The energy economics of our civilization’s digital cathedrals

Danny Crichton:

Hey, it's Danny Crichton, and this is the Riskgaming Podcast by Lux Capital. Energy is the fundamental unit of civilization, the key ingredient that powers our way of life. What happens in the energy markets affects all of us, which is just one reason why the rapid rise of artificial intelligence technologies the past few years has brought massive attention to the voracious demand for energy. In response, big tech companies like Microsoft and Google have introduced initiatives to build or reopen nuclear power plants, all in pursuit of subsidizing the energy demands of the data centers that power the training and inference of AI models. My guest today, Mark Mills, sees contradictions and errors in the analysis of energy demand that need urgent correction. AI draws abundant attention, yet 90% of new data center demand is for traditional compute driven by the continued digitalization of our lives.

Cities increasingly are offering economic development incentives for data center construction despite their incredible profitability and lack of job opportunities. Analysts regularly talk about a so-called energy transition. Yet we have almost never given up an energy source. Even today we burn more wood for energy than ever before in human history. Mark's got a ton of thoughts and that's unsurprising. For decades, he has been covering the energy industry and he's currently a distinguished senior fellow at the Texas Public Policy Foundation, the executive director of the National Center for Energy Analytics and a contributing editor of the Manhattan Institute's City Journal. We've got a lot of energy to get through, so let's get started. So Mark, welcome to the program.

Mark Mills:

Good to be here. Let's rock and roll.

Danny Crichton:

Let's rock and roll. So let's dive into the hot subject of the day, and by hot literally the heat that's emanating from the data centers just proliferating all around the world. Your work has focused on this category for years, and the data centers was one of those things that in the cloud infrastructure world over the last 10 years, Google, Amazon Web Services, Microsoft Azure, these were important. They were growing relatively rapidly, but they didn't take on quite the centrality that they have over the last couple of years with the rise of AI training in AI inference. We've seen Sam Altman propose a multi-trillion dollar data center plan. He's looking for massive funding from the Middle East and from anyone who's willing to support that to grow this all out. And so the data center has become this unit of economic development and growth that it has never been before. And I'm just curious at the top of the line, what are your thoughts in the middle of 2024 here as of the recording the show of what's going on in this industry and its implications for economic development?

Mark Mills:

Well, as you know, I've written about this for a long time, studied it for a long time, was involved in the data center industry at an architect engineering firm that is on the board of that designed power systems for data centers. In fact, designed the power system with some of the world's big data centers, the first ones. So what's going on now is that people are discovering the existence of a new infrastructure and that once you build infrastructures at scale, and there are only two kinds of infrastructures, infrastructures that produce energy and infrastructures that consume energy. And there's far more of the latter than the former 'cause there aren't that many ways to produce energy, just than the physics of the world we live in. So we human beings are really good at inventing new ways to consume energy but not so good at inventing new physics to produce energy. So here we are in a world where data centers roughly consume more energy than global aviation. 50 years ago, they consumed essentially nothing. And so the question you would have if you're an energy guy is, where's that going?

And if you're smart, we know the answer is up because the only calculation you have to have your arms around is the rate of increase in demand for bits. If that's greater than the rate of improvement in the energy efficiency per bit, the net effect is rising energy demand. But that's been what we've seen for the last call it half century. If your typical data center today operated at the energy efficiency of a data center of 40 years ago, computing 40 years ago, one data center would consume the entire country's electric supply. Obviously, we'd never have thousands of data centers if that were the case. So the efficiency drove the demand for the data centers because that made it possible to build them at scale. So we have this discovery and the discovery was animated by AI, as you pointed out, but there's a problem with AI as the pinata for energy, demand for data. And the problem with that is that the sudden discovery that data centers are now the single largest source of new electricity demand in the country. Think about that, the single largest... Not EVs. EVs-

Danny Crichton:

Not EVs, yeah.

Mark Mills:

... they're a factor of 10 less on their impact on the electric... at least a factor of 10 less. So the biggest single vector for new electricity demand in the United States are data centers. And in that the world of data centers, 90% of the net new demand comes from conventional compute. Only 10% of it's coming from AI.

Danny Crichton:

Oh, interesting.

