 4 people that are using that model on those 8 GPUs. The real problem starts out is like when you have something like OpenAI, 100 million users. right? And so how do you actually transpose that into these multiple data centers where hundreds of millions of people are acquiring this, and they're not using one model like specifically, right? Like they were using multiple variations of that model as well. I mean it's an engineering feat, to be honest, the more you get into the weeds of these things. But there is no one size fits all how much RAM you're going to need. You're always going to -- you're never going to have enough RAM. You're never going to have enough compute.

Stewart Webster   00:17:54

Well, and you mentioned that's the trillion-dollar question, right? I kind of want to take a deeper dive into the monetary cost behind the data centers. The famous consulting company, McKinsey & Company released a report earlier this year that stated by 2030, data centers are projected to require $6 trillion worldwide to keep pace with the demand for computational power. Then the report goes on to say data centers equipped to handle AI processing loads are projected to require $5.2 trillion in capital expenditures. So these numbers have -- they're massive at this point. We're past the billions at this point. Now we're in the $1 trillion now. And my question for you is, as a finance person, where is the capital coming from? Is it large corporations? Is it investors? And are utilities and governments prepared to fund this at scale? Because when I -- when you go to the $1 trillion mark, that's more of like a government type expenditure. So where is the capital coming from?

Wish Bakshi   00:18:54

I mean it all depends where you want to go with this, right? Are you talking about electricity bills? Or do you want to talk about the pure infra? I mean, from the AI data centers infrastruct, it's actually like one big happy family where they're all doing handshake deals, right? So Microsoft that owns OpenAI has -- I mean, they had an exclusivity deal before when things kind of fell apart between Microsoft and OpenAI. So it's -- I think the capital is just like recycling between these entities. If you just kind of -- if you really pop the hood and you start looking at it, like NVIDIA has invested in a whole bunch of these AI data like they're called the Neo clouds, right? They all invested in these companies, and they're getting a lot of this. I think Nexus is the one that just kind of got a big deal. So it all -- I think the $1 trillion and when Mackenzie and these consulting comes come in, I don't think anybody knows, to be honest. But what I can tell you is the revenue ops of AI data centers, they're in negative margins right now. they're losing money. There is no doubt about it, right? But what they're trying to go for is market share. They're trying to cannibalize the market and try to take as much market share as possible. And so to give you a long-winded answer, on the AI data center infra spend capital, I think most of it is just coming from VC-backed cash. But here's another crazy part, too, right? Like if you think about it, VCs don't want investments that are not growing like at 10x, 20x, 30x, whatever, right? A lot of these AI data centers with the capital investment, like they're not going to grow at that rate because they're cannibalizing their own market. And so it all depends on how you see it and how you want to play it, but they do know that there is immense appetite for this. Now in terms of the electricity bills, right, -- now that's an interesting one because depending on which market you are in, so let's just take 2 examples. Let's just say ERCOT and PJM. ERCOT is Texas, PJM is where Virginia is, which is the data center hub, and I think New Jersey is over there, too. So now if you start looking at those areas, a lot of that investment is actually being paid by the ratepayers, right? So in PJM, for example, they have this thing called the capacity auction. And it's literally saying, "Hey, you pay for tomorrow's reliability today. That's what the capacity market is. What happens is like because of these mechanisms, the rate payers are saying, "Oh, I'm going to have to pay a premium. But guess what, the ratepayers, me and you are not the ones that are capitalizing on AI data center revenue, right? And so that's where it's a very interesting dynamic to watch how PJM, for example, is saying, hey, we're going to cap the capacity auction at a certain price. Well, then what happens really is like those generators that would have provided reliable generation possibly decide to say, I'm not going to play in it. So then suddenly supply goes down, demand is up. Supply-demand is off. What do you -- what do you think happens to energy? Prices on the spot market, they go up. Who gets impacted? Rper. So it all depends on where the value chain you're looking at. But the money on the AI data center side and the utilities, it's coming primarily backed by, I would say, debt financing, if you want to put it crudely as a finance.

