Masters of Risk | Season 3 Ep. 5: AI Data Centers: How Energy Keeps Them Running - Part 1

View Full Transcript

Stewart Webster   00:00:00

 Today, we're exploring the physical backbone of our digital world, AI data centers and the complex energy systems that keep them running. Joining us is Wish Bakshi, a leading voice in digital infrastructure and the energy behind has spent his career focused on the challenges of scaling AI data systems in a world where demand is exploding and where the risks are increasingly physical, geopolitical and deeply strategic. In this episode, we unpack everything from grid stress and financial risk to supply chain pressure, talent shortages and what keeps them up at night about the future of AI data infrastructure. Wish, thank you for joining us today.

Wish Bakshi   00:01:31

Thanks for having me. It's a pleasure.

Stewart Webster   00:01:33

Absolutely. And it should be a really interesting conversation for our audience. Before we get into the complexity of data centers and AI data centers and digital infrastructure, can you tell us a bit about your background? Looking at your CV, you've got enough advanced degrees to rival some of the world's top scientists. So a little bit about your background. And then also what first got you interested in AI systems? And how did that path lead to what you're doing today?

Wish Bakshi   00:01:59

No, I appreciate the kind words. I mean it was a lot of luck, frankly speaking, and a lot of good sponsors and leaders who kind of took me out of the wheel under their wing. One of the best ways I can explain this is like I started off in the ones and zeros world, right? So, I have a background in computer science, statistics, economics, I have 3 degrees and then a 4-degree masters in management of AI from Queens. And so when I sort of started off, I was in the ones and zeros world, focused mostly on computer science and doing a lot of equity algorithms for large shops. And then I kind of parachuted into a power generation company in Canada and Alberta, one of the largest ones.

And over there, it was very, very fortunate. This is almost like a decade ago, where I got to work on combined cycle power plants, right, like which happened to be the talk of town now. And what I did was I built algorithms over there for wind turbines or combined cycle power plants. Perhaps most importantly, forget the math for a minute. I just think about when you work on these physical hard assets, right, with billions of dollars, the real thing that really matters is the hard hats, the blue-collar workers who are on site whose kids are born like just in a 5-mile radius over there. They've been working on these assets for over a decade, sometimes even 2 decades. And they taught me everything, not just about these physical assets, but they taught me about the grid, the electrical engineering component of it, like how you learn very quickly that math is only a very, very small subset of this problem.

So, I got to build some really cool AI systems and AI algorithms to kind of manage these assets. And then under the sponsorship of a great leadership team, one of the VPs on the trading desk kind of came knocking on the door and said, "Hey, why don't I just -- you want to get seconded into commodities trading, right? -- opened a lot of doors again for me, again, a lot of luck. And I got to work with -- on the pit with a lot of the power and gas traders. And over there, I kind of moved from the ones and zeros world to the electrons, right? That's kind of how the story starts. And then over there, I kind of got to learn, okay, well, how do you build algorithms? You fundamentally have to understand the power and the gas market. And so after that, what happened was I built some really cool AI algorithms for commodities trading and learned a lot about risk. We are after all in a risk podcast and how to manage those positions. And then they kind of -- I kind of started off my own shop. And again, a lot of luck and there's a bunch of folks on the -- who sit on a bunch of boards.

They said, "Hey, do you want to come and build some cool stuff. So, they open up more doors for me. So now working in states for large hedge funds, we're building AI systems. The analogy that I kind of give over there is they work -- they get -- there's way smarter people than me, right, who are building the Lamborghini's and Ferraris of the commodities trading world. I'm the person who gets to test drive those Lamborghini and Ferraris, right?

 So, there's folks with double PhDs from Cornell, Stanford, et cetera. I get to test the cars and see if they're built up to spec. And then coming back home to Calgary in Canada, I got -- I'm getting to work on a lot of the upstream. So again, the story goes from 1s and zeros to electrons and now working on the molecules, right? So holistically speaking, across North America, I'm now getting to work on LNG projects and AI data centers in Virginia. So again, a lot of doors open. And when I founded my company, I had no idea where it was going to go. But it's been mostly the sponsors who said, "Hey, we have some hard problems. Do you want to solve it? And here I am.