Mark Mills:

And 20 years from now, depending on how you make the forecast, everybody has their own forecast, 70% and change of the net new electric demand for data centers will be conventional compute and a third will be AI. Now, this is hard to make a forecast on because AI, and this is maybe an analogy that I think is pretty... analogies are tough, but AI is to compute, conventional compute what jet engines were to conventional aircraft circa 1960. So the development of the Boeing 707, which is the first 1956 or '07 in a commercial service was the first viable commercial jet engine put on a conventional airframe that at that time were all except for the military since all propeller powered. Why did that change the game? There was already a lot of aviation in the 1950s. There's already been 30 years of robust growth in global aviation and personal air travel, but air travel and aviation did a knee in the curve. even though it was growing hugely- to use.

Danny Crichton:

Bigly, yes.

Mark Mills:

... bigly, to use a technical term up until that point, took off. No pun intended, or maybe pun intended. Because what the jet engine did is it enabled far more economic utility for the conventional airframe, which did not change, the same airframe, same aluminum bodies, all those things. Infrastructure, they're still the same. So it democratized conventional air travel. It accelerated it. AI is like a jet engine to conventional compute. So what's going to happen is we're going to need more conventional compute than we would otherwise have needed because AI is data hungry and has to be mediated and managed by conventional compute. All AI systems have conventional compute on both sides of it. It's like an accelerator, if you like, for certain aspects of data as you know. So we're going to accelerate not only the demand for conventional compute to support AI, but as you know, most people who are cognizant I know, AI when it operates roughly speaking, is tenfold more energy intensive per chip or per square foot, however you want to measure it, the conventional compute. So it's a double whammy.

But my point earlier was that you have to realize we're not just getting... AI is not the problem or the primary vector. It's an accelerant for a trend that was already in place. So when I wrote my book a couple of years ago, it came on the market right before ChatGPT became a known word, but my book has a lot of AI in it because everybody in the compute world knew AI was on the cusp of widespread high utility to accelerate the compute world. No one could have predicted exactly which company would become the popular one or would animate the discussion, and there are hundreds of AI tools, thousands in fact, not just ChatGPT or chatbots.

That's a long way of by framing the issue here is I think, and this is, again, why I wrote my book, we are at the end of the beginning. And so we're at the beginning of a new phase equivalent in aviation terms to 1960s for the next 60 years, which is far greater growth, far more economic utility from aviation, far more economic and social impact from aviation than the previous 30 years by a factor of 10. That's where we are. This is pretty cool. So where does it go exactly? Well, there's some things that are hard to predict, some that are easier, but we're at a pivot. But we overuse that word so much that people have a hard time believing when you're actually at a real pivot as opposed to a PR hyperbole pivot. But we're at a pivot.

Danny Crichton:

Let me ask you. There's always been this TikTok, this swing back and forth between centralization and decentralization and compute. So we start with mainframes, we go to personal compute, we can go to the cloud, there's mobile. As we're recording this, Apple's releasing iOS 18 and is releasing the iPhone 16, and it's going to have on-chip inference processing as well as a chip in the cloud. How much of this is going to be driven by data center compute expansion versus on-site, on client expansion? Is there a rebalance towards centralization that's permanent because that's what we've seen in the last couple of years, but do you predict that it continues to go in that way?

Mark Mills:

So George Gilder and I have had this argument, he thinks the era of the data center is essentially over, beyond Google. So he and I disagree that it's over, but I agree with him that there's going to be increasing decentralization. As compute function invariably continues to follow Moore's law, and Moore's law is not done... My first job as a young physicist, 'cause I couldn't get a job as a scientist, was as an engineer designing microprocessors and large-scale integration. So that's a few decades back, I must say. We're not at the end of the physics of Moore's law. So it's insight, let's just say. There's no infinite improvement in anything including computer chips, but we're certainly couple decades away from really hitting some hard limits in the current silicon compute architecture to follow Moore's law. What that means differently is obviously, if my smartphone, which today has roughly 10,000 times the compute power of an IBM mainframe, do I get it 10,000 times better in the next 20 years? Maybe only 1000 times better. Well, okay.

Danny Crichton:

Right.

