Casey Handmer on Dwarkesh Patel: Why Solar Beats Gas for AI

Dwarkesh Patel 0:00 Today I'm interviewing Casey Handwer. Casey has worked on a bunch of cool things. Um, Caltech PhD on some gravitational wave black hole gimmick stuff.

Casey Handmer 0:07 Mm-hmm.

Dwarkesh Patel 0:08 Then hyperloop, then the Jet Propulsion Laboratory at NASA, and now he is founder and CEO of Terraform Industries. Casey, welcome.

Casey Handmer 0:16 Thank you. It's great to be here finally.

Dwarkesh Patel 0:18 (laughs) Big picture question I'm interested in, to the extent that AI just ends up being this big industrial race, who can build the most solar panels, who can build the most batteries, who can build the most GPUs and, um, transmission lines and transformers and et cetera, et cetera-

Casey Handmer 0:35 Mm-hmm.

Dwarkesh Patel 0:35 ... this is not what the US is known for, at least in recent decades. This is exactly what China is known for and right, where they have like 20X the amount of yearly solar, uh, manufacturing the US has.

Casey Handmer 0:46 Yep.

Dwarkesh Patel 0:46 Obviously we have export controls right now, but over time, SMIC will catch up to TSMC's leading edge. So, what is the story exactly of how the United States wins this? Like, why does China just not win by default?

Casey Handmer 1:00 Do you think that China is better at capital allocation than the United States? Do you think the Chinese business environment is better for business than in the United States?

Dwarkesh Patel 1:06 Every ... You make these first principles argument about these other industries where they're killing it, but like, doesn't seem to have hampered BYD or CATL.

Casey Handmer 1:12 Well, people say, "Oh, look, they're so much better at building high-speed trains than the United States."

Dwarkesh Patel 1:15 Right.

Casey Handmer 1:16 Right. I would never like hold up a flag saying, "I'm really good at building high-speed trains." Like that is just a sign that you're really bad at capital allocation. Like why would you devote in 2025 so much industrial effort and money, right?

Dwarkesh Patel 1:26 They're devoting a lot to solar over capacity, which in your opinion, is the key to future industrial growth.

Casey Handmer 1:34 I think they might be like accidentally correct.

Dwarkesh Patel 1:35 They call the most important thing correct, right? Which should count for something.

Casey Handmer 1:37 Well, they're in a similar situation to Europe, but unlike the United States. So the United States is, is the luckiest goddamn country on earth because it's surrounded on two sides by oceans and on the other two sides by like friendly allies, right? China is surrounded by 15 countries who are mostly hostile to it, right? With no good like mountain ranges or rivers or anything to really separate them. And then they get all their oil, al- almost all their oil from the Middle East, right? In countries that they don't control, don't have strong diplomatic relationships with, on fleets of oil tankers that they can't defend because their navy doesn't have-

Dwarkesh Patel 2:08 Right.

Casey Handmer 2:08 ... the ability to operate effectively in-

Dwarkesh Patel 2:10 Yeah.

Casey Handmer 2:10 ... the Indian Ocean.

Dwarkesh Patel 2:11 But they're working on this, right? If you, if you get synthetic fuels working at Terraform-

Casey Handmer 2:14 Personally, yes.

Dwarkesh Patel 2:14 ... do- doesn't that asymmetrically help China? Which might be fine. Like that-

Casey Handmer 2:17 It does. It absolutely asymmetrically helps China.

Dwarkesh Patel 2:19 Yeah.

Casey Handmer 2:19 We're not currently working with China. We don't plan to, but like the physics is very obvious. You know, and synthetic fuels have been around for 100 years. Like there are projects in China right now working on synthetic fuels. It would not surprise me if they were thinking pretty seriously about this.

Dwarkesh Patel 2:30 Just to spell out for the audience, if China has all this electricity production, and the bottleneck is that only a third of final energy use in a modern economy comes from electricity, the rest you need gas and whatever to transport things.

Casey Handmer 2:44 Or coal. They use a lot of coal in China.

Dwarkesh Patel 2:45 Right. And what Casey is inventing is the technology to turn that electricity, which only can supply a third of end users right now, into synthetic fuels, which can supply 100% of the electricity-

Casey Handmer 2:57 Yup.

Dwarkesh Patel 2:57 ... your civilization needs. So then China's energy advantage then becomes overwhelming.

Casey Handmer 3:04 This, this technology levels the playing field. It levels the playing field a lot.

Dwarkesh Patel 3:08 Right.

Casey Handmer 3:09 But at the end of the day, you know, China still contains the poorest Chinese people anywhere on earth.

Dwarkesh Patel 3:13 I mean, b- but-

Casey Handmer 3:13 Never, never underestimate the capacity for an autocratic dictatorship to shoot itself in the foot.

Dwarkesh Patel 3:17 I don't know. I, I w- whenever I, I agree that they've obviously m- uh, made bad decisions, but even if you have the poorest Chinese people anywhere in the world, they can still be quite rich. You know, like Singapore's rich or whatever.

Casey Handmer 3:28 Yeah.

Dwarkesh Patel 3:28 Um, also, there's parts of China which actually contain quite rich Chinese people.

Casey Handmer 3:33 Oh, yeah.

Dwarkesh Patel 3:33 So you had to compare not all of China against the US, but Shanghai and Guangdong against the United States. So you can have a part of China that is as big as America and as wealthy as America and as inno- innovative as America.

Casey Handmer 3:44 Like the Indian middle class is larger than the US middle class.

Dwarkesh Patel 3:46 But also it's not, it's nowhere near as wealthy. Whereas in, there are parts of China which are humongous, which are actually as wealthy as the United States, and in many cases, as innovative, et cetera. Um-

Casey Handmer 3:56 Yeah. No, I'm saying don't underestimate it, but at the same time, like, you know, uh, we wanna find the truth here, right?

Dwarkesh Patel 4:01 Right.

Casey Handmer 4:01 And, and the truth is like we should not count the United States out of the battle and just give up. We're very much still in the race now, provided we don't, you know, take extra effort to shoot ourselves in the foot.

Dwarkesh Patel 4:09 So right now, um, we are export controlling chips for the purpose of we want to keep our AI lead or stay in the lead in AI, and we recognize this is a key input in our ability to compete in AI.

Casey Handmer 4:23 Yup.

Dwarkesh Patel 4:23 So we are going to export control China's ability to have these chips.

Casey Handmer 4:26 Mm-hmm.

Dwarkesh Patel 4:26 Um, energy is also a key input in this AI race. And if China wanted to do the converse of what we're doing to them with these cheap imports, what they would do to us is to export control solar and batteries. Um-

Casey Handmer 4:41 It would be asymmetrical. It would hurt them worse than us.

Dwarkesh Patel 4:43 If they did tariffs?

Casey Handmer 4:44 Yeah, well, China obviously depends upon like the US export market for its economic dynamism.

Dwarkesh Patel 4:48 Right.

Casey Handmer 4:48 Right? And, and the United States, it would, like it's gonna hurt both parties to like sever the link. But if you sever the link completely, China's ability to make the, uh, advanced chips right now is like basically not there, whereas the United States can make them. And United States' ability to make solar arrays is embryonic but it's, you know, it's actually not that far behind China's, right? It's maybe five years behind.

Dwarkesh Patel 5:09 If, if we decided we want to produce 100 gigawatts of solar capacity every single year-

Casey Handmer 5:13 We're already on track to do that.

Dwarkesh Patel 5:14 Okay. Is it gonna be as cheap as it is to do in China?

Casey Handmer 5:17 Um, my, my views on this are somewhat different from the mainstream-

Dwarkesh Patel 5:20 Okay.

Casey Handmer 5:21 ... which is great 'cause this is-

Dwarkesh Patel 5:21 (laughs)

Casey Handmer 5:22 ... a podcast. Um, so the mainstream view would say, "Well, China has cheaper labor," which is no longer true, uh, 'cause they compared to Mexico, and it's got lower environmental regulations, uh, which is true, and that it is more business-friendly, which is absolutely crazy. I've never heard, like there's, there's no way you could justify that like your company having to have a, a like, uh, an inspector from the CCP on its board who like harasses you about Xi Jinping every day, like helps you do your business. Um, and, and also, like their rule of law is not great, so like you're constantly having to pay bribes to people in order to stay in business. Right? The idea that United States cannot compete against that with like mostly or fully automated solar panel manufacturing in the United States, which has cheaper natural gas by far, abundant oil, abundant human resources, great financial capacity, you know, world-leading automation, et cetera, et cetera, is crazy. Like we could literally copy paste s- solo manufacturing factories.

Dwarkesh Patel 6:05 How much additional-... solar power capacity do you think we could be putting on that's manufactured in the US by 2028?

Casey Handmer 6:13 This is a good question. When Russia invaded Ukraine, I thought, "Ah, finally, the Europeans will see sense and they'll, like, pull the trigger on, like, 'We need to localize production of solar panels from dirt to the finished module.'" Which is, like-

Dwarkesh Patel 6:24 Right.

Casey Handmer 6:24 ... roughly a four-stage process. Um, they didn't. Uh, so they're still paying Russia, like, a billion dollars a day, uh, for the privilege of being invaded. But at the time I thought they could probably do that in about two years. And I think the United States could probably do that in two years or less if you started today. Like, it's currently 11:00-

Dwarkesh Patel 6:40 (laughs)

Casey Handmer 6:40 ... so, like, we're gonna start cutting checks by noon, um, and, uh, and yeah. I think you could, you could ramp up pretty quickly. Well, a lot of technology already exists here. It's not like it has to be rein- in- invented from scratch. It's mostly just a case of, like, putting in phone call to all the different manufacturers here in Germany and so on and saying, "We need you to 10x the size of your factory starting today. Blank check, go."

Dwarkesh Patel 6:59 Right. I- I guess l- a lot of your predictions seem to be not predictions, but more like if we had World War II levels of motivation, if we had Manhattan level project level intensity around being of a specific thing-

Casey Handmer 7:11 Yeah.

Dwarkesh Patel 7:12 ... how fast could we do it? Um, which is like an Elon... Like, if Elon was running government, how fast could it happen? Versus-

Casey Handmer 7:18 He was for, for a brief period there.

Dwarkesh Patel 7:18 So may- maybe, maybe then we should put it like if Elon ran the government like he ran SpaceX, as opposed to the question of, like, "Okay, what is actually practically likely to happen given that-"

Casey Handmer 7:26 Yes.

Dwarkesh Patel 7:26 "... we are not treating it with-"

Casey Handmer 7:28 Okay.

Dwarkesh Patel 7:28 "... World War II level intensity?"

Casey Handmer 7:30 So if you look at xAI, which Elon is involved in obviously-

Dwarkesh Patel 7:32 (laughs)

Casey Handmer 7:32 ... what are they actually focused on right now? Right? They're focused on the chips, right? Because they understand, like, the key bottleneck is the chips, not the solar power.

Dwarkesh Patel 7:39 Right.

Casey Handmer 7:40 Right? Because even if Trump puts in a 200% tariff on Chinese solar, right, and we're not able to bypass it by, like, Vietnam or something, it's still a bargain. It doesn't matter. Like, if you, if you need solar to run your data center, it's, it doesn't hurt on te- in terms of the overall cost picture. It doesn't matter at all.

Dwarkesh Patel 7:55 Right.

Casey Handmer 7:55 What matters is having the chips at, like, competitive, uh, capabilities per chip and enough of them, you know, installed in your PCBs, in your data centers, hooked up to your liquid cooling, ready to go. Right? And that's actually something that, that, you know, Elon and his companies are great at. It's like figuring out this mass production, semi-automated mass production. They've got this, um, facility in Texas which is making the, um, Starlink receivers-

Dwarkesh Patel 8:15 Mm-hmm.

Casey Handmer 8:15 ... completely automated. But, like, at what point does, like, "Oh, we don't have a solar panel factory," become on the critical path? Right? I very much doubt it's ever gonna be on the critical path.

