电网中隐藏着200GW，钠电池降价10倍，波浪供电数据中心：Ramez Naam｜EP #280

1. The scale of current and future solar installations

In the U.S., if you put in a request for hundreds of megawatts of power to build a data center today, good luck getting that power before 2031. That’s the situation we have today. New technologies like sodium-ion batteries could drop the cost of batteries by a factor of 10. We already have the very first solar-plus-battery baseload power plants, and they’re affordable. Batteries are plunging in cost and are ultimately going to drop another 10×.

Do you believe the thesis that solar, in the long term, is going to dominate beyond everything else?

The reality is, everybody’s heard Elon talking about space-based data centers. Ten gigawatts a year is like 5 or 6 Starship launches a day.

Unless we hit that exponential absolutely full-on and go right down that path, this looks prohibitive for 15 to 20 years. The biggest unlock that we cannot predict in AI and power will be...

Today, we’re going to do a deep dive into the innermost loop: all things energy. I’m here with my extraordinary Moonshots mates, Dave Blundin, Alexander Wissner-Gross, and Salim Ismail. Welcome, gentlemen. Good to see you all. Looks like you’re at your normal haunts.

We’ve got a friend with us today. It’s an honor and a pleasure for us to invite Ramez Naam. Ramez is a computer scientist, investor, and author—one of the clearest thinkers on the future of energy. After a career at Microsoft, Ramez became a leading voice on the exponential decline in the cost of solar, batteries, fission, and fusion. He’s the founder and managing partner of Planetary VC, investing in energy companies.

2. Guest introduction: Remez and his work on energy futures

He’s the author of The Infinite Resource and one of my favorite-ever science-fiction novel series, the Nexus trilogy. If you’ve not read Nexus, I cannot recommend it enough. On Book Corner, where Alex and I talk about our books, I’ve mentioned Nexus a few times.

Today, we’re going to explore the innermost loop: why energy abundance may arrive faster than most forecasts, and its impact on AI, economic growth, geopolitics, and our future. Again, my mission here is that, at the end of this podcast and the brilliant dialogue my Moonshots mates are going to bring to the table, whether you’re an investor or a builder, you’re going to understand: What’s the alpha? Where is it going? What are the real timelines for everything from building out nuclear plants and fusion plants? Because sometimes there’s hype, and sometimes there is an overwhelming abundance of energy coming our way.

First off, Ramez, welcome, pal.

Peter, it’s an honor to be here. Great to be here with friends. I look forward to it.

You do have friends here.

I have a quick story.

Yeah, of course you have a story. [laughter]

I’m worried. I remember we were presenting to one of the top oil and energy companies in the world—one of the top 3 or 4—and they were like, “Who’s this Ramez fellow? We want to grill him before we let him in front of the key people here.” We were like, “Fine, grill Ramez.”

We were talking a lot about solar, and they’re an oil and gas company. After about 2 hours, they were like, “Okay, we need to get in front of him.” [laughter] That was it. It was an awesome session.

Yeah. One thing I failed to mention is that Ramez was part of our founding faculty at Singularity University. He really led the whole energy conversation there and has been onstage at the Abundance Summit a number of times. Hopefully, you’re back again in 2027.

Ramez, first of all, I just need to try you: You need to write a 4th, 5th, and 6th trilogy. [laughter]

As soon as AI and energy get less exciting every single week, I will make time to write another novel.

How much time do you spend tracking what’s going on in the innermost loop here?

It’s every day, all day. That’s what we all do, right? Living in the singularity. Yeah, we’re living the singularity.

Before we get into it, Alex, do you want to add anything to the conversation up front?

I’ll just add: Welcome to the terror dome. One of my favorite popularizations of yours is that now-infamous chart of the price of solar going down to zero.

Thank you, Alex.

3. AI's power hunger and the need for grid expansion

Amazing. I can’t take another second away from you, pal. Jump on in, and we’ll grill you along the way and make the points, shall we say, in a stellar fashion.