Mark Mills:

So we've slowed it down by tenfold, but if your smartphone had 1000 times more compute horsepower, you now have data center capacity in your pocket or your purse or hand or whatever pocket. So this is consequential. An awful lot of compute function that is in the cloud will move native to your hands or to devices as devices get smarter. But that doesn't change the fact that the head end the same 1000x is advanced. We know for example, that there are myriad compute functions, AI functions we would love to engage in that require a thousandfold improvement in the efficacy of computing to actually do. We just know that now.

For example, modeling in silico the effects of a new drug on you personally, not on humans at large, but on you personally, if I have enough data in theory about you and I have a digital twin of you, I can do an in silico test that's a preliminary decent indicator, not perfect of whether or not you're a good candidate for that drug. We know that kind of thing is possible now 'cause we're doing things like that. We also know we need a thousandfold increase. So instead of exascale class computing, we need yotta scale class computing. Well, if a yotta scale computer requires one computer, not a data center, one computer requires a gigawatt, nobody's going to build it, right?

Danny Crichton:

Right.

Mark Mills:

Data center, you might 'cause it serves not one person. So I can't have a gigawatt virtual model of me. I need to have maybe a megawatt or 10 megawatt. But the point is they both expand, so the idea that everything gets pulled out of the center. The utility function for so much of compute, which is what a data center is, is a compelling function. It's an architectural nature of engineering. That's like say, well, why doesn't everybody get produce their own water or their own food? You can, but utility functions and utility infrastructures, lower costs, lower impact, lower land use, they're inherently better for society at large, but there's always niches where the utility function doesn't work and those niches will continue to exist, but we'll see both.

And the fact that we'll see both means that the trope that we won't have data centers is silly. The idea that a lot of the current data center leaders will necessarily be, if you're an investor, you're assuming company A today will be the dominant company 20 years from now. Well, maybe, because if the core function of what that company, it was a data center function today migrates to the end user into their handheld or into their tools, then that company's business got just wiped out by the cost trends. And if they don't have another model for what they're going to do with their hardware and their employees, they'll cease to be dominant. That's what's happened through all human history. It'll happen again. We could guess what some of those functions are, but we could also guess what some of the functions won't be. Storing large amounts of data at scale still requires massive physical infrastructure even as we get better and better at it.

Danny Crichton:

One of the questions, when I think about this, years ago when we were moving towards Amazon Lambda and this idea of serverless compute, this idea that you could write a functioning code, throw it into the cloud, it would run, you could run millions of copies of it, whatever you wanted, I always thought that there was going to be a, in the same way you have a memory profile to figure out how much memory your software is using, you would've an energy profiler where you could actually look line by line and say, "Look, this line of code is very non-performant. It's costing a lot of money, uses an immense amount of energy, et cetera." I'm curious to see to what degree the pressure on AI, the algorithms, the training, the inference where the compromises come, where people say, "Look, we get you a better answer, but the energy costs, the data center costs, the on-chip costs are just not worth it."

Or the on-chip is just restrictive and therefore you just don't have enough power to do it, and so the compromise, maybe it's an 80% accurate answer, is enough for you. And if you need a scientific in silico around drug usage in a particular person, yes, that goes to the data center that goes somewhere else, but I'll be curious to see how that grows. The other question I have here around here is around economic development. So we've seen a number of counties, a number of states, even nations saying, "Look, we're going to become a data center leader. We want the tech industry, we want to grow jobs, we want to grow spillover effects, positive externalities, you name it. There's also power infrastructure that might get grown here. Google, Microsoft, others are investing in nuclear power plants and other forms of clean energy, so this might spark green improvement to our city." How much of this is quote, unquote, "hot air" coming out of those nuclear power plants, and how much of that is real?

Mark Mills:

There's a lot of very silly things being said because of category errors. People make category errors based on history and an analogy, and it's human to do that. One of the functions of being human is to put things in categories. Before I get to the category error about all the excitement about being a tech-centric county or state by inviting data centers into the tech community, first, let me make a quick observation about your energy thought. Yeah, you can track the energy use of code, energy use of chips. But I'll tell you, having been involved in this for decades, the energy metric, the reason I got involved in computing and energy is because everything is about energy, the goal in compute was always to make the logic operations faster. And the way you make logic operations faster, it has to do with just physics. You make the switch smaller. A smaller switch can move faster just simplistically, and a smaller switch that moves faster uses less energy. So the energy savings, energy efficiency, compute got better, not because we were chasing energy efficiency per se, but because you wanted speed.