Dwarkesh Patel 8:23 Mm-hmm.

Casey Handmer 8:23 There's dozens and dozens of manufacturers of solar panels worldwide that are all competing against each other.

Dwarkesh Patel 8:28 So you're a big solar bull.

Casey Handmer 8:30 Yeah.

Dwarkesh Patel 8:31 Um, right now the hyperscalers are making decisions about the data centers that they're building that are gonna be one gigawatt, two gigawatts, in Meta's case five gigawatts, how they're going to be actually powered.

Casey Handmer 8:41 Mm-hmm.

Dwarkesh Patel 8:41 And the people with actual money on the line are choosing natural gas and it's not like they can't see the learning rate and the... I mean, th- they are building things which will be online in '28 or '30 or something.

Casey Handmer 8:52 Mm-hmm.

Dwarkesh Patel 8:53 So why are they wrong and you're right?

Casey Handmer 8:55 I mean, that's their job. They probably know more about it than I do. But no, in, in all seriousness, if you're like xAI right now trying to build Colossus I data center-

Dwarkesh Patel 9:02 Right.

Casey Handmer 9:02 ... in downtown Memphis, right, you wanna get it done super fast. So you're like, "What are all the different things we need? What are the factors of production to build this? We need a building. Well, we don't have time to build a building. We'll buy a building, okay, and we'll adapt it. Uh, we need power, we need thermal cooling." That's stuff you can deliver on a truck, so that's what they did.

Dwarkesh Patel 9:16 Right.

Casey Handmer 9:17 Need access to gas. They had access to gas there. They could tap into a local gas line and actually, if you can tap into a gas line, generally speaking you, you can get enough power. Like, the trans- the energy transmission capacity of, like, your regular gas delivery pipelines is way, way higher than electricity overhead lines and it's easier to upgrade or, or whatever. And so if you're in this situation, uh, right now you say, "Well, you know, are we constrained by our ability to go and rent, uh, gas turbines?" And no, they're not because there was enough available once, maybe twice, right? But at a certain point you realize, well, as, as you grow, you start to touch all these additional constraints. Um, and some of those constraints include gas availability. So there's a lot of chat about doing this in Pennsylvania, for example, where there's quite a lot of stranded gas and parts of Texas. Um, but at the same time, United States is gearing up in its ability to export natural gas overseas, so y- the price will not be infinitely low forever. Um, you, you start to run into constraints around turbine manufacturing rate, around, uh, transformer, you know, uh, transformer production rate, around grid capacity, um, and, and also kind of running into problems where, where the AIs and, and the humans who depend on, like, kind of legacy electricity, uh, production and delivery utilities are kind of competing with each other. We just saw this recent, um, forward auction in, in PJM result in, um, i- in like kind of very high, like unsustainably high prices for, for consumers who depend on cheap electricity to, like, you know, heat then cool their houses-

Dwarkesh Patel 10:35 Mm-hmm.

Casey Handmer 10:35 ... and have general prosperity. So, so if you kind of look far enough in the future you say, "Well, you know, you can just turn up the dial arbitrarily high." Like, you could say, "Well, we're gonna put in a gigawatt a year." Well, we can, we can meet that constraint with gas turbines, right?

Dwarkesh Patel 10:48 Mm-hmm.

Casey Handmer 10:48 We're not gonna run out of natural gas at one gigawatt per year indefinitely. Okay, what if we're doing five gigawatts per year? What if we're doing 50 gigawatts per year? What if we're doing 100 gigawatts per year? Like-

Dwarkesh Patel 10:56 Right.

Casey Handmer 10:57 ... uh, you can, you can just break the situation. (laughs)

Dwarkesh Patel 10:59 Yeah, yeah.

Casey Handmer 10:59 Does that make sense?

Dwarkesh Patel 11:00 Yeah.

Casey Handmer 11:00 Um, uh, not to reach prematurely for analogies, but, but, like, Henry Kaiser set up the, uh, shipyard in, in Richmond just down the road here, uh, in San Francisco, uh, over near Berkeley. Um, and uh, uh, initially making ships for the British and then by the end of the war, four separate shipyards operating in parallel to the point where, like, he was bottlenecked on his supply of steel because steel was rare enough in the war, 'cause everyone was using it-

Dwarkesh Patel 11:21 Right.

Casey Handmer 11:21 ... for different things, that Kaiser Industries went off and built not only a steel mill but also a steel mine, right? They went and started digging rocks out of the ground to turn into ships.

Dwarkesh Patel 11:31 Yeah, yeah.

Casey Handmer 11:31 Right? And it's the same sort of situation you have here where these, these massive industrial verticals... And here I'm quite bullish on xAI in particular because the Elon cinematic universe has just done so much industrial stuff compared to, you know, the Googles and Metas of this world, can kind of reach all the way through down into primary material supply if they need to.

Dwarkesh Patel 11:48 Mm-hmm. The reason that these current plans are being done based on natural gas is that this is a sort of like, um...

Casey Handmer 11:55 Well, PJM has all kinds of different sources of power, right? They have nuclear as well. They have gas, they have coal, all kinds of stuff.

Dwarkesh Patel 12:01 Mm-hmm.

Casey Handmer 12:01 A- and actually, this, this price here is, is probably driven more by the delivery cost growth.... than by the generation cost growth, if that makes sense. So when you pay your utility bill, the cost is sometimes broken down by, like, a delivery cost and a generation cost-

Dwarkesh Patel 12:18 Yeah.

Casey Handmer 12:18 ... uh, sometimes implementation costs and things. And so, the delivery cost is the cost of, you know, basically what it costs the utility to build and maintain all the power lines that connect all the houses-

Dwarkesh Patel 12:27 Right, right.

Casey Handmer 12:28 ... to all the power plants in some gigantic area divided by your marginal u- usage with all kinds of other complicated rules designed to make it fairer. And the problem that we see, and, you know, the reason that PG&E here in California, for example, is perpetually on the brink of bankruptcy is that even though, like, the cost of an additional solar panel or additional wind turbine or additional gas turbine or whatever is relatively cheap, getting that power to your house is actually really expensive.

Dwarkesh Patel 12:53 Right.

Casey Handmer 12:54 Why? Because you've got, you know, generally, like, unionized labor that has to build and maintain power lines in areas that already have built up infrastructure. So you have, like, multiple collisions where there's, like, a power pole on your own street or, like, building a new transmission line-

Dwarkesh Patel 13:08 Mm-hmm.

Casey Handmer 13:08 ... which requires you to, like, eminent domain land, so you're in court for, like, years and years and years-

Dwarkesh Patel 13:12 Right.

Casey Handmer 13:12 ... spending public money litigating against other people who are also spending public money to litigate against you on behalf of other interest groups and so on and so forth. And then you've got wildfires and, and it's just, um, just, like, it's like the poster child for barmorecostocese. And so one of the reasons that, like, we're going to see large scale pruning of these grids is that we just can't afford under our current regulatory regime to maintain.

Dwarkesh Patel 13:33 When you say pruning, you're just like every- everything will just go off grid?

Casey Handmer 13:36 Well, I mean, it's fairly clear to me that for really, like, large captive loads, like our data centers or, or, uh, aluminum refineries or whatever, you're going to have to build your own power plant for them, which is how it used to work. Like, if you, if you had an aluminum plant back in the day, you would be building your own power plant for it as well.

Dwarkesh Patel 13:53 Seems inefficient to have redundant power plants at every single-

Casey Handmer 13:58 Well-

Dwarkesh Patel 13:58 ... industrial-

Casey Handmer 13:58 Let me paint a grand vision for you.

Dwarkesh Patel 14:00 (laughs)

Casey Handmer 14:00 Um, it would seem inefficient, but, but actually, like, if you are sensitive to the cost of power expressed maybe in, like, supply listicity or something like that, you just have to do it. There's no two ways about it. Like, um, is it inefficient for the XAI Colossus Data Center to have its own captive power plant? Which it does, right? On the backs of a bunch of trucks-

Dwarkesh Patel 14:18 Yeah.

Casey Handmer 14:19 ... in their parking lot. No, it's not inefficient. It's-

Dwarkesh Patel 14:22 Right. It i- it might be a special case.

Casey Handmer 14:23 It's, it's, it's the cheapest way for them to, for them to get power.

Dwarkesh Patel 14:25 Okay. But, uh, big picture question, across different kinds of ISOs, like from Texas-

Casey Handmer 14:31 Mm-hmm.

Dwarkesh Patel 14:31 ... to Pennsylvania, to whatever-

Casey Handmer 14:33 Yeah.

Dwarkesh Patel 14:34 ... um, people are building data centers which will not be online for many years-

Casey Handmer 14:37 Mm-hmm.

Dwarkesh Patel 14:38 ... and they're choosing natural gas. What's going on?

Casey Handmer 14:40 Well, I think they, the, you know, we haven't completely exhausted the supply of the turbines relative to GPUs.

Dwarkesh Patel 14:47 And do, do you have some estimate of when we'll run out of ... 'cause we can also make more turbines, right?

Casey Handmer 14:51 Well, I think everything before about 2030 are spoken for at this point. Yeah, you could make more turbines.

Dwarkesh Patel 14:55 Yeah.

Casey Handmer 14:56 Um, the funny thing is, like, it's actually relatively expensive, I think, to, like, spool up additional production of, of these turbines, for example.

Dwarkesh Patel 15:03 Hmm.

Casey Handmer 15:03 So, so actually, here's one, here's one thing you have to grapple with sooner or later. Conventional power generation is a steam engine.

Dwarkesh Patel 15:09 Right.

Casey Handmer 15:09 Right? So you, you have, um, some kind of chemical that you find inside the Earth that is out of chemical equilibrium with the atmosphere-

Dwarkesh Patel 15:16 Mm-hmm.

Casey Handmer 15:16 ... and you burn it, it makes heat. It could be coal, could be gas, oil, whatever.

Dwarkesh Patel 15:19 Okay. You're gonna be the true birds and bees here. (laughs)

Casey Handmer 15:20 Yeah, exactly. And it makes heat, and you boil water, and the water goes through some kind of mechanical contrivance that creates motion-

Dwarkesh Patel 15:26 Right.

Casey Handmer 15:26 ... and then that motion twists a magnet and generates an electrical field-

Dwarkesh Patel 15:30 Mm-hmm.

Casey Handmer 15:30 ... which then pushes electrons down wires, which then push electrons through a series of gates that then approximate thinking. It's kind of, kind of (laughs) complicated.

Dwarkesh Patel 15:38 (laughs)

Casey Handmer 15:38 Um, but, but the key, the key step, uh, in this, which is converting heat into electricity in the most efficient, most common, and this is the same for if it's a nuclear plant or a gas plant, combined cycle plant or, or a coal plant or whatever, is what's called Brayton cycle. Um, uh, jet engine on an aircraft is Brayton cycle as well. A- and just any time you have a Brayton cycle with, like, a bunch of internal spinning at high speed, it's just gonna cost you, like, a bunch of money.

Dwarkesh Patel 16:03 Hmm. Sorry, because it's, it's inherently inefficient or what?

Casey Handmer 16:06 It's just inherently expensive to build.

Dwarkesh Patel 16:07 Okay. Well, what is the cost of s- uh, like, GE makes these 100 megawatt gas turbines, right?

Casey Handmer 16:14 I guess, yeah. I don't actually know what the retail price is. I would suspect that if their prices is, um, flexible, it would have gone up a lot. But I ... If I recall correctly, it's like 35 bucks a megawatt hour is just the flag four cost for-

Dwarkesh Patel 16:25 How m- how much, sorry?

Casey Handmer 16:25 35 bucks a megawatt hour.

Dwarkesh Patel 16:26 Okay.

Casey Handmer 16:27 Just for, just for the Brayton cycle, right? So we're not talking about the fuel, we're not talking about the heat exchangers, we're not talking about, you know, the cooling ponds or any of the, uh, anything like that. Uh, just, just the amortized cost of the high speed, high temperature spinning components is 35 bucks a megawatt hour.