Stewart Webster   00:22:11

Yes. And speaking from a financial perspective, we have an industry that's rapidly increasing. And as you mentioned, they have negative earnings, right? So from S&P's perspective, I don't envy the credit analysts for trying to figure out this puzzle here. That's quite puzzling in my eyes. But shifting from Wall Street to Main Street on the consumer side, U.S. electricity bills are spiking. So why I live in New Jersey, in June alone, the households saw a 20% increase in their monthly payment for electricity bills. That's well above inflation. If we're hovering at 3% inflation, that's 7x the inflation rate. So -- and you talked about this a little bit in the previous answer, but what's behind the surge in these energy costs? And how do you explain that to the average household?

Wish Bakshi   00:22:55

Okay. So the best way I can explain this is like a perfect storm. Like New Jersey is actually a really good example of a perfect storm. What happened was about 6 years or 5 years ago -- I mean, sorry, so there's been a big anti-nuclear and anti-fossil fuel sentiment that's been there for the last 10 or 20 years, right? I think possibly -- nuclear was like started in 1980s. And so what happened was in New Jersey, folks -- they retired a bunch of nuclear units. I think they required 1 nuclear unit over there, if I remember -- if memory serves me correctly, but they retired also a whole bunch of gas generation. Now if you just look at it just from a pure numbers perspective, that's a lot of supply coming offline, okay? So that's a supply problem. Then what happens is, while they say, "Hey, listen, PGM, which is the data center alley, Virginia, it's a one big happy family. And last year, when they had the capacity auction, it hit a record high. It went from, I think, $3 billion to $14 billion, right? And then people were upset. So this year, when they had the PJM capacity auction in July, which is basically again, all it is, is when you need power, we will -- we as in the utility, will guarantee to be there and in exchange, we are going to get a guaranteed payment of x million dollars. That's all it is, right? You're paying them in advance today to get paid in -- to make sure that you have electricity and you don't have brownouts in the future. So they'll make the investments. Now because PJM decided to cap the capacity prices this year, a lot of folks thought that this would help, right, as maybe like, hey, prices aren't going to go up again to -- like from $14 billion to $28 billion, for example. But this actually like masks the problem because what happens is like more units, more generation, reliable baseload. Baseload is what you need, right? Because renewables when the wind don't blow and the sun don't shine, you need to find energy somewhere else. Batteries are great, but they are not long duration, meaning they don't last for 5 days. They last for 4 hours, et cetera, and PJM doesn't have a lot of those batteries. But what happens is like because a lot of units came offline, so supply got reduced, prices skyrocketed in this thing called the energy market. So PJM also has an energy market, which is the spot market price, right? So that's where energy is being traded on a day-to-day basis. And essentially, what happens is that people were just like or utilities or generators were like, well, I'm capped in certain areas in the zones. Now here's what's unique about New Jersey, which is the perfect storm. it's on the edge of the PJM grid, okay? That means that not only you're not -- so think about like a highway system and you're parked on the side. So people have to go out of their way to get to you, right? And so the other problem with New Jersey is like population density is there. So it has to import a sizable amount of electricity. So guess what? You don't have a lot of generation, right, in New Jersey, and you don't have a lot of transmission and distribution because that stuff doesn't get built very easily today. On top of that, New Jersey phased out nuclear and gas units. So supply came offline, highways got congested. And then on top of that, the incentive structure for baseload generation went off the map, too. So like I said, like New Jersey literally got the worst of it, which is that you guys were sadly stuck in a perfect storm because you parked on the side and all these other things are playing out.

Stewart Webster   00:26:09

It's interesting you said that about them stopping the liquid natural gas production because actually through the years driving up and down the New Jersey Turnpike, I noticed that those large white double-wall tanks, they started taking them down. So they had excavators actually taking them down. And I'm not sure what they replaced them with, but now I see why they were doing that because as you mentioned, they stopped a lot of the natural gas production. Now you did touch upon nat gas, you talked about wind. So those are all different methods of production for energy, right? When we look at those energy production, that type of energy production across coal, nuclear, solar, wind and natural gas, -- where do you see the highest efficiency per dollar invested? So in layman's term, where do you get the best bang for your buck?