Stewart Webster   00:05:49

Absolutely. And you talked about some of the financial professionals you work with. I would say, as a financial analyst, I would categorize myself as an end user of AI models, right? And also my clients, too. So whether I'm working with, let's say, an investment banker or a hedge fund manager or an institutional trader, we're primarily concerned about the output of the AI model, right? And we don't necessarily understand what drove that output. So for example, I would as a model, given a certain change in the crude oil price or given the Y percentage change in the price of natural gas, how does that affect the earnings of an energy company, for example. But I don't know exactly what drove that response. So for our listeners who may not live in this world of data centers, can you bring us up to speed on what data centers actually are? And then also, how are they helping humanity?

Wish Bakshi   00:06:41

I love that question. You're kind of going on the first principles component. fundamentally, like the easiest way to see it is that for the first time ever, humans have managed to build machines that turns energy, in this case, electrical power electrons into intelligence. Now we can get all philosophical that this is really intelligence or just pattern recognition. I'm a computer scientist. So I lean towards the former. Regardless, all I have to say is like for the first time in human civilization, we now have the ability to democratize intelligence to not just large enterprises, but everyday people for a flat fee of $20 to $30 a month for ChatGPT or Gemini.

I mean just think about that, right. Like having that amount of knowledge capability kind of at the tip of your fingers literally on your phone, I mean, it's never been done in the -- in human civilization before. But now like let's dive a little bit deeper, right? So, there are 4 pillars to AI data centers. And again, there's a dichotomy between data centers and AI data centers. The data centers that we had traditionally, there were storage layers. A very good example is how you save files in one drive, right, and share point.

So those are going to traditional legacy data centers, and they don't really have peak usage. It's just a storage layer with compute, obviously, but not -- nothing like NVIDIA GPUs or AMD GPUs or GC or anything like that. Now the 4 pillars of AI data centers are quite fundamentally different is that there's a compute layer, there's a storage layer, there's a networking layer and then the support system. So, the way I kind of used to explain this is on the compute layer, which is the processing power, that's the brain. okay? Now on the storage layer, that's your memory.

So just think about the compute layer, pair that with memory. That's basically how these systems are working. But on the networking side, those are your neurons in your brain, right? How these systems kind of talk to each other, how the compute and the storage layer are talking to each other, that's networking. And then the fourth one is your support system, your power, your cooling, your security. Now those things are like your heart, your lungs, your kidneys, right? Like you can't function without them because they keep your mind and your memory and your network neurons kind of running. So that's the analogy I used to go with. But -- and then the complexity kind of rises when all 4 of these have to run in a very, very hyper performance way. So, the next analogy I would give is like, do you know how you write exams, right? You're just sitting. You're not doing much, except that you are writing an exam that's very comprehensive. And you can see that you are on overdrive. -- your mind is cooking, you're heated up. you're obviously like sweating a little bit. Well, why is that happening? It's because you're processing power, your brain and your storage layer, your networking and your support systems are on hyper drive, even though you're not moving. It's very similar to these NVIDIA GPUs is that steady state, 0%, 5%, they're fine.

It's when you start training these AI models on them or doing inference, we can talk about that in a bit. That's when they go to 100% output. And that's when the systems start stressing a lot of the electrical engineering inside a data center and also, obviously, more importantly, on the power grid, which we can talk about in a little bit. You did ask about how does it benefit humanity. I think it hasn't really benefited us on scale. We are in the very, very early innings of AI adoption. I would argue like it's in an infancy. Frankly speaking, the tech world is benefiting greatly from Gen AI and LLMs like talk about the coding because they are originally the creators, right? The tech the Silicon Valley, for example, they're the creators of this technology, and they have high adoption rates. And therefore, the early adopters are only on the valley side or the tech side.

 Outside of it, especially in the area that I come from, which is energy, there's a lot of noise around these agentic AI systems and LLMs, et cetera, but it hasn't really started yet. And I'm -- frankly speaking, I'm an AI engineer working in building these energy AI systems. So, I would love to tell the story of like, yes, these things are tangible and providing value, but that story is going to take some more time. And the reason is not because it doesn't work, it's because you're working on physical assets.

The last thing you want is having an agentic AI system telling a power operator to do something and it hallucinates, makes up stuff and then boom like you have a different problem, right? Now why -- like any time in history when a new tech comes along, people are usually scared. That's another reason why adoption is pretty low outside of just generally speaking, -- this time, the real bottleneck actually is not because of change. It's totally different actually.