Dwarkesh Patel 16:43 Do you think the hyperscalers are being irrational or do they have some reason not to be

Casey Handmer 16:47 Well, to be clear, they don't care about cost, uh, for cost of power. This is, this seems extremely, like, it's very counterintuitive. So for, like, Grandma Kettle in Pennsylvania, she's very sensitive to electricity cost and we don't really want her to, like, suffer in her retirement-

Dwarkesh Patel 16:59 Right.

Casey Handmer 16:59 ... from unaffordable electricity costs-

Dwarkesh Patel 17:01 Yeah.

Casey Handmer 17:01 ... and, like, having to sit there shivering. Like, that's not the image that we want. At the same time, like, wha- what, what is the, what is the economic value to you of using, I don't know, Claude or Grok or whatever you use, like, on a monthly basis?

Dwarkesh Patel 17:10 A lot, yeah.

Casey Handmer 17:10 But, like, it's obviously much more than the subscription.

Dwarkesh Patel 17:12 Yeah, that's right.

Casey Handmer 17:12 But, like, is it, like, ten times more than the subscription maybe?

Dwarkesh Patel 17:15 Yeah, easily.

Casey Handmer 17:15 Okay, let's say, so let's say the subscription is on the order of 10 bucks. The value is on the order of 100 bucks, no, actually probably more like 100 and 1,000. How much does it cost XAI or Anthropic or whatever to serve your usage?

Dwarkesh Patel 17:25 Yeah, the mar- the marginal variable cost of, like, serving it for, uh ... Like, in, in electricity it's, like, less than 10% of the actual cost of, uh-

Casey Handmer 17:31 Well, their cost of serving it is, is like maybe a buck per million tokens or something like that.

Dwarkesh Patel 17:35 Right.

Casey Handmer 17:35 And the cost of electricity is about 10% of that.

Dwarkesh Patel 17:37 Yeah.

Casey Handmer 17:37 So it's like 10 cents of electricity is generating $1,000 worth of economic value.

Dwarkesh Patel 17:42 Right.

Casey Handmer 17:42 Right? So it's very obvious that Anthropic could be like, "Oh, our electricity cost basis has increased by a factor of 100."

Dwarkesh Patel 17:48 Yes.

Casey Handmer 17:48 And now instead of paying 10 cents on your $100 bill for, like, for power, you're now paying 10 bucks.

Dwarkesh Patel 17:52 Right.

Casey Handmer 17:52 Right? So we're putting up your, your, uh, your subscription up to $110 for an electricity capacity charge, right? And then they could go out and buy turbines for prices that would make your eyes water.

Dwarkesh Patel 18:00 Okay

Dwarkesh Patel 18:01 so, um, then why are we gonna get the solar future? Like, in 2032, we're gonna have hundreds of gigawatts of extra demand for data centers and at that point most of it's coming from solar, and why is that?

Casey Handmer 18:14 Well, you just, there aren't enough turbines being, being manufactured.

Dwarkesh Patel 18:17 Right. But also I think, um, in the early 2000s if you j- uh, we can probably overlay the graph of, like, how many turbines are being manufactured. Like right now we're at a historical-

Casey Handmer 18:26 They've ramped up basically to like the early 2000s rate again.

Dwarkesh Patel 18:28 But I don't know, you gotta make more solar panels as well, right? Like, there will be, uh, supply elasticities for both solar and natural gas. Is there some reason to think that it's worse for the supply chain involved than having a natural gas powered data center than a solar-

Casey Handmer 18:42 Yeah, I do. Um, the learning rate for, for natural gas is nowhere near as steep as solar, which just tells you that it's easier to make solar panels.

Dwarkesh Patel 18:48 Hmm.

Casey Handmer 18:48 Much easier to make solar panels. There are actually very few manufactured products which are easier to make. Like the learning, the learning, the, uh, the right solar coefficient is 43%, so every time we double communal production we get a 43% reduction in cost.

Dwarkesh Patel 18:59 And what is the basis of, um, like, why are we finding 43% worth of things that can be made cheaper or more efficient every single year?

Casey Handmer 19:07 Well, I mean, roughly speaking there's like 10,000 manufacturing process engineers working on this full time.

Dwarkesh Patel 19:13 Right. But that can, that can be true of any process, right? But no other process sees-

Casey Handmer 19:16 Well-

Dwarkesh Patel 19:16 ... uh, the kinds of learning rates-

Casey Handmer 19:17 It's partially true. Um-

Dwarkesh Patel 19:18 ... that solar is seeing?

Casey Handmer 19:19 Yeah, so in order to, like, sustain this over a long period of time you obviously need to have, like, a demand elasticity that exceeds, like, your learning rate.

Dwarkesh Patel 19:26 Yeah.

Casey Handmer 19:26 Right? So- so otherwise you would, after, you know, a couple of booms you would, you would saturate your market at the current price-

Dwarkesh Patel 19:32 Right.

Casey Handmer 19:32 ... and you'd have no additional growth. But in this case, like, you know, every, uh, roughly every two years we're doubling production to a two to, two and a half years doubling production, the price is coming down by a factor of 40, 40 odd percent.

Dwarkesh Patel 19:41 Hmm.

Casey Handmer 19:41 Uh, so like roughly 20%, 15, 20% per year. And then just as a result of that price reduction, demand skyrockets by like probably six times more than that additional marginal capacity, like production capacity increase.

Dwarkesh Patel 19:52 Right.

Casey Handmer 19:52 And this is one point where I'll say actually the pros, so-called pros, are definitely wrong. Conventional wisdom is like, "Oh, solar demand's gonna saturate, like, this, this week. It's gonna, it's gonna saturate." We've got a graph here somewhere that's like, "This year is it. It's never gonna grow anymore." And instead it's just like blasting out the top of the graph. And uh, uh, this conventional wisdom is wrong. I- it is not only the case that solar adoption production price decreases are continuing, they're accelerating and the rate that they're accelerating is still accelerating. Right?

Dwarkesh Patel 20:17 Wait, sorry. The rate that it's accelerating is accelerating?

Casey Handmer 20:20 Yes.

Dwarkesh Patel 20:21 As measured in the total fraction of energy that's coming from solar? Or-

Casey Handmer 20:25 In, in the sense that its fitness for the markets that it is being produced for is increasing over time.

Dwarkesh Patel 20:31 Mm-hmm.

Casey Handmer 20:32 So it is, it's still extremely early. Right?

Dwarkesh Patel 20:34 Okay.

Casey Handmer 20:34 We're still at like the Apple II computer era of-

Dwarkesh Patel 20:37 Wait, wait. Okay, so-

Casey Handmer 20:37 ... of solar.

Dwarkesh Patel 20:37 Backing up, if the story is that the reason solar is getting cheaper is because there's a lot of demand for, um, more solar and that demand can sustain economies of scale or whatever is going on.

Casey Handmer 20:49 Yes. I'm gonna go out on a limb here and, like, agree with Elon Musk on this. Yeah.

Dwarkesh Patel 20:53 Then shouldn't that also be true of gas turbines and transformers and power stations and whatever else that's required for the, um, the non-solar future? 'Cause like even there we're expecting AI to drive up demand for power regardless of the source. So-

Casey Handmer 21:08 Yes.

Dwarkesh Patel 21:08 ... to the extent a story for solar becoming cheaper over time is just that like, well, demand will go up and that will drive efficiencies.

Casey Handmer 21:15 Yeah.

Dwarkesh Patel 21:15 Why- why isn't that true for-

Casey Handmer 21:16 Well, let's say you're a bank and you're trying to decide whether to lend, I don't know, like GE a bunch of money to like-

Dwarkesh Patel 21:20 Right.

Casey Handmer 21:20 ... expand production of, um, of their gas turbines.

Dwarkesh Patel 21:23 Yeah.

Casey Handmer 21:24 You can, you can write them the check today and then they'll start scaling up their factories and then they'll, you know, they'll start to see the benefits out in three or four or five years, right?

Dwarkesh Patel 21:32 Right.

Casey Handmer 21:32 You don't know if the AI bubble will have burst by then. You don't know if China will have invaded Taiwan by then.

Dwarkesh Patel 21:38 Hmm.

Casey Handmer 21:39 You don't know if Siemens or Philips or someone will have out-competed you. Right? You don't know if GE's like major looming structural problems will cause it to be unable to compete-

Dwarkesh Patel 21:47 Right.

Casey Handmer 21:47 ... as it has in the past. You- you also don't know, like, in order to make that money back you have to then operate that plant at that capacity for 20 years. Right? And if I was looking at the same charts as they're looking at right now I'd say, "What are the odds that in 25 years time we can produce gas turbines at a price that is relevant in a world where solar is already at its current price?"

Dwarkesh Patel 22:06 Right.

Casey Handmer 22:06 And batteries are at a price where they're already. Like-

Dwarkesh Patel 22:08 How does-

Casey Handmer 22:08 ... you cannot win.

Dwarkesh Patel 22:09 I feel like there was actually a similar discussion a couple years back, or a year back, uh, when AI people were like, "No, AI is real. This is gonna happen." And then SK hynix, Samsung, et cetera were like, "We're not ramping up HBM production because like, oh, HBM is all, used largely for AI workloads and if this demand doesn't continue then our manufacturing, additional manufacturing capacity for HBM will not have been worth it." And then there was another bottleneck with co-op.

Casey Handmer 22:33 What- what happened after that? Did they end up indeed ramping up their production?

Dwarkesh Patel 22:37 I think so.

Casey Handmer 22:38 Well, so when someone says, "We can't do it. We won't do it no way, no how-"

Dwarkesh Patel 22:41 I mean, not in the US right now.

Casey Handmer 22:42 ... what they're saying is, "Write me a check."

Dwarkesh Patel 22:43 Right.

Casey Handmer 22:44 Right? And they did. And now Samsung's coming on board in the States to build T6 with xAI I think.

Dwarkesh Patel 22:49 Hmm.

Casey Handmer 22:49 So like they all got there in the end.

Dwarkesh Patel 22:51 Maybe- maybe it's worth going into the numbers, right? So right now 43% of US data center power consumption is from natural gas. And basically you think asymptotically that will be like 100% solar? If you go to like 2040-

Casey Handmer 23:05 Yeah. Yeah, so I mean obviously like legacy production coal stuff's gonna retire over time.

Dwarkesh Patel 23:09 Right.

Casey Handmer 23:09 Right? Um, and- and if a gas plant is still making money people will keep operating it. But- but at a certain point, like it is the case right now that operating a coal plant costs more than building a new solar plant, so it's just cheaper to-

Dwarkesh Patel 23:19 So-

Casey Handmer 23:19 ... turn it off.

Dwarkesh Patel 23:19 Yeah, I wanna know what the-

Casey Handmer 23:21 And obviously capacity's gonna increase a lot, so that helps to dilute the existing, um, production rates.

Dwarkesh Patel 23:26 Yeah. And- and also the, um, the amount of use is gonna increase a bunch, right? Like the- the amount of like data center use of energy will just be-

Casey Handmer 23:34 Yeah, for sure.

Dwarkesh Patel 23:35 ... exponentially higher, so-

Casey Handmer 23:36 Yeah.

Dwarkesh Patel 23:36 The- the new stock matters a lot as compared to the existing stocks. Anyway, so I- I wanna know 2027 what fraction is natural gas? 2030, what fraction is, uh, natural gas versus solar?

Casey Handmer 23:46 For- for new load or f- or for, uh, for the whole-

Dwarkesh Patel 23:48 Let's say new load.

Casey Handmer 23:49 For new load. Um-

Dwarkesh Patel 23:50 2035, et cetera. Basically like okay if- if eventually you're right that we'll pave the earth in solar panels to sustain our quadrillions of AI souls, what- what is the, what is the pace of that?

Casey Handmer 24:00 Well, I think the question to ask is like what is the, what is the major constraint on that ramp up?

Dwarkesh Patel 24:04 Right.