Wish Bakshi   00:26:56

So what I'm about to say is slightly controversial, maybe a lot of controversial. It will make one camp really, really happy. And another camp, it will make them really upset. So before I say this, I think it's important to state that newer technologies are becoming incredibly more efficient to scale, produce and the costs are dropping exponentially, especially like if you look at BEs like energy storage, those technology stacks like have dropped exponentially in prices, right? But to answer your question directly, efficiency per dollar today, there is nothing better. So today, for example, combined cycle power plants, CCGTs are around $3,000 per megawatt hour, okay? And actually, for CCGTs, the costs haven't actually gone down. They've gone up because of procurement issues, GE, Ver Nova, Siemens, they have a -- there's a bottleneck there. Despite this, gas turbines, including simple cycles and natural gas, like recip engines, for example, they are still the single best reliable stack, hands down. And because this is because they provide the grid with baseload and inertia, which we were talking about, right, which is like the heartbeat, the grid has a heartbeat of 60 hertz and you drop down to even 59.95, like you're talking like such a small number, the heartbeat of the grid will struggle. So like the analogy that I usually give is you know how like people can have 1 or 2 red Bulls, the grid can't even have a Coke Zero, right? Because that caffing is going to give it a heart attack. And so what the combined cycle power plants or gas turbines do is they provide that steady state 60 hertz so that if there's too much or too little, the baseload will kick in and absorb it or also kind of like keep the inertia going. Otherwise, you have like this situation that we talked about, which is in Spain and Portugal, for example, like 6 million people were out of power. It's called the Iberian crisis, although it had something to do with the reactive power as well, but inertia was part of it. Now that being said, I am a huge fan of energy storage. So highest efficiency per dollar invested. They're absolutely the best technology to provide for exactly the topic we're discussing, which is AI data centers with -- because they give you this thing, which is harmonics modulation. So again, the last thing you want is too much frequency and voltage fluctuation in these very, very expensive hardware and batteries stabilize that at a millisecond level, right? And they respond very quickly. So if there's something bad about to happen like a trip, batteries will kick in. There's a UPS as well, but batteries kind of stabilize it. The problem is that -- and they have enough power to run the facility when the prices are too high. That's another good arbitrage strategy, which is like, hey, during the evenings, you have this thing called a super peak when people are going home, they're switching on the lights. Power prices are pretty predictable in that 4 to 8 p.m. time because they're going to go up, especially during like peak season, so like super cold, super hot. And so you can use batteries during that time and make sure that your costs are down. And it just works. It doesn't cost like 7 years to produce or even site. They're just perfect. And the problem is like 4-hour duration batteries are just not enough to power the grid, right? Like you're not going to power it for 5 days straight. But what is an interesting trend that people are noticing now that is probably not even out is that their batteries are becoming this thing called like they're behaving like these peakers. So peakers are like simple cycle turbines. They ramp up very quickly, and they ramp up very down very quickly. combined cycle power plants take 10 minutes, depending on the size and all the specs, I'm not going to bore you with it, but the batteries can do that in milliseconds. This is why Elon was able to build this Colossus X AI. And if you look, he's parked a whole bunch of Tesla super packs over there, right? Because -- so they solve like so many problems. The only problem is that you can't make them, use them for 5 days.

Stewart Webster   00:30:50

That's incredible. You're making me realize I should have got a masters in electrical engineering, right, instead of going for masters and finance. I mentioned this before. When I think about data centers, the first thing that comes to mind is the massive amount of physical space required to build them. And on the surface, it seems like a natural opportunity for institutional real estate investors. I'm thinking data centers could fit neatly into a commercial real estate development fund, maybe a real estate investment trust strategy. But is there really a real estate play here? Is it that straightforward? Or is it a little bit more complex than it appears?