Speed's what matters, and you get it from that efficiency. But there are two sides of the same coin. But what that causes is growth and the demand for the thing, the compute function, which is where you get net energy demand increase. I will tell you, and I will be a cynic, most people developing hardware and software for the compute world that are doing the foundational work could give a rat's patoot about what you described, obsessively tracing what's using less or more energy. What they're doing is trying to get performance. Performance matters. That's how you win. That's how you always win. Now once you've got the performance, can you build it and sell that at a price people will pay? And the cost to operate energy is a feature, not a bug. And having less energy costs matters, but it depends on what all the other ancillary costs are. So this is where the cooling part of data center gets interesting because they all throw off a lot of heat.

But anyway, so you can do those things and people will and are increasingly, this is fine, but all it does is improve the efficiency. And I just want to come back to the classic Jevons paradox. Anything you do, whether it's AI looking at your code to see where you've been, energy inefficient by the AI coder helping the human coder write better code that's more resource efficient, i.e., electron efficient, all you'll do is increase demand for that particular chip and code and history has shown the increase in demand for the chip or code is far greater than the energy savings. And that's just a fact of history. People think that relationship will break sometime in the future. I would take a bet like that with anyone anytime on a two-to-one basis, I'll bet two Krugerrands to their one. We can pick the metric, pick five years, and I guarantee you that I win just because. But back to the overall effect of this.

So businesses now, and I'm partly responsible for this in a minor sense, I stole, I created the word, the phrase digital cathedrals in one of my books, and I used it as a chapter of my book to describe data centers because the first skyscraper, skyscraper was a Woolworth building in Manhattan at the beginning of the 20th century and New York Times when it was finished, I got finished of course in record time, this massive skyscraper, tallest building, habitable building in the world at that time. And for a while until the Empire State Building famously came shortly on its heels and the New York Times called it a cathedral of commerce. And they did that with full knowledge of history because in those days, journalists actually knew their history. So I'm being unkind to journalists. You know your history, you function as an editor and journalist, so you understand history. A lot of journalists don't know much about history, and the reason they use cathedral commerce is because the cathedrals of the Middle Ages were the iconic, not just religious structures of the time, but the economics of building them was staggering.

In today's dollar terms, there are multi-billion dollar edifices. In fact, they would cost that to build today in real dollar terms, but they also, it only functions as a epicenter of commerce for the people who built it in the whole number of jobs that were created that rippled out from these cathedrals. But the town fairs and the town markets all the weekends, buying and selling took place in the grounds around the cathedral as well, sometimes inside of it. Famously in biblical terms, when the money changes were in the temple, those kinds of, and they were the cathedrals was where commerce took place. So here comes New York Times saying skyscrapers of cathedral of commerce, and everybody chased the metric of subsidizing funding, encouraging skyscrapers to be built in their cities because they were cathedrals of commerce. They attracted a high density of high paid people, not just the construction of a skyscraper, but by bringing all those people together at such high density, it created all kinds of economic synergies.

So I called data centers the digital cathedrals of our era because they are the iconic structure. In fact, people, some people genuflect to the digital industry and the icons of power much as people didn't do, but that's a good analogy in broad economic terms because they are where we're spending the most capital. We're spending more building digital cathedrals, more capital building data centers globally than all the world's. The communications systems around it combined than all the world's electric utility are spending combined on infrastructure to produce electricity. So these are epicenters of growth, but so you would want skyscrapers to be in your city or your town, and now you're going to do the same with these cathedrals of commerce of our digital era. The problem is, as you know, except for building them and they get built very quickly, year to two, most usually, usually about a year.