Casey Handmer 24:04 Right? And then everything else will just-

Dwarkesh Patel 24:06 Sure.

Casey Handmer 24:06 ... draft in behind. Um, and I suspect that the hardest thing to make will always be the- the silicon, like the, um, the GPUs. Right? So the question is really...... how, how quickly does TSMC ramp up its production of GPUs? And that's a question for you, not for me.

Dwarkesh Patel 24:20 Um, I'll, I'll u- I'll just use some numbers that AI2027 used for their compute forecast, which even if you don't buy their singularity thing, I think they did a reasonably good job with-

Casey Handmer 24:29 Sure.

Dwarkesh Patel 24:29 ... like, crunching the numbers on their compute forecast. And I think they said there's on the order of 10 million H100 equivalents in the world today, and I think they said by 2028, there'd be 100 million.

Casey Handmer 24:39 Mm-hmm.

Dwarkesh Patel 24:40 So basically, 10X more H100 equivalents in the world.

Casey Handmer 24:43 About a kilowatt each, something like that?

Dwarkesh Patel 24:45 Yeah.

Casey Handmer 24:45 Yeah.

Dwarkesh Patel 24:45 Yeah.

Casey Handmer 24:46 Okay, so it's like 100 gigawatts.

Dwarkesh Patel 24:47 Right, okay.

Casey Handmer 24:48 And that's, that's, yeah, I mean, that, that sounds roughly right. Um, you're not the first person to give me a call and ask me about this, I'll put it that way.

Dwarkesh Patel 24:53 Yeah, yeah.

Casey Handmer 24:53 I'm not gonna name names, but like, pretty much all the names you've heard of have given me a call and said, like, "We know that you're a minority voice on the paper that came out recently with, uh, scale micro-grids talking about how you could do 90% solar, 10% gas."

Dwarkesh Patel 25:04 Mm-hmm.

Casey Handmer 25:04 And I said, "You can go all the way, 100% solar." I wrote a blog post about it. And so they always call me up and say, "What about this?" And they're all talking like five gigawatts in the next few years. So, and that's just like 90-plus percent solar for just those. So I think within a few years, we'll probably see that like the majority of new DCs that are going in will be mostly solar.

Dwarkesh Patel 25:22 Within h- sorry, h- how long?

Casey Handmer 25:23 Let's say by, let's see, what's a tw- for 2027, majority of new DCs going in at that point would be mostly solar.

Dwarkesh Patel 25:28 Going in as in like new, um-

Casey Handmer 25:30 As in ground-breaking at that point.

Dwarkesh Patel 25:31 But if you're ground-breaking in '27, you're probably, like, planning it now, right?

Casey Handmer 25:35 Oh, that, that's why they're calling me.

Dwarkesh Patel 25:37 (laughs)

Casey Handmer 25:37 Yeah.

Dwarkesh Patel 25:39 (laughs) Um-

Casey Handmer 25:39 My consulting fees are extremely affordable. (laughs)

Dwarkesh Patel 25:43 (laughs) Um-

Casey Handmer 25:44 But I, really, I, I don't have deep visibility 'cause I'm not in the same room-

Dwarkesh Patel 25:47 Right.

Casey Handmer 25:47 ... with like, with the meta people as to like when we're gonna hit the wall on transformers and when we're gonna hit the wall on, like, just how much, like, municipal, like, peak, peak load-

Dwarkesh Patel 25:54 Right.

Casey Handmer 25:54 ... can we shave off, um, which is the latest thing that's been doing the rounds.

Dwarkesh Patel 25:57 Yeah.

Casey Handmer 25:57 Which it turns out, like, there's a handful of, of places in the United States, and by a handful, I mean, like, literally a handful, um, where, like, there might have used to be an aluminum smelters as a bunch of like latent capacity in the grid, and there's also a bunch of generators on the grid that are notionally turned down and they like operate at, say, 40, 50% capacity factor.

Dwarkesh Patel 26:13 Yeah.

Casey Handmer 26:13 But they max out at about 80% capacity factor 'cause you gotta bring 'em down for maintenance pretty often, right? Especially if they're old. Um, and so they're saying, "Well, you know, we could pay you just to like operate this old coal plant or something at higher capacity. It'll go down this power line to this place where the smelter used to be. We'll set up there and then we promise to, like, curtail when you need the power."

Dwarkesh Patel 26:30 Right.

Casey Handmer 26:31 Which basically means they just have like a massive captive battery plant as well, which is fine. You just buy that and deli- arrives on a truck. The major advantage to doing that over the pure solar play is that the power's already there, so there's no risk there. There's no... And then you don't need a massive amount of land. Like, the problem with the solar approach is that there's no two ways about it. It, it's just, it's a farming operation. You need a huge amount of land.

Dwarkesh Patel 26:52 Yeah.

Casey Handmer 26:52 Right? The total amount of land that you're using, less than 1% is, uh, under batteries, under roads, under data center, uh, structures, et cetera, et cetera. It's mostly solar.

Dwarkesh Patel 27:01 Right, okay. So, uh,

Dwarkesh Patel 27:02 l- l- let's get into what this, um, if you've got a five gigawatt plant you wanna build-

Casey Handmer 27:07 Yes.

Dwarkesh Patel 27:08 ... break down the numbers for me in how much land in terms of solar you need to farm this out. And especially, I was talking to somebody in this space and they said, "Look, the big problem is not, obviously, cost for, the cost of energy for these data centers is a small fraction of the total cost. Most of the cost is going towards chips. So then the issue is just can you, can you make the energy available?" And they were saying even though solar panels themselves you can acquire, the issue is getting that much contiguous land and getting the permitting to interconnect it or whatever the word is, is like apparently a big hassle.

Casey Handmer 27:45 Yep.

Dwarkesh Patel 27:46 Um-

Casey Handmer 27:47 It's kind of a nightmare.

Dwarkesh Patel 27:48 Yeah. And so they're like, well, at that point, is it actually easier than just getting on the grid or, uh, but yeah. What's your, like if you, if you need tens of thousands of acres of solar, where can you do that without a, g- and get like-

Casey Handmer 27:58 Basically in Texas.

Dwarkesh Patel 27:59 Okay.

Casey Handmer 27:59 I mean, uh, well, like, there's this very popular misconception that like there's not enough land to do solar.

Dwarkesh Patel 28:04 Right.

Casey Handmer 28:05 Right? This is garbage. If you've ever flown in an aircraft in the United States and you've ever looked out the window, you'd be like, "Oh, wow, look, there's a lot of land you could put solar on," especially west of like 110.

Dwarkesh Patel 28:15 Yeah.

Casey Handmer 28:15 Right?

Dwarkesh Patel 28:15 Does it need to be flat or no?

Casey Handmer 28:17 No.

Dwarkesh Patel 28:17 Okay.

Casey Handmer 28:18 It doesn't matter.

Dwarkesh Patel 28:18 (laughs)

Casey Handmer 28:20 Um, like, do trees grow on mountain slopes?

Dwarkesh Patel 28:22 (laughs)

Casey Handmer 28:23 So it doesn't matter.

Dwarkesh Patel 28:24 Right.

Casey Handmer 28:24 Um, and the total amount of land, so just for reference, Nevada is something like 80 million acres or something like that.

Dwarkesh Patel 28:29 Yeah.

Casey Handmer 28:29 So just Nevada, which is like 90% federal land-

Dwarkesh Patel 28:31 Yeah.

Casey Handmer 28:31 ... is 80 million acres. And, uh, I would, I would never say that we should, we should sacrifice Nevada to the AI and, uh, and pave the entirety of Nevada from one wall to the other. But I just saw a bunch of things in my feed the h- last couple of days that like, you know, Vegas is, is falling apart and-

Dwarkesh Patel 28:46 Really?

Casey Handmer 28:47 Well, like, no one, like, the Boomers are retiring-

Dwarkesh Patel 28:48 No one's there.

Casey Handmer 28:48 ... no one goes there anymore.

Dwarkesh Patel 28:50 Yeah.

Casey Handmer 28:50 Um, people would go to see like the-

Dwarkesh Patel 28:51 The solar

Narrator 28:52 (laughs)

Casey Handmer 28:52 ... the 100 million acres of solar.

Dwarkesh Patel 28:53 The massive solar farm?

Casey Handmer 28:53 Yeah, okay. So five gigawatt plant. So-

Dwarkesh Patel 28:55 So even if you did it in Nevada...

Casey Handmer 28:56 We can do it anywhere. You can do it anywhere you can find land, right?

Dwarkesh Patel 28:59 Yeah.

Casey Handmer 28:59 People say, "Well, you can't do this in Europe because Europe doesn't have solar power." Europe has solar power. I've been to Europe in summer.

Dwarkesh Patel 29:04 (laughs)

Casey Handmer 29:04 It's like sunny for like 20 hours of the day.

Dwarkesh Patel 29:05 Right.

Casey Handmer 29:06 It's a bit seasonal, right? But that's not a big deal.

Dwarkesh Patel 29:08 Wait, I mean, it is if, 'cause y- because energy is a small fraction of the cost, you care more about making sure the chips are running all the time, right?

Casey Handmer 29:14 Yeah, so in practice what happens is let's say Europe hypothetically, like, awakens from its slumber and decides it wants to participate in AI. Okay, I hope it does. Uh, they say, "Well, we're gonna have to put, you know, 100 gigawatts down as solar at some point to build these data centers." It'll most likely be in Southern Europe. Spain is not particularly heavily populated. That's a great place to start. So we put in 100 gigawatts of solar, uh, solar data center in Spain. Um, and then in order to achieve, you know, basically if you're spending like AI hyper-scaling money on your-

Dwarkesh Patel 29:43 Mm-hmm.

Casey Handmer 29:43 ... GPUs, you wanna have like four nines uptime in order to-

Dwarkesh Patel 29:46 Right.

Casey Handmer 29:46 ... maximize your like tokens per dollar spent.

Dwarkesh Patel 29:49 Yep.

Casey Handmer 29:49 Okay? Uh, tokens per dollar spent on the entire project, not just on that. Uh, this is a very subtle point. I can go into vast detail on it later on maybe. Um, but let's just say you need four nines uptime. In order to achieve four nines uptime in like the middle of winter, you need to have a, a lot of solar overbuilt, right? Is solar overbuilt a bad thing? No. Is the fact that we produce 40% more food than we need a bad thing? No.

Dwarkesh Patel 30:09 Right.

Casey Handmer 30:09 It's much better than producing 40% less than we need. Okay? And it just means that effectively you have a giant captive power plant attached to a data center that-... 99.9% of the time produces more power than it needs, and 99% of the time produces much more power than it needs. And it can now actually be the source of power that the local utility, instead of being like, "Naughty, naughty data center, you must disconnect, uh, when we tell you to," they say, "Hey, data center, I notice you've got like a bunch of power you're not using 360 days of the year. Would you mind ever so much if we threw a power cable over the wall and we powered our entire town off your spare power at like essentially zero marginal cost-"

Dwarkesh Patel 30:47 Right.

Casey Handmer 30:47 "... plus, you know, whatever residential batteries that we need-"

Dwarkesh Patel 30:50 Yeah.

Casey Handmer 30:50 "... in addition to local power supply?"

Dwarkesh Patel 30:51 Yeah. Uh, Brian Potter had a good analogy in his blog post about this where he's like, "I don't know. My MacBook has a terabyte of storage."

Casey Handmer 30:58 Mm-hmm.

Dwarkesh Patel 30:58 "And I use like 100 gigabytes and I just got the terabyte version because it's cheap enough and I might need it at some point that it's worth it." And so you're saying solar will get- gets so cheap that it's- it's the way we'll treat hard drive space, is they get a bunch of excess, um-

Casey Handmer 31:12 Yeah. But also like the- the- the market will be made at the place like the- the new- the new marginal like consumption and production, right? So like all the people-

Dwarkesh Patel 31:24 Right.