Wish Bakshi   00:31:25

Hell, no. I mean I'm so happy you asked this question, too. Like everybody in the inner circle talks about how there were so many CRE commercial real estate folks that kind of decided to play in this field. So like high level, they have learned the hard way that this is not a real estate play, okay? And the reason is because along with siding, you have to figure out fiber optics, you have to figure out water, you have to gas infra, then there's all these other problems that you have with kind of just like making sure the engineering is there, right? But from a financial standpoint, like there's 3 things that you look at. They have a high -- AI data centers are -- like they have a high depreciation rate. They're low revenue, and it's a sunk cost possibly, right? So they require enormous amount of upfront investment. So for example, like chips, $50 billion per gigawatt, right, just in chips, but they depreciate very quickly, like 2 to 4 years, right? Like what commercial real estate is going to be okay with that type of depreciation rate. And it can lead to, and I'm hypothesizing here, but like there's going to be players that are going to be doing a lot of massive write-downs because their revenues -- like the write-downs are going to outpace because they're negative margin. And this creates like a negative cash flow problem, obviously. And then it risks massive sunk costs if AI hype also doesn't deliver profitability, right? So there's like all these compounding effects that people thought that were not there in AI data or they were not even aware. Now 70% of an AI data center is like chips, wiring, cooling, services, et cetera, right? And that's not like concrete that you put up in an office building. Like it's just completely fundamentally different. And it's not even a typical industrial load because of the depreciation rate. We don't know if it's revenue positive and for how long, right? The infrastructure changes every 3 to 4 years. That means that you need to upgrade the facility. We didn't even talk about like the new racks that are going to need the HVDC lines. And to earn a normal return, get this number, to a normal United States data centers would need about $480 billion revenue in 2025, far above the current run rates, right? And that's not even considering the black -- the Rubins NVIDIA chips that are coming out Q4 of 2026, which are like even more pricy. So factor all of that, no, it's not a commercial real estate play.

Stewart Webster   00:33:51

Okay. That's good to know. On the topic of, I guess, building and construction, I wanted to talk about nuclear reactors. And it seems like nuclear reactor building in the U.S. has been extremely slow. From what I saw in 1973, we started building a reactor in Tennessee. That project was halted in 1985, and then it was subsequently completed a couple of decades later. Then we had Vogtle 3, Vogtle 2. Those were completed in 2023 and 2024, and I think they took about 10 years to complete. So correct me if I'm wrong, but it seems like since 1973, the U.S. has only built 3 nuclear reactors. -- whereas in one of your recent publications, you noted that in China, there are currently 29 nuclear reactors under construction. and they're going to generate 33 gigawatts. So in short, what is the U.S. getting wrong in terms of construction? And what is China getting right?

Wish Bakshi   00:34:43

I think 2 things. You said in short, but it's challenging to ask. It's just a sort of macro question in short, but I'll tell you this. So U.S. has 27 design specs for the same nuclear reactor build. France has one, okay? I mean just think about it, how much regulation and rect is that? Crazy, crazy. Like there's no reason for that. Now the other part too is in China, what they're doing is, like, for example, they added 4 gigawatts of nuclear in just in 2024 at 40% lower cost than U.K.'s Hinkley Point C, like 40% lower cost. So they know what they're doing from a perspective of like nuclear, and they know that nuclear is part of the story, if not the story. The third thing that U.S. has done is because of the 1970s, 1980s, like anti-nuclear sentiment, all the talent left, right? If you look at it, like even if you want to build nuclear, like just think about semis, why does Taiwan own the entire value chain with TSMC, et cetera, and then ASML, obviously, in Europe. It's because like a lot of this talent is built in-house through culture, through real-world practice. America stopped building any. So we don't have the talent anymore. They all left all the nuclear physics, all the nuclear experts left. And then obviously, like the infra people who actually build these nuclear facilities also left. So these policies have a very long-term effect. And one of them has been that we need to grow like get this talent because Korea has it, France has it, China has it, U.S. does not.