It's because there's not enough AI talent to go around, which I'm sure we're going to talk about a little bit, too. Now with respect to benefiting humanity, that is a real thing. I mean, look at the protein folding problem that AlphaFold solved, right? It's allowed us to get one step closer to using powerful tools for structuring -- structure prediction and drug discovery. stuff that used to take us like hundreds years or maybe thousands of years, you can now do it in an hour. Think about the power of the drug discovery, right? So it saved us a lot of time. And then the other thing, maybe like closer to the average folks like me and you, in a few years, we're going to have variables with AI that learn from our heart rates, our HRV, our deep sleep and even blood glucose, it tells when we're about to get sick or overexerting ourselves, right?

And so think about it like we're going to have a custom AI model on our wearable, on our watches, on our phones that's basically going to say, "Hey, every night, I'm going to train on the latest data that I got from you. And next morning, I'm going to be custom fitted just for you. And these wearables are already predicting when you're about to have the next target event. So I think we're in the early innings, but it's coming.

Stewart Webster   00:12:24

Absolutely. That's fascinating stuff. And as you mentioned, we're in the kind of like the infancy stage of AI data centers, but we are scaling faster than almost anything else right now, right? We're experiencing this rapid growth. And it seems like every day, we're hearing about new data centers popping up. And actually, just less than 24 hours ago, Mark Zuckerberg announced that Meta will be building a new 5-gigawatt AI data center. I think it's called Hyperion. And so we're always hearing about the success of building the data centers. But what are some of the operational risks that might not be getting enough attention in terms of physically constructing these centers?

Wish Bakshi   00:12:59

You certainly asked very interesting and hard-hitting questions. And this is a beautiful question, by the way. Because at the end of the day, people can deploy capital, but the physics and the engineering is all that matters at the end of the day. So this is where the fun starts, right? So I'm going to get slightly technical, but keep it at an 80,000 level. And it's just so that the readers can actually understand when they hear these numbers, what do they actually really mean sorry, the listeners. So first off, let's baseline ourselves on what these machines are capable of, right? So listeners can relate. Like NVIDIA H100 GPUs, which are being replaced today with the newer ones that are coming in, they're about 700 watts, okay? Now on average, a washing machine is anywhere from 400 to 1,200 watts. So let's just say that the 700 watts is washing machine. So when I'm talking about 700 watts, 1 H100 GPU, we're not going to name any of that technology. We're just going to use the analogy of a washing machine, okay? Now these GPUs are put into a rack where they can talk to each other. So just think about our physical rack.

I'm grossly simplifying it. But essentially, a bunch of these GPUs are put inside a rack, so they can talk to each other and pass information to each other, right? These are called the weights. That's literally what they're trying to optimize for so that they can learn something. Now these racks that have a bunch of these washing machines in them, they're about 130 kilowatts, the newer ones that are being deployed today. Now that means that a rack now has 200 washing machines, okay? Think about that. So, these racks have about 200 washing machines. And now again, I'm assuming that's 100% power utilization by these GPUs, which is not the case, but it keeps the math simple. Now the newer ones, the Rubins that are going to come out end of next year, they're going to be about 600 kilowatts, okay?

 So, we went from 130 kilowatt to 600. So, we have about a 4x jump. So, we went from 200 washing machines to 1,500 washing machines worth of power in 1 rack, okay? Now by the time you get to a 1-gigawatt facility, right, the reason I'm giving this analogy is because 1 gigawatt is a reference range that everybody talks about. That's about 1.5 million washing machines worth of power, okay? Now you might think 1.5 billion washing machines or this like, yes, that's not a big deal. But that's roughly a nuclear power plant worth of power being sucked by these washing machines, right? When was the last time United States or in North America, we built a nuclear power plant. That's a bal, and that took decades to build, right? Now you asked about operational risk. And the reason I use the washing machine is actually has a very specific reason. So let me ask you this. How many times -- how often do you run your washing machine a week?

Stewart Webster   00:15:47

I would say probably -- so it's a family of 4, probably once a day.

Wish Bakshi   00:15:53

Okay. Once a day, so that's 7 times a week, right? And now think about 1.5 million washing machines. These washing machines are going to run 365 days, 24/7. That's how much you're running it, right? And we all know that these washing machines take a lot of energy. Now imagine -- and again, another complexity that adds it is that these 1.5 million washing machines are literally in about maybe 5 to 10 block radius. So, it's not even spread out, right? It's literally in a 10-block radius. Now imagine what happens when just 50% of these 1.5 million washing machines switch on simultaneously at the same time, right? That's a small city's worth of power being drawn instantaneously.