Casey Handmer 31:24 ... who are working this space right now are like, "Oh, I'm in, I'm in the business of delivering power, storing power. I'm going to serve the AI market-"

Dwarkesh Patel 31:31 Yeah.

Casey Handmer 31:31 "... because that's where all the growth is occurring."

Dwarkesh Patel 31:33 Right, right.

Casey Handmer 31:33 That's where all of US GDP growth is occurring-

Dwarkesh Patel 31:35 Right.

Casey Handmer 31:35 ... right now. Like that's basically-

Dwarkesh Patel 31:36 So, but, uh, y- y- I guess you didn't answer the question of, yes, theoretically we could do this, but h- is it gonna be possible to get the permitting to have tens of thousands of acres of contiguous land?

Casey Handmer 31:47 It doesn't need to be contiguous. Well, I mean, it helps if it's contiguous. It doesn't need to be convex. Um, so this is, you know, you can have a- a bit over here and a bit over there and you can wire-

Dwarkesh Patel 31:55 Mm-hmm.

Casey Handmer 31:55 ... them together relatively easily. In fact, in the limit, in the limit you have fields upon fields of solar arrays with-

Dwarkesh Patel 32:02 Tell me your dream case here (laughs) .

Casey Handmer 32:02 Okay. Fields just- just solar arrays as far as the eye can see.

Dwarkesh Patel 32:06 (laughs)

Casey Handmer 32:06 A- a- and then within the solar arrays, roughly in the middle of them, you have your- your batteries and your-

Dwarkesh Patel 32:11 I play Factorio. I remember this (laughs) .

Casey Handmer 32:12 Okay, good. I think it's-

Dwarkesh Patel 32:13 I remember this optimal layout of like batteries and-

Casey Handmer 32:16 Yeah.

Dwarkesh Patel 32:16 ... solar.

Casey Handmer 32:16 So you've got your batteries and you've got your data centers. So in terms of ground, like floor area, you know, it's roughly, let's say, 10%, um, racks, 10% access to the racks, um, maybe like 50% batteries stacked up on top of each other and the rest is cooling, something like that in terms of what- what sits in the- in the like- the centralized node, right?

Dwarkesh Patel 32:35 Mm-hmm.

Casey Handmer 32:35 And that could be 100 megawatts or it could be 10 gigawatts depending on how you want to scale this. But then all you need to connect that to the outside world is like an optical fiber, right? An optical fiber cable, which you can string up on poles, you can run underground. You could even use microwave links if you really wanted to. You could use Starlink if you really wanted to.

Dwarkesh Patel 32:53 Mm-hmm.

Casey Handmer 32:53 I don't know what the... But Starlink wouldn't be fast enough. I'm not sure if it's like capacity is high enough.

Dwarkesh Patel 32:57 Mm-hmm.

Casey Handmer 32:57 You could use laser links if you really needed to. And that's- that's it. It's like this completely self-contained world of like-

Dwarkesh Patel 33:04 Because it's off grid.

Casey Handmer 33:05 ... yeah, of computation that occurs off grid, like on private land somewhere in the backwoods of Texas where like no one lives and no one will ever live-

Dwarkesh Patel 33:11 Yeah.

Casey Handmer 33:11 ... because it's completely inhospitable to humans.

Dwarkesh Patel 33:12 In- in terms of the ratios, it's o- one trend that was impressed upon me is that the power density of racks is increasing a lot-

Casey Handmer 33:20 Yes.

Dwarkesh Patel 33:20 ... as the fLOPS per GPU are increasing. So we've, um-

Casey Handmer 33:23 It's like a megawatt per rack is what they're heading to now, which just seems bananas to me.

Dwarkesh Patel 33:27 I think it was even more than that, but yeah, yeah.

Casey Handmer 33:31 Okay.

Dwarkesh Patel 33:31 It's, yeah.

Casey Handmer 33:32 So let's- let's get concrete here for a second. You got... Let's say you got one rack and that's one megawatt and I'll leave the cooling to like someone who specializes in air conditioners-

Dwarkesh Patel 33:38 Right.

Casey Handmer 33:38 ... but it's- it's basically throw air conditioners at the problem. And then you have batteries. So in order to get four nines of uptime on this, you need... In- in South Texas, you actually need less than this, but let's just say 24 hours worth of-

Dwarkesh Patel 33:48 Mm-hmm.

Casey Handmer 33:48 ... of battery storage, right? 'Cause that means it'll get you through two of the... two- two bad nights in a row basically. Um, and, uh, and actually it turns out that you can significantly decrease power consumption with a very small reduction in overall compute. So if you- if you've got like three really bad days in a row or something, you can actually just like, you can dial back your power usage quite a lot without compromising your inference or- or, um, or training. Okay, so you've got say a, a Tesla Powerwall, something like four megawatt hours. So one- one megawatt rack and then six Tesla Megapacks, each of which is roughly one truckload worth of stuff. So like one truckload worth of rack and then like six truckloads worth of batteries. Uh, and then in order to operate this at an average power of- of one megawatt, your solar arrays in Texas would be something like 25% utilization. So on average, you know, if- if the sun came up every day and the day was the same length all the time, you would need four megawatts of solar arrays, which is about four acres of land. But in practice, because you're aiming for four nines instead of like one nine, uh, you need an overbuild about two and a half X. So you've got about 10 acres of solar. Uh, so 10 acres of solar, uh, six truckloads of batteries, one truckload of- of, um, of data center and- and some cooling stuff. And that's-

Dwarkesh Patel 34:57 For how- how- how- for how big of a data center?

Casey Handmer 34:59 One- one megawatt. That's just a one megawatt, right?

Dwarkesh Patel 35:01 Okay.

Casey Handmer 35:01 So 10 acres, one megawatt-

Dwarkesh Patel 35:02 Yeah.

Casey Handmer 35:03 ... uh, kind of situation at, um, at four nines. Yeah. So if you want five gigawatts and that's 5,000 times-

Dwarkesh Patel 35:08 50- 50,000 acres.

Casey Handmer 35:09 ... times 10, so 50,000 acres. And- and actually at larger scale you can get, uh, you can probably cut all those numbers down by 10, 20%, but like on that order. And like 50,000 acres sounds like a lot.

Dwarkesh Patel 35:17 It- it- it does sound like a lot.

Casey Handmer 35:19 Right?

Dwarkesh Patel 35:19 Is it not? (laughs)

Casey Handmer 35:19 The amount of land, the amount of land put aside for Oak Ridge was 100,000 acres. The amount of land put aside for Hanford was about 100,000 acres.

Dwarkesh Patel 35:25 What's Hanford?

Casey Handmer 35:26 Hanford was where they made plutonium in the- in the- in the Manhattan Project.

Dwarkesh Patel 35:28 Okay.

Casey Handmer 35:29 Yeah.

Dwarkesh Patel 35:29 But I don't know how big that was (laughs) . Was it like-

Casey Handmer 35:31 It was about- about 100,000 acres.

Dwarkesh Patel 35:32 Okay.

Casey Handmer 35:32 I mean-

Dwarkesh Patel 35:32 But I mean, is it like...

Casey Handmer 35:34 Well, it was-

Dwarkesh Patel 35:34 Is it- was it like, "Oh, this is so small," and then you're like, "Oh, but it's 100,000 acres," or... (laughs)

Casey Handmer 35:38 Well, so the reason they... And it's still largely unpopulated now-

Dwarkesh Patel 35:40 Right.

Casey Handmer 35:40 ... 'cause it's a national laboratory. But the reason they did that was they thought, "Oh, we're gonna need four piles to produce plutonium." These are not, um, these are not nuclear reactors that produce excess thermal energy.

Dwarkesh Patel 35:50 Yeah.

Casey Handmer 35:50 So you can't actually make nuclear power with them, but you're making plutonium with them. Then in the end, they only needed two, I think, and they wanted them spaced out 'cause they felt like they might just like spontaneously like explode.

Dwarkesh Patel 35:58 Right.

Casey Handmer 35:58 And a bunch of other facilities and plants and stuff as well.

Dwarkesh Patel 36:00 Uh, uh, Austin Vernon had an interesting blog post where he said that if, uh, if you allow for... if you have like diesel generators or something which can-... take over the generation for, like 10% of the generation during these, you know, with the winter or something-

Casey Handmer 36:16 Yeah, yeah.

Dwarkesh Patel 36:17 ... uh, then you can have a 60% reduction in the amount of solar panels you need to install because you don't need to have, you don't need to plan for that contingency.

Casey Handmer 36:25 Yeah, yeah. So I mean, basically there's a balance here. So this is not a very complicated optimization problem. Like, for people who do optimization problems for fun, this is how you do it. You start off with a bunch of en- like, NREL data on like what your solar abundance is in this particular part of the world. Uh, and then you just start throwing solar panels and batteries at it over like the course of a one-year simulation until you hit the number of nines you want, right? And then you, to an extent you can trade the amount of panels and the amount of batteries you've got back and forth and there's like a very, very broad optimum, right?

Dwarkesh Patel 36:51 Right.

Casey Handmer 36:52 Or you can throw in a third thing like a diesel back- backup or a gas turbine or whatever.

Dwarkesh Patel 36:56 The issue here is, look, if these, uh, if META or Microsoft or whoever just wants to get something off the, off the ground, um, this might be low op-ex to have this huge solar farm, but high cap-ex, um, where you need to hire like 30,000 people to go out in the middle of a desert-

Casey Handmer 37:15 Mm-hmm.

Dwarkesh Patel 37:16 ... and install 50,000 acres worth of solar panels and they're like, "Why would I not just buy like 50 gas turbines instead?"

Casey Handmer 37:23 Why, why not just outbid, like Microsoft or META outbids Google or something for the, for the last gas turbine that's available that year? Yeah, I mean, totally. The, the thing that, the thing that like I think META has realized is like, is like Zuck is running out of time to spend his money to win. The cap-ex is not crazy high, just to be clear. Like, the cap-ex is still dominated by just the GPUs.

Dwarkesh Patel 37:40 Right.

Casey Handmer 37:40 Right. So like how much does five gigawatts worth of GPUs cost?

Dwarkesh Patel 37:43 I, I, I'm not m- I don't know if my numbers will be wrong, but like 250 billion or something?

Casey Handmer 37:48 Yeah, 250 billion sounds about right.

Dwarkesh Patel 37:49 Yeah.

Casey Handmer 37:49 Okay. So like is 50,000 acres gonna cost $250 billion in Texas?

Dwarkesh Patel 37:55 That's so much money. Wait, I (laughs) I didn't ma, I didn't do the math in my head. I'm like, "Oh, that's a lot of money." (laughs)

Casey Handmer 38:00 No, we're talking, we're talking like, like maybe 100 million, hundreds of millions of dollars. Something like that.

Dwarkesh Patel 38:02 Right, right.

Casey Handmer 38:02 So it's like literally .1% of the cost is land, right? How much does a megawatt of solar cost? Well, if you go and ask the usual suspects, they'll tell you a mil- million dollars. But this is one of the things that, that breaks my brain at Terraform, which is my day job, which is that the, the modules themselves if y- without tariffs would be like eight cents a watt. So that's like $80,000.

Dwarkesh Patel 38:20 Eight cents a watt?

Casey Handmer 38:21 Yeah.

Dwarkesh Patel 38:22 But they're like a dollar a watt, including installation and everything.

Casey Handmer 38:26 Including installation and everything.

Dwarkesh Patel 38:27 Yes.

Casey Handmer 38:27 But like the panels are the magic part. They're the thing that turns sunlight into pure electrical energy-

Dwarkesh Patel 38:34 Yeah.

Casey Handmer 38:34 ... at 25% efficiency. Everything else should be like less than that.

Dwarkesh Patel 38:37 Right.

Casey Handmer 38:37 And that's what we're... If you wanna work on that project, come and work with us at Terraform 'cause we're very cost sensitive. (laughs)

Dwarkesh Patel 38:42 We'll give you opportunity to shill. Don't worry. Um...

Casey Handmer 38:44 Constantly.