Great. Let’s just start. We’re going to hit a few different topics here with the intersection of electricity—really, energy—and compute. A few years ago, as an investor in clean energy, that was a fringe sector to some people, though it’s a $3 trillion sector. But now that we see that AI depends upon electricity, it is everything. Value flows to that which is scarce, and right now power is scarce.

We’re going to cover 6 topics. At the end of each, we’re going to pause for discussion. Number 1: AI is power-hungry. 2: speed to power and the grid—that’s everything; it’s not cost. 3: behind-the-meter power—that’s how it’s happening. 4: making the grid better is totally undervalued, and that’s where the near-term wins are. 5: solar. 6: fission and fusion. And 7, finally, the out-of-this-world ideas: launching compute into space or launching it into the oceans.

Let’s cement ourselves on AI as power-hungry. You have to exponentially increase compute to get linear gains in AI. There are some ways to cheat that curve, which we’re doing, but that’s the basic phenomenon here, and I don’t think we fully grasp—most people, anyway—the relationship between these things.

First, I want to be clear that power is cheap compared to GPUs. If you look at building a gigawatt data center, you’re going to spend $50 billion, with $35 billion of that going to chips. When you compare the ratio of the all-in capex of your data center to your 5-year energy cost, it’s amazing how little energy costs.

So, when you say AI is power-hungry, it’s not really a cost issue. It is that energy is the bottleneck for AI. This has a lot of ramifications because these numbers are in billions or tens of billions of dollars. Every hyperscaler has whole teams devoted to optimizing the cost of energy.

But if you tell OpenAI or Anthropic today, “Look, we can give you power at twice the cost, and it’s on tomorrow,” they’ll take it. They won’t tell you that, of course, but they will take it, because the ratio of the revenue you can generate from a unit of electricity to its cost is basically the same as this.

So, what’s the challenge? The challenge is that we stopped being able to build out the grid fast. I’m not talking about power generation. We can still do that pretty fast, at least for solar, wind, batteries, and natural gas. But the poles and wires are a huge problem.

We talk about the interconnection queue, which is the queue to get your new project hooked up to the grid. This is for the generation side: If you’re building a new solar plant, wind plant, or natural-gas plant, how long does it take before you’re hooked up to the grid so you can deliver power to your customers? That’s gone from 15 months 20 years ago to now coming up on 45 months.

Regulations—what is it?

4. Log-linear relationship between compute and AI progress

It’s regulation, and it’s also that demand growth has slowed, right? U.S. demand growth per year is much slower than it was in the ’80s, even let alone the ’50s. Utilities have just re-engineered themselves. They’re more oriented toward customer service, meeting the regulators’ demands, and so on, than they are toward building stuff fast.

That has gotten in the way. But permitting is also a huge issue—not utility regulation per se, but permitting issues for land controlled by the state, the county, and the federal government.

If we’re just going to jump in, led by Peter’s example, I have to ask you: You sort of flew by this. You mentioned or alluded to the notion that intelligence was somehow proportional to the log of compute. I know a number of executives have also pushed the narrative that maybe one could naively extrapolate some law that looked like that from scaling laws in machine-learning training or machine-learning inference.

Do you think that’s actually true? And if you do think it’s true, do you think it continues to be true in an epoch of recursive self-improvement?

It’s an awesome question, Alex. This is core to the big questions of AI: Are we going to have recursive self-improvement to ASI?

Everything in machine learning since 2000 has shown something like a log-linear relationship, really between training-data size and the precision of a model. You’re alluding, I think, to first Kaplan scaling, then Chinchilla scaling, and then post-Chinchilla scaling.

Long before Chinchilla scaling, with single-layer neural nets, we were finding this in the early 2000s, right? Compute is used to convert training data into a model, into a neural network, and that has a roughly log-linear relationship.

But we cheat—and by that I mean we keep finding ways to make it more efficient. So, is it actually log-linear? No, it’s a little bit faster than that because we keep finding ways, as we see with DeepSeek Flash, which just came out, and as we see with Kimmy K3. We keep finding ways to bend that curve.