 It's crazy, right? Because at what point do you run that many washing machines at the same time and switch them all on and then wait for -- just think about the power draw, right, 1.5 million homes pulling that much power at the same time from the grid for 365 24/7. I mean it's a pretty big number. Now imagine what that would do to the inertia of the grid, right? That inertia has -- the grid has a heartbeat; it's called 60 hertz. So when you start pulling this much power instantaneously, the heartbeat gets disrupted. And so that's where the challenges kind of start coming up is like because there is not enough -- like there are no transmission distribution lines, substations that are meant to supply this much power in such a small radius instantaneously. And the opposite is also true.

 So now imagine like you switch off 1.5 billion washing machines at the same time or even half of them. That just means that suddenly the power draw that was being supplied has an instantaneous way to kind of disperse on the grid, and it all has to happen in milliseconds, right? So, you asked about operational risk. I got a little bit technical, but hopefully, with the washing machine analogy, people get it that these problems are -- at the end of the day, they're all manageable, right? The thing I'm trying to highlight is that the response times are not in minutes for these things. They're in milliseconds. So, it's going to take some really like really, really cutting-edge engineering to make this work because when these things switch on, they go 0 to 100 in milliseconds, not in like 5 or 10 minutes. And there are some techniques you can kind of imply, but this is the crux of it.

Stewart Webster   00:18:13

Wow, that's fascinating. So I just want to put that number in perspective for our listeners. So you said it's about equivalent to 1.5 million wash machines and 1 rack. One data center. Okay.

Wish Bakshi   00:18:29

No, that's in 1 data center. So 1 gigawatt Yes. So 1 gigawatt data center. So rack will have, I think I said roughly about 2 -- so current ones have 200 washing machines in 1 rack. -- worth roughly speaking, the new ones, the Rubins that come out next year, that's about 1,500 washing machines in One rack worth of power.

Stewart Webster   00:18:53

Wow, that's insane. So the connection I was trying to make here is I live in the New York metro area, and there are about 8 million households in this area, right? So that's 8 million washing machines if every single household has one washing machine. So just with a few data centers, you're equating to the same amount of power as New York City essentially.

Wish Bakshi   00:19:13

You know what, you actually took it exactly where I was thinking about, and you made it more concrete. But yes, you're talking about like 8 of these data centers in that region is going to be that much, roughly speaking, assuming 100 -- like there's some assumptions I've made just to keep the math simple, but it's roughly in the ballpark of that much, yes, 1.5 million homes.

Stewart Webster   00:19:33

So with the AI data centers consuming so much power, how real is the risk of blackouts and brownouts in these high-density areas, right? And what's coming to mind is the last time we spoke, we were talking about the Iberian blackout, which many people may not have heard of, but that's something that actually happened earlier this year. So when it comes to building new capacity, what are some of the types of energy sources that are being used? And then also, what does the threat look like for blackouts?

Wish Bakshi   00:20:00

Yes. So there's 2 parts to your question, right? Like one is like what is the risk of the blackouts? And then the second question is like what energy sources are stable. Now I want to be careful how I answer this because people are a little bit sensitive around when you kind of start picking the types of energy sources. In the community, it seems like you're picking aside when you're really not, you're just being factual. But for your first part, like how real is the risk of blackouts. So the -- every jurisdiction has its -- they're called the ISOs, right, independent system operators, and they're mandated by FERC. And they do a pretty good job. So like, for example, Virginia, it falls under PJM, Texas falls under ERCOT.

There's a whole -- like California is called CISO, then there's SPP MISO. So these are independent ISOs, and they kind of operate interchange like with each other, but ERCOT is an exception, Texas, which is kind of like islanded with a few HVDC lines going across. But they do a pretty good job of managing these sorts of risks. So -- because they have one mandate and one mandate only, which is grid reliability, right? It's literally their main job. And they do these complex powerful models to make sure these sort of things like brownouts, blackouts don't happen.

This is why like approval rates are so -- they take such a long time. And there's such a big -- like there's such a big queue, right, both on the interconnection side, even on the load gen side, like the whole portfolio. They have to look at it holistically. Now that being said, the risk is definitely not non-zero. One of the challenges is that how do you model something of this magnitude when no one has done it before, right? So for example, I mentioned ERCOT. ERCOT, Texas has about 135 gigawatt worth of grid connection requests right now. That's about 110 to 135 nuclear reactors worth of power that they need to serve their -- mostly the AI data center load. And again, like I mentioned, the last nuclear power plant that we built in the U.S. took 2 decades and it was $15 billion to $20 billion over budget because regulators couldn't make up their mind on the specs of the design and they just -- there was anti-nuclear sentiment around that and all the other funds aesthetics that at the end of the day, shouldn't really have mattered.