Dwarkesh Patel 38:45 (laughs)

Casey Handmer 38:45 Um, but, uh, no, in, in, in all seriousness, like the, the, the central takeaway is that, is that the hyperscalers are not power cost sensitive. They are power availability sensitive.

Dwarkesh Patel 38:57 Right. Okay. So-

Casey Handmer 38:58 And for all these things, you just run into this like supply elasticity wall at the sort of rates of increase that we're talking about.

Dwarkesh Patel 39:05 Right.

Casey Handmer 39:05 And, and solar is by far the best option for like fire hosing energy at a given problem-

Dwarkesh Patel 39:12 Right.

Casey Handmer 39:12 ... 'cause it's just, it rains down from the sky.

Dwarkesh Patel 39:14 If you're running a frontier technology company, you know how essential it is to recruit the world's best talent, but this requires navigating the Byzantine US immigration system. Not only do you not have the time to deal with this yourself, you just don't have the tacit knowledge to maximize the probability of success. But given how critical exceptional talent is, you can't take any risks with visa approval. You need to work with the best. Lighthouse handles everything better and faster than you possibly could and they do it all with minimal input from you. Lighthouse knows the nooks and crannies of the immigration machine. They know how to frame publications and awards or how to structure comp comparisons versus benchmarks and how to put everything together into the most compelling story possible. And they've even optimized the tiny details like the tone they use when they draft support letters to help US immigration officials understand the importance of tech in startups. Companies like Cursor, Together.AI, and Physical Intelligence are all already working with Lighthouse. You can join them by visiting lightousehq.com/employers. All right, back to Casey.

Dwarkesh Patel 40:24 Between the fact that the, um, maybe solar prices will go down and the fact that demand is gonna go up, do you think electricity prices are likely to rise?

Casey Handmer 40:33 Yes, but electricity prices at this point are a reflection of a regulatory irrationality, right? It's the same situation in Europe, and Australia for that matter. Like, your prices will rise until you've had enough and you say, "No, we demand that you allow us to take advantage of power technology that's been invented in the last 50 years." In terms of things that are causing us to lose to China, tariffs are neither here nor there because as we've discussed-

Dwarkesh Patel 40:58 Interesting.

Casey Handmer 40:58 ... we're not sensitive to cost on power, but the environmental regulations that are actively preventing us from deploying renewable energy in the United States, like this is the reason Texas is winning. Texas is out deploying California ten to one. The regulatory environment around solar is just insane. It's insane.

Dwarkesh Patel 41:15 Say more.

Casey Handmer 41:16 The regulatory environment.

Dwarkesh Patel 41:16 Yeah.

Casey Handmer 41:17 Okay. So in the United States, part of the reason that solar has not been deployed at massive scale yet is that, uh, a bunch of laws went into, into action in the early 1970s that were intended to protect our environment-

Dwarkesh Patel 41:28 Okay.

Casey Handmer 41:28 ... and that makes a lot of sense and our environment's a great thing and we should protect it.

Dwarkesh Patel 41:30 But how, how do, how... I mean, I think people will be familiar with NEPA and whatever-

Casey Handmer 41:32 Yeah.

Dwarkesh Patel 41:32 ... but like how is it especially impacting solar?

Casey Handmer 41:35 Well, let's say you've got a bunch of private land out in the middle of no- nowhere, right? And you wanna build it, solar on it. You'll probably end up triggering NEPA-

Dwarkesh Patel 41:41 Mm-hmm.

Casey Handmer 41:42 ... right? At which point you now have to do what is not in the law but considered necessary, uh, under current regulations, which is like your four-year environmental impact review which generates like so much paper that just the environmental impact of producing the report, 'cause you have to cut down trees to make paper, is more than the environmental impact of just deploying the solar.

Dwarkesh Patel 42:00 Mm-hmm.

Casey Handmer 42:00 Like this is bonkers. It is crazy town, right? Uh, the thing that drives me particularly crazy in southern California is that just because solar's kind of new and off-grid solar is very new, unless you're very, very careful you end up getting regulated as though you're trying to build a chemical plant...... even though it's solar ray. And the impact of solar ray on desert is arguably positive because it shades the ground and i- improves like, uh, soil moisture retention. Um, there are, like if you wanted to reverse desertification, you would basically just place solar panels on it and that would pay for the process. Um, but you, you end up having to go through more stringent environmental review process than if you just wanted to like grade the whole thing and cover it in concrete.

Dwarkesh Patel 42:35 Mm-hmm.

Casey Handmer 42:36 Or grade it and then park a bunch of like old rusting cars that are dropping oil into the aquifer.

Dwarkesh Patel 42:40 Right.

Casey Handmer 42:40 Right? Which in many cases you don't need a permit for at all. Right? But to build the solar, it- you have to go through this whole process. And- and like if there's one thing that anyone listening to this can do, it would be like have a categorical exemption for solar deployment or- or just like, if I put money in an escrow account that says like if after 20 years we have to pull this out, like we'll pull the solar out of the desert and it goes back to being desert, right? I will do that in a heartbeat. But like if I have to hire another biologist for $10,000 to be like, "Well, on that 40-acre plot, we found a tuft of grass which we believe might be a, uh, a critical, you know, one of the 20 species that this particular species of bee, uh, sometimes eats, and this species of bee is not technically endangered but it might be at some point in the future, uh, therefore you can't deploy there," even though it's like zoned unrestricted industrial-

Dwarkesh Patel 43:25 Right.

Casey Handmer 43:25 ... and it's sandwiched between a rocket test stand and like a chemical plant, for example-

Dwarkesh Patel 43:30 Right.

Casey Handmer 43:30 ... in an industrial part of the desert, I will, like, I'm gonna become the Joker.

Dwarkesh Patel 43:33 (laughs)

Casey Handmer 43:33 It is insane. It is like, I just think we need to be a bit balanced about this. It's like, I don't want it to like drive species into extinction, but like the meta problem here is if we don't move our industrial stack off fossil fuels in 10 or 20 years, first of all, we'll get poor the same way UK- uh, UK did, right? Because they ran out of coal basically. And the second thing is, we'll get poor because we'll flood our coastal cities in Florida underneath climate change.

Dwarkesh Patel 43:57 Mm-hmm.

Casey Handmer 43:58 Um, we need, we need solar synthetics for that part, we also need to do sulfur injection and a couple of other things.

Dwarkesh Patel 44:04 Yeah. People will point out that transmission line growth has been stuck in a rut for decades and we have all these bottlenecks and stuff-

Casey Handmer 44:10 Yeah.

Dwarkesh Patel 44:11 ... and s- and in terms of substations and, um, transformers, et cetera, et cetera. Why will this not hamper this, uh, abundant solar future?

Casey Handmer 44:21 Um, that's a really great question. Um, so actually, you and I had a conversation along these lines almost two years ago when we first met.

Dwarkesh Patel 44:27 Mm-hmm.

Casey Handmer 44:27 And, um, and it- it caused me to go and write a blog post. So-

Dwarkesh Patel 44:31 Mm-hmm.

Casey Handmer 44:32 Um, this is, this is a good way of thinking about it.

Dwarkesh Patel 44:34 I think there's another blog post you wrote though which was also to our conversation we had, which was How to Feed the AIs.

Casey Handmer 44:38 That's much more recent, yeah.

Dwarkesh Patel 44:39 Yeah.

Casey Handmer 44:39 That was after dinner, I think.

Dwarkesh Patel 44:40 Yeah.

Casey Handmer 44:40 Um, and- and to be fair like, I usually am fairly clear in my blog posts if I'm posting or if I'm serious.

Dwarkesh Patel 44:48 Mm-hmm.

Casey Handmer 44:48 Um, but, uh, but this one actually I'm dead serious on. It's actually the one where I'm, um, it's like the most out of the money bet as well, like-

Dwarkesh Patel 44:55 Mm-hmm.

Casey Handmer 44:55 ... e- e- everyone else that I consider to be like a respectable forecaster in this area disagrees with me on it. So I'm, that's one side. Um, well, so the grid, we know why the grid's expensive. It's a lot of wires strung up in like hard to reach places that are hard to maintain, especially as, you know, the workforce ages and-

Dwarkesh Patel 45:09 Yeah.

Casey Handmer 45:09 ... and regulations and all the rest and eminent domain and so on and so forth. Okay. So the grid's not gonna get cheaper anytime soon or easier to build. And if you look at the projections of like how much grid, you know, uh, DOE would have us needing to build in the next 10 years versus how much is actually being built, it's like, it's not even in the same order of magnitude. So you say, "Well, are we totally screwed?" The answer is no, we're not totally screwed because batteries actually do the same job that the grid does. But this is kind of weird, just hear me out. The grid transports power from one place to another. It transports almost instantaneously at the speed of light.

Dwarkesh Patel 45:38 Yeah.

Casey Handmer 45:38 So it's actually performing a spatial arbitrage.

Dwarkesh Patel 45:40 Right.

Casey Handmer 45:41 Right? So the idea being that right outside the local nuclear power plant, power is really cheap because they make a lot of it, and in your house, power is really expensive because you don't have a power plant in your house, uh, and you- you pay the intermediator a- a small fee and they allow this- this trade to take place. Um, and that's basically how the grid works. And until quite recently, the only way we had of meaningfully storing energy, like storing electricity on the grid was pumped hydro, and that only works in a handful of places in limited, limited capacity. And it doesn't work all that well, either. The efficiency is not great. Now we have batteries. Batteries store power at one time of day and they release it at another time of day. So batteries are performing a temporal arbitrage-

Dwarkesh Patel 46:21 Mm-hmm.

Casey Handmer 46:22 ... an arbitrage over time. But they can be local or they can be more remote. I think we'll end up seeing batteries next to the solar arrays and batteries in the middle of the grid at substations-

Dwarkesh Patel 46:31 Mm-hmm.

Casey Handmer 46:31 ... and batteries on the sites of existing power plants that get turned off and batteries in your house and batteries everywhere in between. One way of thinking of this is what is your per capita allocation of batteries in kilograms per head? And like when you and I were much younger, you know, the lithium-ion battery was in a cell phone size, so you're like, I don't know, 10 grams per person or something. And nowadays, half the people in this town drive Teslas, so your per capita allocation of- of lithium-ion battery is, you know, 100 kilograms or something like that.

Dwarkesh Patel 46:59 Mm-hmm.

Casey Handmer 46:59 So we're talking like four or five ooms of increase (laughs) of total-

Dwarkesh Patel 47:02 Mm-hmm.

Casey Handmer 47:02 ... total battery per person. That trend is only going to continue. And then you say, "Well, okay, we've got batteries that are performing this temporal arbitrage and the sun comes up every day," right? So like the- the- the power swing from like midday you're otherwise curtailing the solar array, to dusk when everyone's watching TV and cooking dinner or running the air conditioners to cool off in the evening is very predictable, right? Whereas like, "Oh, we had like really bad weather so we had to use the power line that runs to the pow- extra power plants over by Hoover Dam or something," doesn't get used nearly as much or like its- its peak utilization is almost never, which means that the utilization of the batteries is on average like let's say 300 days a year and the utilization of your most expensive grid, m- highest voltage grid assets is much, much lower. And that includes like the trans- the s- substations and transformers and stuff that serve that. So it's a really bad position to be in if you're a grid operator, right? Because you've got this- this aging existing thing that the batteries are cannibalizing. The batteries are being installed behind the meter. You don't have a say in whether they're being installed, how they're being used. All you know is that your, the, your utilization of your asset where you get to charge top dollar for it is just dropping year after year after year after year as the same time as your operating costs are increasing year after year after year-So it just very clear that like the average distance the electron is gonna travel between generation and consumption is going to decrease in the future pretty radically.

Dwarkesh Patel 48:23 Yeah, I guess the-

Casey Handmer 48:24 It's already decreasing. It's gonna continue to decrease.

Dwarkesh Patel 48:26 The difference is that... I mean, it's especially helpful for solar but like solar is the one that's most intermittent.