It’s a little bit less bad than log-linear, but it’s still somewhere between a power law, like n to the 5th, and a true exponential, or log-scale, difficulty.

I don't see that changing anytime soon.

This seems—if I understand your broader thesis—almost axiomatic: if we can't bend the curve, we need basically exponentially larger amounts of energy just to make essentially linear or polynomial progress in intelligence. We need more and more energy. We need to achieve a Kardashev Level II or Kardashev Level III–type civilization, with Dyson swarms, just to keep making incremental progress. [laughter] Right?

This is getting at the core issue with superintelligence, actually. To a certain extent, here's my view, Alex: the naive view is that, at any given time, intelligence is basically log-linear with compute and log-linear with data, but we keep making the algorithms better, and that sneaks us toward a polynomial domain. The polynomial domain is still steeply diminishing returns. You're arguing it's polylog; you're arguing that intelligence is polylog in compute.

At best, it's polynomial. I don't necessarily mean polylog, but at best, the very best examples suggest that compute has to go up as n⁴ to get an n-size increase in intelligence.

And you don't think recursive self-improvement, if we are indeed in an era of RSI, causes anything better than polylog?

No. Look, you do the math on RSI, and every way that you improve AI has diminishing returns. So, every model of RSI that does not include hardware—we can set that aside. Every model of RSI in software fizzles over time.

The bump might be so big that we're like, “Wow, this is just over-the-top amazing,” but it always looks concave. There is no mathematical model of RSI that's valid, that I can see, that leads to an actual vertical asymptote takeoff.

I'm going to take us back. So, the grid is the grid. [laughter]

The grid is the bottleneck right now.

Yes, it is. So, we need power in a practical sense. Look, everybody—

And Alex, not that I didn't appreciate your genius in those questions—that was fun. This may turn out to be an entire conversation between Ramez and Alex, but we'll see.

Let's have another way to talk about that.

I have words. [laughter] I think the point here is that the grid is the bottleneck. That's a really important point because it speaks to the infrastructure needs we're going to have for dealing with this. So, let's move on. I'll come back to it.

For the power side, we don't have data that's as clean for the demand side. But here are 3 locations around the U.S. You see that, in this 7-month period between April and November last year, the wait times to get connected for the load side—not for generation, but for your data center or whatnot—went up by 6 or 7 months.

Everywhere around the country, as demand is going up for large-load interconnection, you're seeing longer and longer waits. This is the Texas grid. ERCOT currently peaks out at about 80 gigawatts. They have submissions into their demand-side queue for well over 200 gigawatts of load.

5. Moving data centers to where energy is abundant

Most of this is speculative. Most of these submissions are BS. Not quite anybody can put in a request for a large load and power without actually having financing or a customer, and so on. Most of these things evaporate. But in any case, the Texas grid operator is overwhelmed with these requests for power, and of course, that just jams up everything.

Wait, let me ask a quick question. Are we talking about, “I have a data center, I want to connect it to the grid to get power,” or, “I have a new power source, I want to connect it to the grid to deliver power”? Is it about the same either way?

This chart is generation: I've got a new power source, and I want to connect it to the grid to deliver power. These 2 are demand, or load, interconnection queues, and they're going up.

Honestly, it's much longer than that. Today in ERCOT in Texas—the most advanced, most progressive—

Progressive.

—least—yeah, most progressive in a positive sense, least regulatory-burdened, fastest-moving grid in the U.S.—if you put in a request for hundreds of megawatts of power to build a data center today, good luck getting that power before 2031 or 2032.

That's the situation that we have.

Yeah. When you did a presentation for my Abundance community at our monthly meetup, that was my major takeaway: the issue with energy for AI isn't building solar farms. It isn't fission or fusion. It's the grid. The grid is the issue.

So, in that case, there are people who are watching who are investors and want to understand this. We talk about infrastructure picks and shovels for AI, and we talk about data-center construction companies and all of that. Who are the companies that are building the grid? Is there a set of work orders and purchase orders for building out a more robust grid?