 But then, again, that's -- I digress on that topic. But in terms of like blackouts or brownouts, I don't think that risk was going to stick around by the time we solve this problem. But on the IBR side, right, like the inverter-based resources on the renewable side, solar and wind, they don't provide grid inertia right now. So there's 2 types, grid following and grid forming inverters, right, base resources, sorry. And so that grid forming does not exist today. And it does exist, but it hasn't been rolled out at scale. And so that's where the risk of blackouts or brownouts coming is like if you're too heavily reliant on inverter-based resources like solar and wind and then a massive cloud cover shows up over a solar plant that's 500 megawatts, well, that frequency in the grid inertia does suffer, right?

Because suddenly, your power output for a solar plant goes down. And I've seen this happen in certain solar plants when I was modeling them. These are the type of things that kind of like you have enough these plants where the risk kind of comes in. Not to say that solar and wind are bad, they're not. I think every electron is -- needs to be generated if you want to have a success story. But that's where like some of these risks start showing up. Now when it comes to building new capacity, right, in terms of capability, what types of energy sources you asked about. In the short run, there's only 2 technologies that I would be very, very bullish on, and that's nat gas, combined cycle power plants, simple cycle turbines and then obviously, BES, which is batteries, right? In the long run, I'm very bullish on nuclear, nat gas and BES.

These technologies solve some of the most hard-hitting problems that AI data centers come with. Now on the combined cycle and simple cycle plants combined with BES, they're awesome based on what the market is saying. And just to kind of like hit it out of the park, the best part is that U.S. has a ton of gas. And I mean like U.S. exports are 16 Bcf per day of LNG to Europe and other markets. And it's probably going to go up to 25 to 30 Bcf in the next 30 years like -- in the next 3 years. That's a lot of gas that U.S. produces, right? The world needs. And I don't see why U.S. cannot leverage their own production capability to serve some of these combined cycle and simple cycle turbines to basically serve AI data center loads and a lot more actually.

Stewart Webster   00:24:41

So this is a ton of energy usage, right? So I guess, can you map this to the financial cost? Or can you place a dollar monetary amount to these costs that these AI data centers will cost? And what I'm thinking is, of course, you have to build the -- there's a onetime cost associated to building the AI data center, right? But then on top of that, there's maintenance costs or like you mentioned, there's energy usage costs, so on and so forth. So what do the dollar figures actually look like?

Wish Bakshi   00:25:10

Yes, I wish I could give you a concrete answer, but ballparks, the new chips that are coming out, right, the ones that are just being rolled out for 1 gigawatt. So again, 1 gigawatt is about the nuclear power plants worth of power that requirements, just the chips, just the NVIDIA chips, right, they're going to be anywhere from $25 billion to $30 billion, okay? That's the kind of number. Now the newer generations that are coming out, they're probably going to be -- just the chips are going to be closer to about $50 billion for 1 gigawatt, right? Keep in mind, I haven't talked about gas, nat gas. I haven't talked about combined cycle power plant investments in power plants.

I haven't talked about nuclear builds, right? And obviously, probably the most important part is like transmission and distribution, and I haven't even talked about substations, right? So on the chip side, I can tell you this. Now back in the day, it's funny we say that it's maybe like 2 years ago, combined cycle power plants was about like $1,000. Now it's about $3,000. So prices are doubling now. And so people are trying to figure out like, can I even afford it? That's one. And the second part is because these turbines, the GEs, the Siemens and the Vernova, they're so expensive. They're going up in prices because the demand is so high and the lead times are about 7 years. So you talked about like financial risk, right? I would say like, well, if they're not able to deliver these and you are waiting on these chips and they show up, but your plant hasn't become operational, that's a big problem to kind of manage, right? Because again, some of these chips that are being rolled out, they've been delayed. The NVIDIA GP200s, 300s have been delayed by a quarter. So imagine having a power plant that you invested in and suddenly like the GPU chips haven't showed up or maybe that you didn't get all your inventory in, right? So that's where the financial risk component comes in. The other part is like I'm not going to talk about like the traditional risks such as credit, operational, reg risk. There's 2 types of risks that I would talk about, which is kind of unique that blows my mind. The first is on the depreciation side.