The construction phase employs temporarily say 1000 or 2000 people, but the operations phase employs nobody. The economic benefit from the data center is literally global. It's certainly national and regional, but the buildings themselves in a perfect world have no human beings in it. The fact that there may be 20 human beings monitoring and operating, it is a concession to the fact that robots aren't good enough yet. But as soon as we can make robots operate the data centers, which will happen, there'll be no human beings in a data center. It'll be a sealed building that will have no access for humans any more than you want to be inside of a power plant. They will seal off a nuclear plant and the core of the plant has no human access. That's where data centers are going to go.

So why are we subsidizing the construction of a building that doesn't employ anybody in the long term and worse for subsidizing for industries that are fantastically wealthy? They don't need incentives. I don't mean to be mean to my data center buddies. If I were on their side and you offered me money, I'd take the money, But they don't need the money. They're astonishingly profitable and they don't need the incentives. They're going to build them anyway. They're certainly not creating jobs. And if you feel compelled to spend subsidies to create jobs, then you should either build baseball stadiums, which are also economically inefficient way to create jobs. What you should really be doing is providing the money and infrastructure to build the factories that build the chips that go in the data centers, but that requires even more electricity and energy and natural gas and chemicals than the data centers do. And we are not very friendly to factories by and large in the United States anymore. So I think it's a really bad use of taxpayers capital.

Danny Crichton:

When you think about the... we mentioned earlier that the growth in data centers is outpacing global aviation and it's on trend lines that will outpace a bunch of other industries as well. How does that change the global market for energy? Do countries have to rebuild their grids? Are they rebuilding the infrastructure from ground up? And again, bringing up the idea of nuclear. A bunch of companies and countries are looking at nuclear again that they hadn't looked before because it's one of the only sources that can provide the base level power that is required for these large scale installations.

Mark Mills:

Well, it's not the only source. I'll tell you what's going to happen versus what people think will happen, and another bet I'd be happy to tell. But first, here's what's interesting again, about data centers and about the cloud and the cloud economy is that yes, the infrastructure itself is energy using and it's now become a significant source of electricity use all by itself. But the real benefit from information is productivity, everything about information, the reason we collect information, the reason we do AI, the reason we do all the compute functions we do, whether it's a spreadsheet, whether we're doing zooming, what we're doing is adding economic, it's time and energy efficiency to tasks and creating new modalities and new kinds of tasks. But fundamentally, it's an economic driver.

So if it, in fact, is a good economic driver, and I think it's a very powerful one, information is the most powerful economic driver of all time, full stop. But if we can accelerate the value of information as economic driver, two things happen: We actually consume energy to get the accelerant, the data, but the acceleration of the economy increases wealth, which increases energy and power demands. So it's actually unlike most things. So aviation is similar. Aviation facilitates global commerce, so do cars and they consume energy to do it, and they create an economic benefit doing it. I think the big difference between the two of them is not the energy they consume.

Again, ironically, data center energy consumption is now in the transportation class of energy consumption. The really big difference is that the leverage we get in the economy broadly is going to be higher per unit of capital invested, both because information has higher leverage and also because the net new marginal dollar spent on information infrastructure is far more economically powerful than the last dollar, which is far faster trend than it is with hardware. The net new dollar on a car is a better car, but it's not 10 times better. But your net new dollar spent on a computer can often be 10 times better, certainly twice as good. A new car you buy, it's not twice as good. It's nicer, nicer infotainment, a little more comfortable, but you wouldn't say, "Oh, this is 10 times better." It's new, but it's not economically significantly different.

Danny Crichton:

Right. I usually say it's a capability question, right? And you have the capability of moving from point A to point B. You can literally buy old model T kits, do that and it'll function, it'll work. It's not a great ride.

Mark Mills:

And its economic efficiency is not as good, but it's not profoundly different than driving a Lexus or a Mercedes.

Danny Crichton:

Right.

Mark Mills:

Right? But for compute, we know it's profound. So what that means is I think the world is on the cusp of much faster energy growth than all the modelers are pretending they know about. We're guessing the future. But surely if you're an economist and you are bringing an economic accelerant into the world, which is the global cloud, and you're accelerating the cloud's efficacy and decreasing its costs, where are the modelers that are actually looking for this continuity that will push more growth? More growth means more people will be wealthier and they'll buy things other than compute. They'll buy air conditioners that don't have them. It doesn't really matter whether they buy an electric or gasoline-powered car, they both take energy to fabricate and they both take energy to run. Roughly the same to run, by the way, but far more energy to fabricate the EV than the conventional car. But that doesn't matter. There's lots of energy in the world.