Casey Handmer 48:33 You can predict, you can predict the amount of solar power you're gonna get in three days pretty accurately 'cause of weather prediction.

Dwarkesh Patel 48:38 But you can't like buy more batter- you can't like change the amount of batteries you have.

Casey Handmer 48:41 Well, actually in the limit you can because you can put them on trucks and drive them around.

Dwarkesh Patel 48:44 Mm.

Casey Handmer 48:44 Right? So there could be a capacity market for batteries where you drive them around to people who need them and get

Narrator 48:48 I understand.

Casey Handmer 48:48 ... in a pinch. In practice, it's gonna be cheaper just to double the size of your battery, um, because batteries are gonna continue to get cheaper and cheaper and cheaper and cheaper.

Dwarkesh Patel 48:54 Right.

Casey Handmer 48:55 But what it does mean is you can say, "Well, I know that I'm gonna have, you know, three low days."

Dwarkesh Patel 49:00 Yeah.

Casey Handmer 49:00 "So I will start curtailing now by 5% so I don't have to curtail by 50% in three days. And then overall for the whole year, I'll only curtail, you know, five hours so I'm still at four nines, um, instead of having to curtail 24 hours 'cause I can't predict the weather."

Dwarkesh Patel 49:14 Okay, so let- let's assume you're right and then... I mean, I think at some point you will be right. Like maybe we disagree about... Sorry, I- I'm not qualified to disagree. Maybe you and some other person might, uh, disagree about what year it happens, but I think it's hard to deny that in the asymptote-

Casey Handmer 49:28 Oh, yeah.

Dwarkesh Patel 49:28 ... our civilization is headed towards lots of energy used for AI and a lot of that coming from solar.

Casey Handmer 49:34 Mm-hmm.

Dwarkesh Patel 49:34 Um, in that asymptote I've just like... I wanna get like the crazy nerd sci-fi like what does our civilization look like? What is happening in this, you know?

Casey Handmer 49:42 Kardashev level one.

Dwarkesh Patel 49:44 Yeah. L- l- let's wait til get to turning the entire Earth into an AI factory, but more like, I don't know, the 2030s, um-

Casey Handmer 49:52 Mm-hmm.

Dwarkesh Patel 49:53 ... where you've gotten multiple people who are building on the order of five gigawatt or 10 gigawatt sites. The value of the hardware is dependent on its complement, which is the software, right? Like right now AI models are fine and so the hardware they're running on, the economic value they can generate is sort of bottlenecked by how good the software is. But if you actually had AGI, if you had like a human level intelligence or maybe even better-

Casey Handmer 50:16 Ideally better, yeah.

Dwarkesh Patel 50:17 ... yeah, um, running on an H100-

Casey Handmer 50:19 Hmm.

Dwarkesh Patel 50:19 ... uh, that H100 is worth a lot, right? Like we're paying a lot for humans to do work. Right now I don't think AI is that valuable. Like the models themselves aren't super, super valuable in terms-

Casey Handmer 50:30 Mm.

Dwarkesh Patel 50:30 ... of just pure economic value, right? OpenAI is generating on the order of 10 billion ARR or some- uh, you know, 20 billion ARR. And-

Casey Handmer 50:37 That sucks. It's terrible. How do- how can they sleep at night?

Dwarkesh Patel 50:40 I know.

Casey Handmer 50:40 (laughs)

Dwarkesh Patel 50:41 Um, but like for context McDonald's and Kohl's generate more yearly revenue than that. But I think the promise of AGI-

Casey Handmer 50:48 Yeah.

Dwarkesh Patel 50:48 ... is to automate human labor. Human labor generates on the order of $60 trillion of economic value or like that's how much is paid out in wages to labor around the world, right? So that's what AGI can do and even if you curtail it to just white collar work, there's still tens of trillions of dollars of value. So once we have models which are actually human level-

Casey Handmer 51:07 Hmm.

Dwarkesh Patel 51:08 ... they will be worth at least that, um, pending the fact that you can build them or you can run them.

Casey Handmer 51:12 Well, I don't think we should constrain ourselves to being like, "Oh, well maybe there'll be some fraction of current payroll," right? 'Cause like that- that's kind of a very contingent on like humans being humans thing.

Dwarkesh Patel 51:22 Yeah, no, I- I think that's a lower bound, to be clear.

Casey Handmer 51:24 Oh, yeah. Low- lower bound for sure. But like if you think about like someone- someone trying to like estimate the upper bound for the market value, the market cap of like Caterpillar, right? Based on like, well, it takes, you know, this many men and like-

Dwarkesh Patel 51:37 (laughs)

Casey Handmer 51:37 ... burrows and wheelbarrows to like dig a trench and so, you know, it couldn't be more than that, right? But actually, you know, one way to think about the industrial revolutions is every time you figure out an industrial revolution what you're doing is you're finding some way of like bypassing a constraint or-

Dwarkesh Patel 51:50 Yeah, yeah.

Casey Handmer 51:50 ... bypassing a bottleneck and the bottleneck prior to what we call the industrial revolution was, uh, metabolism, right? It's just like how much like oats can a human or a horse physically digest and then convert into useful mechanical output for-

Dwarkesh Patel 52:04 Right.

Casey Handmer 52:04 ... you know, their- their, um, their peasant overlord or whatever. And- and nowadays, we would giggle to think that like, oh, the amount of food we produce is meaningful in the context of like the economic power of a particular country.

Dwarkesh Patel 52:15 Right.

Casey Handmer 52:15 Um, because 99% of the energy that we consume routes around our guts, you know, through the gas tanks of our cars and through our aircraft-

Dwarkesh Patel 52:23 Right.

Casey Handmer 52:23 ... and in our grids and stuff like that.

Dwarkesh Patel 52:24 Interesting.

Casey Handmer 52:25 And so right now, the AI revolution is about routing around cognitive constraints-

Dwarkesh Patel 52:32 Yeah.

Casey Handmer 52:32 ... that in some ways writing, you know, like writing, printing press, computers, the internet have already allowed us to do to some extent. Uh, a credit card is a good example of-

Dwarkesh Patel 52:40 Yeah.

Casey Handmer 52:40 ... something that routes around a- a cognitive constraint of like building up network of trust, right?

Dwarkesh Patel 52:44 Yeah.

Casey Handmer 52:44 It's- it's a centralized trust.

Dwarkesh Patel 52:45 Yeah, that- that- that's interesting. And wait, wait, wait, it's also really interesting something that came up, uh, and I want to credit James Bradbury and Gordon with making this interesting point when I was talking with them a couple of days ago is if you measured by GDP, AI's outputs might be underwhelming, right? The- one of the complaints that economists-

Casey Handmer 53:01 Hmm.

Dwarkesh Patel 53:01 ... have about the internet is that it's hard to measure the consumer surplus that's created by the internet because a lot of the, um, a lot of the goods and services that are made available-

Casey Handmer 53:09 Mm-hmm.

Dwarkesh Patel 53:10 ... you- you pay zero for them and so they don't show up in GDP.

Casey Handmer 53:13 Well, it's the same with oil.

Dwarkesh Patel 53:14 Yeah, in- in that sense that it's only 1% of G- energy is like 1% of GDP.

Casey Handmer 53:17 Well, oil- oil's like $8 trillion a year or something, right?

Dwarkesh Patel 53:19 Yeah.

Casey Handmer 53:20 But if- if you, if you, if you said, "Well, one day we're gonna consume 100 times more energy in the form of oil than the form of food and the per joule cost of food is, you know, whatever it is, cost of a Big Mac, then oil should be like $800 trillion a year, right? It's 100 times... Like per- per unit energy, oil, like gasoline, is 100 times cheaper than the cheapest food that humans can digest.

Dwarkesh Patel 53:40 Mm-hmm.

Casey Handmer 53:40 So- so- so like does that mean that we've like shot ourselves in the foot by using oil to run our economy-

Dwarkesh Patel 53:46 Right.

Casey Handmer 53:47 ... because it's so cheap? Like, no. (laughs)

Dwarkesh Patel 53:50 Right, right. And- and- and also the amou- the fra- it's fraction of GDP also doesn't correspond to how important it is. For example, oil is like 1% of GDP or something, but, um, if you don't have oil, then you have these oil shocks which cause double digit decreases in GDP. So, the elasticity of demand often matters more than its like raw fraction contribution to GDP. But anyways, on the original point about AI, so you're gonna have this huge deflation of... U- so Gordon put it this way, he's like, "If, um..."If you imagine Dario's data center of geniuses-

Casey Handmer 54:24 Yes.

Dwarkesh Patel 54:24 ... how is that showing up in GDP? Well, it would be the inputs which are the chips, the energy, et cetera-

Casey Handmer 54:28 Mm-hmm.

Dwarkesh Patel 54:29 ... and the outputs which are just the tokens. And neither of those is gonna be that astronomical in comparison to the value of those, that data center of geniuses is producing. Um, so in terms of GDP numbers, like, if that data center of geniuses automates a bunch of, um, or at least compliments a bunch of human work, et cetera, it might actually cause like a nominal decrease in GDP while at the same time contributing massively to, um, what we might think of as the valuable stuff humanity or human civilization can produce. And so, in the long run, uh, it might make more sense to think of the size of our economy or the size of our civilization as the raw energy use that we do-

Casey Handmer 55:11 Mm-hmm.

Dwarkesh Patel 55:11 ... rather than GDP. Because aga- again, GDP will see this huge deflation-

Casey Handmer 55:14 Yeah.

Dwarkesh Patel 55:14 ... because the variable cost of running AI-

Casey Handmer 55:17 Yeah.

Dwarkesh Patel 55:18 ... will just be pretty cheap, um, as compared to like paying humans' wages, et cetera.

Casey Handmer 55:22 I think, I think we're in the, we're, we're at, at, at the point where like you've got a mixed economy with, like, an AI doing my job and also a human doing my job.

Dwarkesh Patel 55:31 I love how this is the, this is the new way we use the phrase mixed economy. (laughs)

Casey Handmer 55:34 Yeah. Then obviously, like, I still have some pricing power relative to humans, and the AI thus has pricing power, right?

Dwarkesh Patel 55:39 Yeah.

Casey Handmer 55:40 But, but if it was the case that, like, a new, um, a new kind of job emerges that AI is really well adapted to, like, the... because it's not competing against humans for most of those roles, it'd be competing against the other labs, right? And so you'd actually see, like, the cost push down to, you know, a mul- a small multiple of whatever the marginal production cost of those tokens is. That would be my guess. So I, I think, I think it might be a mistake to as- to assume that, like, well, if, if we're gonna pay, you know, a top AI researcher $200,000 (laughs) $200,000 a year, well, um, let's say the sort of AI researcher that I could be, $200,000 a year, that if, uh, an AI comes along that's as good as me at that, even taking into account the fact that realistically speaking I only get, you know, maybe 10 hours of really top cognitive work done a week, that it would also be worth $200,000, right?

Dwarkesh Patel 56:24 Yeah.

Casey Handmer 56:24 Obviously, it's, it'd be worth much more than that in, in the sense of, like, you can copy-paste its output-

Dwarkesh Patel 56:30 Yeah, yeah.

Casey Handmer 56:30 ... and much less than that in the sense that, you know, whatever the marginal additional cost of, of spooling up H100s is. Um, and so if, if some, if some kind of role comes along that, like, the AIs are really well specialized at and out-compete the humans quickly, then we'd also expect to see that their, um, both the, the cost of providing that service to drop drastically-

Dwarkesh Patel 56:46 Yeah.

Casey Handmer 56:46 ... at the same time as the overall value generated in the economy by that service would increase a lot.

Dwarkesh Patel 56:50 Yeah, yeah, exactly. Um, if we think that the value of cognition is going to be unbounded and the way to derive cognition, you can just, to the extent you think solar will e- eventually win, you can just, you can derive it from how much land it takes to power an H100, uh-

Casey Handmer 57:07 Yep.