The average life cycle of these chips at max performance is about 3 years. And that's because these are Ferraris that are being run like Mario cards. Like literally, we're buying these Ferraris and then we're driving them like Mario carts around and running them to the ground. the harder you run them, the faster they face performance degradation, right? Just again, remember, like if you're running these Ferraris every day and pressing on the gas pedal and just running it fullrottle, there's going to be more wear and tear.

 Now if you're spending about $25 billion to $30 billion in chips, right, you would think that you would want to take care of these Ferraris. But that's not the case, right? Because time to value and time to market of these new models is what all the big 4, big 5 care about. So now it's like your billionaire trust fund uncle, who can always go out and buy a new one, right? So that's what these Google, all the big fan companies, et cetera, all they have to do is like, yes, we got capital. We'll run this Ferrari to the ground. And then we'll buy new ones, and we'll buy the better ones, right? So -- but now if you're leveraging up, like this is where the risk comes in. But if you're leveraging up to pay for that new Ferrari by collateralizing your house, well, you better hope that Ferrari keeps winning the F1 to keep you a float, right?

Which brings me to the second point, which is what the risk that the market is seeing is that these AI start-ups have now started using NVIDIA GPUs as collateral. So private credit is flooding these GPU lending AI start-ups that are using chips as a collateral. So Core, we've raised $9.9 billion, for example, this Fluid stack that borrowed $10 billion against NVIDIA chips. And then the lenders are charging them like not cheap rates, like they're in the teens, right, due to chip depreciation and like they're ballparking at about 1- to 6-year lifespan ranges, but I'm more like 3 years. Now maybe this is going to be like a nothing burger, and I'm just making something out of nothing. But any time there has been a bubble, it's been strongly correlated to collateralized debt compounding until the bubble burns. And maybe this time is different, right, in quotation, maybe we'll see. But that's the kind of risk that I'm looking out for.

Stewart Webster   00:29:30

Wow, so those numbers, I mean, we're clearly in the hundreds of billions of dollars. And as you said, a lot of the risk lies around leverage, right? And when you think about the Great Depression or the Great Recession, a lot of it was around leverage and these financial engineering, right, behind it. So hopefully, we don't go down that route this time. But one thing that caught my...

Wish Bakshi   00:29:54

This time is different, right? The is different.

Stewart Webster   00:29:57

Yes. So one thing that caught my attention is earlier in the discussion, you mentioned a huge amount of the world's compute runs through Virginia, right? And I saw that there are reports that say that more than 70% of global Internet traffic actually flows through Virginia. And some people are actually calling Virginia data center alley. So we have Silicon Valley on the West Coast and then we have data center alley in the Mid-Atlantic on the East Coast. And as someone who lives here on the U.S. East Coast, I find that statistic quite intriguing. But what kind of strategic position does that give the U.S.? And could it be seen as a threat or concern for other nations? I know you're in Alberta. So from a Canadian perspective, is that a threat to other nations around the world that the U.S. has so much data center concentration in one area?

Wish Bakshi   00:30:45

Yes. What's very interesting is like I'm in Alberta, but I'm also in Texas, right? So I split my time between Texas, Silicon Valley and San Fran and then Alberta, Calgary. So you get to see a little bit of everything. So in Texas, you get to see the oil and gas, the energy, the power utilization. But then in the valley, you get to see all the AI innovation. And then Alberta is where everybody wants a slice of the pie, and I think it's well positioned for AI data centers. But you asked a very good question around the risk component. I think just to like baselines ourselves, right? There's about 12,000 data centers worldwide right now.

And I'm not talking about AI data centers. I'm just like talking about the legacy ones, right, and a little bit of AI data centers. The U.S. has about 5,500 of them. So that's about 50% of the global 12,000 data centers that are there, and this is not even the AI ones. About 1,000 of these data centers are responsible for about 80% of the power usage. Just think about that number. So U.S. is about 50% of them. And out of them, 1,000 of these data centers are responsible for about 80% of power usage. Now strategically speaking, like concentrating that much compute and storage in a centralized location is never a good idea, right? But Virginia did some things very interesting back in the 1990s, which is that they made some strategic decisions, which was that one of them being the relocation of the first large Internet peering exchange.