But all of the metrics, bigger houses, bigger cars, having a house at all, having beyond subsistence food, vacations, all those things which are big effect, the dominant energy consumers in the Western world are not for survival, therefore, comforts, conveniences, entertainment, right? We want a bigger house, a better air conditioning, all that. So if we expand, the number of people who have that because of economic productivity, we're going to get faster growth. This is what data centers are doing. They're consuming energy themselves. And on the grand scheme in the United States, while data centers and communications broadly are the single largest source of new demand, they aren't overall the largest source. If you put all the other sources of new demand together, they're still bigger. We're increasing air conditioning, we're increasing use of manufacturing now 'cause we want to repatriate, all these things matter.

But that synergy gets ignored 'cause people look at these things in silos and think, "Oh, well, I'll get a more efficient operation in the supply chain by bringing AI and data centers into that supply chain so I minimize trucks wasting fuel." Yes, they'll do that. They're already doing that and then you pretend you're going to cut energy demand. Well, no, you did cut the energy demand for a static economy. You increased it slightly for the data centers, decreased it more than that for the delivery of the trucks, but now you make trucking cheaper. And more people start to use trucking so the long run trend is you have depressed the rate of growth. That's how an economist would say it. But they have that backwards. You allowed growth in the first place. If you don't reduce the costs of operations, you don't get growth beyond those who are already wealthy enough to afford whatever that is. So we're enabling growth by adding efficiency.

And this is an inverted way of thinking an energy about how it's traditionally thought about in energy planning circles, because energy planners, I think, and most people think of the future statically. They may babble at there's growth, blah, blah, but nobody thinks about discontinuities because they only happen episodically. So you're guessing when the interregnum ends. So we don't have to guess anymore about electric demand. There was a flat load growth in electric growth in the United States for the last 20 years, and that ended over the last year. And so do we think that just a bubble, it's temporary and it's going to be pushed back down with efficiency? I don't think so, and forecasts are tough, but this isn't, again, another one where I'd happily take the bet. Let's look out three years, five years, electric demand is going to keep tracking up unless we have a gross recession, like a depression or we have a really horrible war, and that changes everything.

Danny Crichton:

We keep talking about the future, and to me, it's really interesting because obviously we have so many of these technologies and we know where they're going to go over the next couple of years. There's work being done by mathematicians, AI. We come to know where the theory is. It'll trickle down into practice and empirics over the coming years. When you think about governance strategy, whether that's at the local level, state, regional, national supernational, there's a lot of discussion about regulating AI, but very little about it in terms of energy consumption outside of maybe a climate change circle, whatever. How do these all get connected and should governments be spending more time thinking about it?

Mark Mills:

I'm reflexively not a fan of regulating new technologies to determine outcomes. There's a self-evident spectrum between doing nothing in thinking about safety issues, and the obvious one is to try to figure out how to detect deep fakes in political campaigns. People have been doing fakes of photographs and political campaigns since there've been political campaigns, but there's a distinction with the difference if you have the accuracy and the velocity you get in our cloud world. So how do you detect those and how do you protect... Those are important discussions to have and probably we're going to end up with some forms of regulations or more likely standards that will then result in regulations. But broadly speaking, especially at the political level, especially at the political level, and we don't have a really good idea what this is all going to mean, how big it's going to become, what its impact will be on industries and people.

And most of the existing laws are sufficient to protect people from new technologies, which has been true for most of the history. And what happens is when new technology is introduced from photography itself to the car to aviation, we have all specific standards and classes of regulations for safety. I suspect we'll probably see some safety class of regulations around compute. I think that feels inevitable but constrained to big picture things or constrained is very specific applications like healthcare, for example, where the outcomes can be very serious. But the idea that we should regulate this is offensive to me because we don't know what we're regulating and the people who are doing the regulating really don't know what they're regulating, so let's just say. Not to, again, be unkind to their intelligence. It's just their domain knowledge isn't there, and we don't know where it's going to go fully yet. I have my own opinions and we'll find out.