Dwarkesh Patel 57:07 ... using solar panels. Th- that is a very interesting derivation that, like, "Okay, well, this is, at a minimum, we're going to just fill up all the land." I mean, at some point, you might have, like, declining marginal value of cognition or something, but-

Casey Handmer 57:19 We kind of discussed this earlier, but if you have, like, a megawatt of, um, 10 acres, 10 acres of, of land feeding one megawatt H100 or something-

Dwarkesh Patel 57:27 Right.

Casey Handmer 57:27 ... that's generating, like, let's say a megawatt is 1,000 humans. So one acre is 1,000 humans' worth of-

Dwarkesh Patel 57:33 Right.

Casey Handmer 57:34 ... cognition. The implicit land value there is a lot higher than it is as, like, undeveloped desert, right?

Dwarkesh Patel 57:40 Mm-hmm.

Casey Handmer 57:41 It's also a lot higher than it is as, like, the most productive farmland that humanity has ever had.

Dwarkesh Patel 57:46 Right.

Casey Handmer 57:46 Right?

Dwarkesh Patel 57:46 At current hardware efficiencies, I don't know if it's worth spelling out-

Casey Handmer 57:50 Mm-hmm.

Dwarkesh Patel 57:50 ... basically, H100 has the same amount of FLOPS as the human brain, but also uses way more energy than a human brain.

Casey Handmer 57:56 Yes.

Dwarkesh Patel 57:56 Uh, it uses like 50X more energy.

Casey Handmer 57:58 Is that right?

Dwarkesh Patel 57:59 So 20 watts versus 1,000 watts?

Casey Handmer 58:01 Human brain, like 20 watts, yeah, could be.

Dwarkesh Patel 58:03 We know hardware can be at least as efficient as the human brain, and the human brain can generate this many FLOPS on 20 watts.

Casey Handmer 58:09 Yeah.

Dwarkesh Patel 58:09 So if you do that, uh, calculation-

Casey Handmer 58:12 Yeah.

Dwarkesh Patel 58:12 ... then that's 50X 1,000, so 50,000-

Casey Handmer 58:15 Yep.

Dwarkesh Patel 58:16 ... AI solars off of-

Casey Handmer 58:17 And it could easily be-

Dwarkesh Patel 58:18 ... one acre.

Casey Handmer 58:18 ... much more than that because neurons are much slower-

Dwarkesh Patel 58:20 Mm-hmm.

Casey Handmer 58:21 ... than, than, um, transistors obviously. Um, so, uh, probably 10 years ago, one of my friends reminded me, um, uh, you know like the way your phone saves power is it goes to sleep between you, like, tapping out hello?

Dwarkesh Patel 58:31 Right.

Casey Handmer 58:31 Like H, it goes, te- it takes a nap for like 10,000 cycles, A, you know? Um, it's kind of nuts. So-

Dwarkesh Patel 58:37 Right.

Casey Handmer 58:37 ... just that, like, humans... I think Elon's talked about this in the context of self-driving cars as well, which is like anything humans do is like glacially slow from the perspective-

Dwarkesh Patel 58:45 Mm-hmm.

Casey Handmer 58:45 ... of a computer.

Dwarkesh Patel 58:45 Let's actually go back to the original point of, like, I was explaining w- why I think it's plausible that there could be more than hundreds of gigawatts of extra demand from AI in the 2030s. I want to understand what that looks like in the real world. Like, at that p- point, it's become basically this industrial problem of can you generate enough solar panels and solar modules and batteries-

Casey Handmer 59:06 Mm-hmm. Yep.

Dwarkesh Patel 59:07 ... and not to mention the chips themselves?

Casey Handmer 59:08 So I've seen an industrial point, and then there's a cultural point as well which, like-

Dwarkesh Patel 59:11 Well, let's start with the industrial point. (laughs)

Casey Handmer 59:12 Okay. (laughs)

Dwarkesh Patel 59:13 Um, I, I, I want-

Casey Handmer 59:14 Okay, let's talk about it.

Dwarkesh Patel 59:15 I want to know what the year 2035 looks like if-

Casey Handmer 59:16 Yeah, yeah, yeah.

Dwarkesh Patel 59:17 ... if we've got AGI and we're just bottlenecked by the ability to deploy it.

Casey Handmer 59:20 Yeah. So you can ask a question like, what do you need in order to run, like, what is the minimum amount of matter that you need in order to perform these, these calculations?

Dwarkesh Patel 59:28 Yeah.

Casey Handmer 59:28 So, so, so right now, we're talking about, like, AI racking and grid and transmission and a bunch of, like, ISOs and all that rest. You don't need any of that stuff, right? And, like, obviously, XAI is on top of this because the first thing that Elon will always ask is, like, "Delete anything you don't absolutely need." Um, so what you actually need is a big slab of relatively cheap silicon to make the power and then a small slab of relatively expensive silicon to do the thinking.

Dwarkesh Patel 59:52 Mm-hmm.

Casey Handmer 59:52 And if it's in space, that's all you need, right, 'cause it's in the sun all the time, so you don't need a battery, right? But if you're on the Earth, you need a battery as well, so you need something to connect. Um, and you don't need a transformer, right? You don't even need a DC-DC converter. You can actually make do with buck converter or with relays or whatever to basically match the, the current output of your solar array with the charge state of your batteries and the power consumption of the, um, of your, you know, uh, GPU or something.

Dwarkesh Patel 1:00:18 Mm-hmm.

Casey Handmer 1:00:18 But, like, a, a solar array about the size of this desk, for example, will generate about 500 watts in, in, in full sun. So you could actually imagine, like, a-... aliens who have different silicon technology stack, building their systems as, like, an integrated solar array with, you know, a bit of, uh, a bit of, you know, computronium in the middle, for example-

Dwarkesh Patel 1:00:35 Mm-hmm.

Casey Handmer 1:00:35 ... uh, in, uh, on the same wafer. Uh, but that's, that's basically all you need, and then-

Dwarkesh Patel 1:00:38 On the same wafer, this-

Casey Handmer 1:00:39 Yeah.

Dwarkesh Patel 1:00:39 ... because it's bo- it's all silicon?

Casey Handmer 1:00:40 It's all silicon.

Dwarkesh Patel 1:00:41 Yeah.

Casey Handmer 1:00:41 It's all silicon all the way down. And what's silicon made of? Well, it's an, it's an element. It's, uh, chemically in the crust. Uh, there's no shortage of it.

Dwarkesh Patel 1:00:47 This is a, this is a great prompt for a sci-fi exercise because if, especially in space, you don't need batteries, so you just like... The, the prompt is, uh, the, the future TSMC just manufacturers integrated solar, uh...

Casey Handmer 1:00:58 And they can fly around.

Dwarkesh Patel 1:00:59 Solar dyes.

Casey Handmer 1:00:59 Right? They're, they're solar cells.

Dwarkesh Patel 1:01:00 Yeah.

Casey Handmer 1:01:00 Right? And they don't, they're, they're relatively dense, so they don't fly crazy fast, but they don't need to because they're immortal.

Dwarkesh Patel 1:01:04 Is, is this what the Dyson sphere will be made of, Casey? Is it just, is it just gonna be comped chironium at the center of, uh, solar cells?

Casey Handmer 1:01:09 They can fly closer to the sun to get more power.

Dwarkesh Patel 1:01:11 Right.

Casey Handmer 1:01:11 Right? Up to the thermal limit. And they can fly further from the sun to go and explore or, like, fly to other planets or something.

Dwarkesh Patel 1:01:16 Huh.

Casey Handmer 1:01:16 And they can, they can adjust the or- orientation of the solar sail with, uh, LCD panels so that it can be integrated into the, into the wafer itself.

Dwarkesh Patel 1:01:23 Mm-hmm.

Casey Handmer 1:01:23 So, I mean, that's a, that's actually... If you say what, what's the, what's the post-human state, that's it. You know? Like, that's... We will, we will eventually be-

Dwarkesh Patel 1:01:29 A solar sail with a silicon die in the middle for comp compute.

Casey Handmer 1:01:32 Yeah. One, one human's worth of, of computation, one human brain, you know, can be simulated in roughly a square meter of silicon floating in space. Um...

Dwarkesh Patel 1:01:38 Wait, how, how much? Sorry.

Casey Handmer 1:01:39 One, one square meter of silicon.

Dwarkesh Patel 1:01:40 Mm-hmm.

Casey Handmer 1:01:40 Like a, like a, the thickness of a sheet of paper floating in space.

Dwarkesh Patel 1:01:42 Right. Yeah, interesting.

Casey Handmer 1:01:43 That's it. That's a future human form. That's, that's my final form.

Dwarkesh Patel 1:01:45 That, that's the attractor state. (laughs)

Casey Handmer 1:01:46 (laughs) Yeah. That's assuming, that's assuming a little bit of software improvement. But like, I don't think that's-

Dwarkesh Patel 1:01:50 (laughs) The, uh, the, all that's assuming (laughs) is software improvement.

Casey Handmer 1:01:54 Yeah. Um...

Dwarkesh Patel 1:01:55 Um, the Dyson sphere, that's all it needs, is a little bit of software, a little bit of tweaking of the algorithms.

Casey Handmer 1:01:57 Well, I mean, if, if the soft- like, the, the, the area of the panel, you know-

Dwarkesh Patel 1:02:01 Mm-hmm.

Casey Handmer 1:02:01 ... is, is kind of the, the variable there.

Dwarkesh Patel 1:02:02 Right.

Casey Handmer 1:02:03 Um, so you can ask, well, what, what do you need in order to make the silicon? Uh, making, uh, solar rays, making chips is a kind of multi-stage process, but, but basically you start off with, uh, silica- silicates, which are rocks, uh, ideally in a relatively pure form. You, uh, chemically reduce them. Um, there's a couple of different processes that can do that. And then purify them into, like, ideally like six nines of purity for solar ray, maybe nine nines for like really nice computers. Um, silicon purity and, uh, you can grow crystals, cut wafers, et cetera, et cetera. So then the, then the constraint is like, well, how quickly can you convert the crust into enough silicon to support, you know, silicon thought?

Dwarkesh Patel 1:02:35 Yeah.

Casey Handmer 1:02:35 Um, what does the silicon ecosystem look like?

Dwarkesh Patel 1:02:37 Any thoughts?

Casey Handmer 1:02:38 Well, it's, it's pretty quick, right? (laughs) Like if you, like how many... If you have like one kilowatt per square meter and then you use that just to like rip oxygens off the underlying dirt, you know-

Dwarkesh Patel 1:02:46 Mm-hmm.

Casey Handmer 1:02:46 ... it doesn't take all that long to... You only need about 20 microns of silicon to make, uh, solar PVRA. So it's-

Dwarkesh Patel 1:02:52 So like from dirt as you mean like actual dirt?

Casey Handmer 1:02:55 Yeah, actual dirt has plenty of silicon in it. Yeah.

Dwarkesh Patel 1:02:56 Huh.

Casey Handmer 1:02:57 So for example, setting up a brand new silicon refinery takes about 18 months.

Dwarkesh Patel 1:03:00 Right.

Casey Handmer 1:03:00 Right. Um, but that's just with the current technology, right? This... I, I actually think we may find ways... So one of the nice things is if you have infinite, infinite free solar power, approximately free solar power, you can revisit a bunch of legacy industrial processes that are, have been optimized for efficiency and say, "Well, what if we just like use twice as much power and we just want to do them faster and cheaper?" Right? Like less capex, less lead time, more power, right?

Dwarkesh Patel 1:03:25 Mm-hmm.

Casey Handmer 1:03:25 Well, you can start solving problems. It turns out that, um, that if you want to chemically reduce silicon, you can do it electrolytically, like with less efficiency and a bunch of like... under a hydrogen ret- rich atmosphere or something. And then, uh, so one of the ways that silicon can be refined is by turning it into silane, which is a silicon tetrahydride.

Dwarkesh Patel 1:03:41 Mm-hmm.

Casey Handmer 1:03:42 I'm not really a chemist, but I think that's