 And so essentially, what happened as a secondary second order effect was that it allowed them to get a lot of infra investment to flow into there, so including power and fiber, the 2 fundamental pillars again. And now a lot of the global traffic, like you said, routes to Virginia, which obviously certain nations would like to avoid. And so what they do, like, for example, China and Europe, they have their own exchanges and fiber optics to achieve that.

 Now when we mentioned Virginia, it's actually not just Virginia, it's specifically Northern Virginia. That's where it's all concentrated in like these 2 areas. And so what's very interesting about that is strategically the reason why Northern Virginia is actually even better is because there is a region that is called Marcellus. And Marcellus around that in Pennsylvania has, I think, the largest gas reserves in the world. And the problem is that they can't get that gas out because of some regulations, et cetera, and environmental considerations. So -- but you will find like most of the AI data centers are kind of trying to go as close to that area so they can get gas out to feed these combined cycle power plants that need gas. So strategically speaking, like U.S. is very well positioned both from a capital infra and then obviously, the fuel, the molecules that need to go into these power generation to feed the AI data centers. In terms of a threat to other countries, I think that's -- I think we have bigger problems in terms of seeing threats with the other nations. But strategically positioned, like U.S. has it all. I think the only part that U.S. is not doing correctly is not building fast enough.

Stewart Webster   00:33:53

So you just gave us some insight on the geographic concentration, right? But then also, I'm thinking about corporate concentration, right? Do you see any concentration risk when it comes to who controls AI data centers from the corporate perspective, right? And just so that our audience understands the power behind these companies, last week, NVIDIA became a $4 trillion company, which is equivalent to the GDP of Japan, right? And just to put that number in further perspective, there are only 3 other countries in the world who have a GDP that's larger than NVIDIA's market cap, right? So that's the U.S., China and Germany. And it's the same story when you think about Microsoft, Amazon, Alphabet, Meta, these huge corporations. So are we too reliant on a handful of companies to power the future of digital infrastructure?

Wish Bakshi   00:34:44

This is, again, a sensitive question. Always, like I said, you asked very interesting questions. I would say the risk is there. This risk is never there. However, that being said, I think the only companies that can truly execute quickly are these big companies that we just listed, right? Because they have the capital and most importantly, they have the talent, right? Now if you kind of wanted to diversify and split it, a very good example or like deregulated, look at Europe. Europe is a very good example of how not to do AI because they passed so many AI regulations that they just killed all innovation. And then now they've started finding U.S. companies to take market share away from them and make money out of it, right? Now is that a good thing or a bad thing? I'm not going to take a position on it, but I can tell you this, Europe is far behind on all things AI related despite having the capability to kind of at least try to catch up, right? Now in terms of the other part, which is the corporation risk and being very, very, let's just say, very pointed, and I'm sure you're going to talk about this, is that if you want to counter the China in terms of their innovation and their throughput, you have to concentrate a lot of this capability with certain few companies that can move really, really fast. And the U.S. tech companies are so well positioned because the China takes a centralized approach, right? They throw capital, they back their companies, et cetera. U.S. is, hey, we'll remove the red tape for you.

Now you guys figure out how to deploy it, right? And the only way you're going to do this is by letting corporations, and I can't believe I'm saying this, but having oligopolies, et cetera, in the AI space to move really fast. And this is why you'll see like Meta, Mary Zuckerbig on a tear of just hiring all this really, really good AI talent from everywhere, right, like OpenAI is like the main target right now. And then just like they're just getting so much talent in-house now. And the idea like they just post Apple's main AI lead, I think the number that's been thrown around is about $200 million, right, all in. And so -- and before that, the rumors are obviously like the scale AI lead getting about $100 million package. Again, these are rumors, so I don't want to confirm it. But this is the kind of consolidation you can do for AI talent for these companies who have that kind of capital. If it wasn't this way, then I think we wouldn't be moving fast enough.

Stewart Webster   00:37:23

Yes. So you're talking about the talent and the hardware engineers and AI engineers who are physically designing these systems, right? I'm just curious, are we lacking hardware engineers? Because when I think about some of the hottest jobs on Wall Street over the past decade or so, it was always the software engineer, right? And when we think about Fortune 500 companies and the ones that have strong recurring revenue, it was always the SaaS companies, Software as a Service, right? So there was a huge focus on software. And are we limiting ourselves on the hardware side? And because of that, do we see any long-term risk in terms of a talent gap with hardware?