I think insofar as the implications for energy, what it's done is created the obvious collision between the fiction that has been peddled, that we can have growth in data and stuff insulated from normal demands for energy. I've never believed that was a case. I think it's been largely a hidden issue, and we've also exposed the fiction that we can supply it all for data, never mind the rest of the country with wind and solar. If it were the case that wind and solar plus a battery was the cheapest way to power a data center, you could buy this stuff. Now, every data center could just never connect to the grid and just get big bad batteries and just go ahead. Why are they talking about nuclear power plants? Because at scale for highly reliable power, the only option now not in terms of the future, is natural gas turbines.

I say only 80 to 90% of the net new demand globally is probably going to be met outside of coal using regions by natural gas. By coal region regions I mean Indonesia parts of Asia. So natural gas is the vector because you can build gas turbines quickly. They'll run at 80 to 90% capacity factor, and they're very low cost, extremely high reliability, and they run when you need the power, and they're cheaper despite the fiction that the batteries make it cheaper. They just don't. The arithmetic doesn't work. Physics doesn't work. Nuclear plants can do the job too if we can build them fast enough and cheap enough. What we don't have any evidence yet is that we know how to build small nuclear reactors inexpensively, safely and quickly because very few of them have ever been built.

We built these in the '50s, by the way. Nuclear power plants in the one megawatt, half megawatt electric range were built, a half dozen of them by the Army and they powered remote bases. They did it for, in some cases, for two decades. They got decommissioned because of those particular designs at that time with what we had were more expensive to run than burning diesel and diesel generators, so the all-in costs, maintenance and labor and fuel. But I think that's not true anymore. We're on the cusp of seeing that that's not true anymore. But somebody has to build them at scale. That hasn't happened yet. The really big nukes we do know how to build, so that's inevitable. That's what's going to happen, so I'm delighted that the data center community has embraced nuclear energy again.

Danny Crichton:

Let me close out, you just had testified in front of the Senate Energy and Natural Resources Committee I think back in May of this year. And the very end of your opening remarks, you said that policymakers can no longer entertain the idea of an energy transition, and this is a cause of cloud and AI. Give a little bit of a sense, 'cause obviously we've been hearing about an energy transition for at least 10, 20 years. We're going to migrate from old technologies to new technologies, but you're putting a foot in the sand and saying that's not going to happen.

Mark Mills:

Yeah, I know it's not popular position to take because everybody on both sides of the political aisle genuflecting to the narrative of an energy transition. The debate doesn't so much seem to be around if there is one, but how fast it'll happen. That's essentially how we frame debate. And so let me say two things that are relevant to this, and this first one is easy. Over all human history, without exception, economic growth and the development of new technologies brings new energy demands. It's never been otherwise. So a future with more technology and more people and more wealth will need more energy, not less; not the same, but more. There is zero evidence that we have broken that linkage, zero evidence, and there's no evidence that we know how to break that linkage in the physics of the universe that we live in. But people say it can be through all this magical hand waving and PowerPoint presentations, really silly stuff. It hasn't happened. Not going to happen.

So then what you would look at is how has the world expanded its supply of energy over the decades and centuries and what you'll see in the data. We'll do the recent data. You could do a century or two centuries. We could do the last 20 to 30 years that we've been talking about an energy transition. We begin talking about an energy transition in different words 50 years ago after the Arab oil embargo where the west convinced itself that we were running out of oil and therefore we had a transition away from oil dependency into something else because we were running out of oil. And that's essentially what all the legislation in the United States and Europe has been about since the Energy Policy Act of 1974. It's a really long cycle. What has happened over that time period is the consumption of oil has increased, not decreased, but the supply and consumption of other energy forms has increased. So what I would say is we know two things: There's never been an energy transition in human history, full stop, ever.

Well, and I have to be honest, there's one exception. For, again, historians, it's whale oil for illumination. Whale oil was a primary form of high- quality illumination replacing candles during the age where we horrifically harvested whales for their high-quality biofuel. It burned very clean, high energy density liquid natural hydrocarbon. That era ended because of the development of the synthesis of kerosene from coal; not drilling for oil, but synthesizing kerosene. And if you overlap the graph of the whale production and the arrival of kerosene, they map by exactly that year collapses. The first oil wells weren't drilled for 20 years after the collapse of the oiling age, so there's only been one energy transition. The oldest source of energy for the world other than human labor and animal muscle is burning wood. So today, burning wood in absolute terms, BTUs or tons or cords of wood is roughly the same as it was 200 years ago. We did not transition off wood.