Stewart Webster   00:36:18

Yes, you're right about the talent. When I think about the university level programs for engineering, right, you have the civil engineering, electrical engineering, industrial engineering. But the closest I've seen to this topic is petroleum engineering, right? I've never seen like you can get a degree in nuclear engineering in the U.S. And like you said, it's really like a talent gap at this point. On the topic of China and geopolitical issues, if China continues to add gigawatt capacity, what does that mean beyond AI, right? So for example, what does that mean for their GDP? What does that mean for defense and military? Or what does that mean for the long-term strategic position relative to the West?

Wish Bakshi   00:36:56

I mean, this is such a macro question. It's hard to predict the future, but I can tell you this. from a strategic standpoint, let's just say that there's a whole thing that's happening between the current administration, not wanting NVIDIA chips to be sold to China. So they have like these lower-tier GPUs that are being sold to them. Huawei may or may not catch up. But guess what, their chips don't have to be that good for power efficiency because they have so much more power. If you look at the trajectory and just Google it on the website, just look at China's added gigawatt capacity versus U.S. and France, et cetera, whatever. It's just like exponential curve. So even if they don't build the best GPUs with the most power efficiency racks, it doesn't matter because they have enough power to go around, right? Now from a GDP perspective, if you buy into the AI productivity game, and another thing, too, is like robotics. -- they are literally leading the field right now, right? And U.S. is kind of in the same tier, but I would say not at par, but not better probably. So that gives them an edge over there, too. It gives them an edge on the EV side. So I think from a long-term strategic position relative to U.S., they are going to become the pioneers if they're not already considered the pioneers, and they are going to take the lead if we don't start building faster, remove the red tape, get the talent in-house and start just kind of like innovating a little bit more. But on a more positive note, I would say that the capital investment and the talent, there's no better place in the world than U.S., right? Like think about what Elon has done, whether we like him or not, with xAI, SpaceX, et cetera, like so -- and a lot of that talent, they flock towards pioneers like that, which you don't see a one-to-one. There's no Elon of China, right? There's a whole bunch of them like Alibaba, et cetera, whatever. But Elon and a lot of these other pioneers, they attract the talent, they attract the capital. It's just a matter of like removing the red tape and opening the doors to say, listen, we're just going to build and then everything will work out. It's not all bad, but China is leading it in capacity, and they are going to overtake if we continue this trajectory. I mean I would tell you this, and this is like, again, very frank, nobody knows what's coming next, okay? Not -- there's no expert knows, nobody can predict it. And to be honest, like a lot of these numbers, et cetera, everybody is just guessing. Like it's literally a ground zero. And I can tell you this that everybody is still trying to figure this out. And it's not all bad. The media kind of has taken and sadly like the super intelligence, AGI jobs. What you're seeing at the enterprise levels right now is that like those jobs -- job losses that they predict, et cetera, they're not happening. because adoption is slower, people are trying to figure out -- here's a very good tangible example, right? You don't want an AI model controlling a power plant, no matter how good it is because you're one mistake away from bringing the plant down. The full automation of AI -- like it's not going to happen at least in the near term. And I also think that from an AI data center story, the capital intensity is humongous. It's probably understated. But there's also this thing called Phantom AI loads. So we don't know where the phantom AI loads are and where the reality is, but they're coming. They might not just be at the rate of speed that people predict.

Stewart Webster   00:40:28

That brings us to the end of part 2 of our conversation with Wish Today, we dug into the forces reshaping the future of energy and AI infrastructure from the capital needed for CapEx to the surge in consumer electricity cost from the role of real estate in the data center boom to the geopolitical stakes as China ramps up power and AI capacity. If Part 1 laid the foundation, Part 2 shows us just how high the stakes really are. Data centers aren't just about technology, they're about grids, households, national strategy and the balance of global power. Thank you to wish for sharing us his insights and helping us see the risks and opportunities from so many angles. If you enjoy this episode, don't forget to like, share and leave a review. Thanks for listening to Masters of Risk. I'm Stuart Webster, and we'll see you next time.