Wish Bakshi   00:37:59

You got it. Like this is like literally the core of the problem that we have is we used to build amazing great things, manufacturing used to be the main core -- and we outsourced it a lot. Now I'm not saying outsourcing is not a bad -- it's a bad idea, but there are strategic discussions you have to make. Like for example, just 2 or 3 decades ago, manufacturing used to be the lead employer in all U.S. states. Today, it's hospitals for every state. Like think about that, right? Like first, what does that say about the state of the population? And then second, what does that say about innovation, right?

So that's one. Now second is another data point is Foxconn literally calls Apple the great teacher, right? And why do they do that? It's because Apple, when they set up shop in China, they brought in all their hardware expertise and taught them how to build hardware. And then the Chinese basically perfected the art of scale, scale manufacturing of hardware. And so we passed on a lot of that domain knowledge, a very good example of this capability being lost in the West. is nuclear reactors in the U.S., right? Like I keep going back to this. But because we stopped building nuclear reactors, the talent pool is gone.

They all left to different industries. So now when we want to build things, like we can't even find the talent. And when it comes to AI hardware, it's such a niche area. And you can see that, right, like TSMCs in Taiwan, et cetera. And so the real problem starts happening is around the -- even if you wanted to build this hardware side, we don't have the capability, and we need to kind of bring that onshore and have that strategic perspective. Now on the AI hardware side and the AI systems side, it gets even tougher because a good example would be like -- well, in the case of AI hardware, -- we outsourced it to China, and they basically now have an absolute advantage in that space.

Now for AI algorithms, here's the fun fact, right? You talked about the software in the OS side, the operating system side. Let's call it like loosely software. AI algorithms equates to software. We used to have this advantage, but China has also caught up on that area. And I would say, in a lot of cases, they're even better than us, right? I mean, look at what Deepsea quants did. It's a hedge fund. And they achieved with a fraction of ApenAI's budget. And again, the media doesn't want to talk about this, but they basically took over the LLM leaderboards, right? And today, if you look at the LLM, large language model and GenAI leaderboards, 12 of the top 17 models are from China. That's the stuff that nobody talks about in the media, right? And they're so cheap. So think about those optics. Now people will say the cybersecurity concerns, et cetera, yes, sure. But the idea is like -- and here's another fun fact is they have basically figured out how to not only scale the hardware and optimize it, right, on like inferior NVIDIA chips, but they've also figured out how to do the software and the AI algorithms. And now people would say, "Oh, a lot of these Chinese engineers are hardware to Stanford educated. That is also not true. They're all homegrown. They're all coming from top-tier Chinese universities. So the idea that we lost hardware, but we're still like guarded software, I would say we're losing in both. And we need to change that. And I think what we need to do is like give -- open a lot of doors for our talent. with hands-on experience. So less theoretical, more hands-on, and we need to truly focus on STEM programs now more than ever.

Stewart Webster   00:41:37

Absolutely. I completely agree with that. And my last question, when you think about the future of AI data centers, what keeps you up at night? Is it corporate power that we talked about, geopolitics, talent, regulation? Or is it something no one is really talking about yet?

Wish Bakshi   00:41:52

So if you would have asked me this question like maybe 18 months or 2 years ago, I would never say this. But I would say this today, what I -- what keeps me up is this thing called super intelligence. As computer scientists, like in a nerdy cryptic way, we like to -- and physicists like to call it the singularity. I think what it really means is that when AI becomes self-learning, self-improving, then it becomes very hard for us mere models to predict what comes next. Now I'm not saying this is all bad. I'm saying we don't know. Having built these AI systems and agents, working with reasoning models and just how rapidly they're getting smarter on intelligence and even EQ tests, there is no doubt we are closer than ever to these type of systems now. And here's the thing, right? I'm not worried about these systems and super intelligence. I'm more worried about our reaction to these systems. When humans have faced uncertainty or something we don't understand, we often react irrationally. I mean think of chicken little, right? The sky wasn't falling, but chaos did. So we have to think, are we chicken little or can we keep calm and carry on.

Stewart Webster   00:42:59

From physical infrastructure to geopolitical tension and talent gaps, this conversation shows that the future of our digital world is grounded in some very real, very high-stakes decisions. Wish Bakshi helped us see that data centers aren't just technical facilities, they're strategic assets. And the risk tied to them, whether in energy, supply chain or scale, will shape how resilient our systems and societies really are. Thanks again to wish for joining us. If you enjoy this episode, don't forget to like, share and leave or review. Thanks for listening to Masters of Risk. I'm Stewart Webster, and we'll see you next time.