In fact, the consumption of wood for energy today provides the world more than double the amount of energy that all the world's solar panels and wind turbines combined provide. So where did the wood transition? We didn't transition off of wood. We added coal, then we added hydropower, then we added coal and the oil and gas, and we've added nukes. So we've added wind. So what the data show is that whether you pick 20 years, 50 years, 100 years is we keep having additions to the world's energy supply. We change the ratios then it depends on the country, but there's been no absolute decline in the consumption of any energy fuel except we flatlined nukes because everybody got silly. We flatlined largely hydro dams for a while because the number of really useful hydro dam sites that are close to populations are relatively limited and people didn't want to dam up the rest of the world's big rivers.

That's a decision, but it's doesn't mean there's a lot more, but not a lot more. But we didn't decrease it. We didn't transition away from it. We added to it. So the data show no transition, none. These show no transition. At any time period you choose to pick, if you pick the last 20 years where trillions of dollars have been spent in Europe and here to affect an energy transition, there is no energy transition. Europe is using more hydrocarbons today than it did 20 years ago, not fewer, which is in the data. And the consumption of oil per capita, which would be a measure of a transition, is largely unchanged for 40 years or 50 years. Presumption of natural gas per capita is rising and the consumption of coal per capita is declining slightly. I'm talking globally. So the data show no transition.

Then you, "Well, what about the future? We could always do more." Well, yes, we're going to build more windmills and lots more solar panels and lots more electric cars instead of burning gasoline. It's all going to happen because they're all useful now. They're all cheaper than they used to be, but they're not causing the abandonment of hydrocarbons. In fact, the last fact point I would give you is that even in the most ambitious plans to, quote, "decarbonize," which is the language that's driving the transition narrative, the most ambitious plans, which come from the International Energy Agency, their net zero plan, it imagines one thing that's impossible and another thing that's silly. The impossible thing that net zero plan imagines is an absolute reduction in the world's consumption of energy by 2050, even with a bigger world and more people and more wealth.

I'm not buying that, but that's their assumption. And they also imagine that all the things that are the replacements will keep getting cheaper at the rate they've been getting cheaper in the past, which isn't happening. It's not in the data. They're on asymptotes. But with those two assumptions combined in their scenario in 2050, over half of the world's energy still comes from hydrocarbons, coal, oil, and gas. Well, in absolute terms is only... not in terms of percentage terms, it was absolute terms, that's only slightly lower than, it's about the same as today. So it's not a transition away from today's infrastructure. In their world, we'll need less new stuff and all the new stuff will come from the magic unobtainium of you improvements. But we still have a huge infrastructure, essentially the same as today's native oil, gas, and coal. That's not a transition. You could call it a net zero plan if you like. That's their nomenclature. And I'm not arguing the merits of less or more carbon dioxide or fewer hydrocarbon sources, just the fact that that scenario does not get you a transition.

So the word energy transition as a idea that we're genuflecting to is, well, I guess the only word I could use, silly. It isn't happening. It isn't going to happen. It's not going to happen even in the most wildly optimistic and silly scenarios, so why do we keep talking about energy transition? 100% of everything on the planet that's produced or operated uses hydrocarbons, not some, but everything in the supply chains. That's the world we live in. You could not like that and say we ought to transition away from it. But PowerPoint plans hand waving about how all wind and solar with batteries will magically decarbonize the world and eliminate hydrocarbons is really dangerously silly because it'll put the U.S. on a bad economic geopolitical track. I get why no one wants to hear that. And so far, I feel like the kid in the child's fairy tale saying the emperor has no clothes. I know that's what this is. I get it. But that's what it is.

Danny Crichton:

Well, Mark, there may not be an energy transition, but we do have to transition to the end of the show. Mark Mills, thank you so much for joining us.

Mark Mills:

Thanks a lot, Danny.

‍