It's a really good question. The grid—the poles and wires, the distribution grid in particular—is dominated by regulatory monopolies, the local utilities. I'm not going to comment on their current prospects, whether I think those stocks are buys or sells, but the regional monopoly utilities stand to make a huge amount from this in the areas where data centers can be built.

There's a separate issue, not in my slides, of more and more voters pushing back and saying, “We want to stop data centers from being built.” There's a lot of psychology behind that. I don't think the reasons are necessarily valid.

Even in Texas yesterday—maybe it was the day before—Governor Abbott sent out a letter pausing data-center requests.

6. The potential of large-scale solar projects and their costs

Oh, no. Not them, too.

It wasn't quite a pause. It's an audit of all data-center requests in Texas. In Texas—a red state, the most libertarian state in the country—it's political cover. Abbott knows that his voters have an anti-tech sentiment that translates to AI data centers because they're an obvious target. So, he wants those data centers built, but there's an election coming up. He's got to cover his ass for a bit by making it look like he's serious about this. That's the politics right now in the country.

So, basically, we're doomed. [laughter]

I don't think we're doomed. We will have space. We'll always have sun-synchronous orbit.

We will have sun. [laughter]

The AI doomers would say, “Thank God we're saved. The AI god won't be built.”

And even if not, they'd propose orbital bombardment of the data centers. There's no way of winning. [laughter]

Let me talk about the grid. Look, I'm going to show you a lot of sci-fi stuff and awesome stuff, but, Peter, what you're saying—the grid itself, the poles and wires—are the limit.

I've talked for years about the exponentials in solar and batteries. We'll talk about fission and fusion, but the poles and wires have thus far not become an exponential technology. That's something I would love to solve. I have not seen a lot of startups in that space.

So, I mean, aren't we moving the data centers to where the energy is, so you don't need to set up—

Or disconnecting them from the grid entirely?

Exactly.

Let's move on, and I'll get to that. This is a more practical forecast. You see, even by 2028, we'll build—this is probably a little bit low. The orange is how much we'll build. Maybe it'll be 20 or 30 gigawatts, whereas demand could be much higher.

This is an interesting slice. By the way, I'll tell you, every forecaster—Morgan Stanley, whoever—differs somewhat. But this is an interesting slice: the blue bar is if you sum up all the GPU manufacturing scheduled between now and 2030, primarily NVIDIA, but also AMD, Cerebras, whoever, versus the expected pace of U.S. grid build-out. The chips are more than twice the pace of grid build-out in terms of their power draw.

Put simply, the AI demand, which is chip-limited, shows roughly 200 to 275—call it 230—gigawatts of power demand based on the chips. You bought the chips, you've installed the chips—can you power the chips? There's 230 gigawatts of demand there, and U.S. grid build-out is projected at roughly 100 gigawatts.

7. The role of nuclear fission and fusion in future energy

Some of you might remember that, about 6 months ago, Satya Nadella, the CEO of Microsoft, made this comment: “Look, man, warm shells are our limit. We've bought the chips. We don't have warm shells to put them in.”

Right. That is the limit for everyone at this moment.

Yeah. Why is that discrepancy there? Eric Schmidt told us his number was 100 gigawatts, or 96 gigawatts, of additional power by 2030. The 230 is just based on chip manufacturing. So, either more of the chips are being kept domestic—which wouldn't surprise me—or the fabs ramped up, which would surprise me. Where does that discrepancy come from?

Every single forecaster has a different number, and I think some of them base their estimates just on announcements by companies, whether they're chip fabs or utilities. Some of them base them on their discounted projections of what they can actually achieve.

Also, I will say there's a big miss in power demand from chips. A lot of people just say, “How much power can my Blackwell GPU draw?” Multiply that by how many you're going to build, and that's the power demand.

No, you're missing almost half the power because you have to add the draw of the rest of the IT equipment in the data center, cooling, and so on. That nearly doubles the total power use.

You know, that would make sense. Those Cerebras chips just suck down power, and they run the transistors much more efficiently than the prior-generation A100s and H100s from NVIDIA. The transistors are actually doing a lot more work, which is fundamentally better, but of course that's going to draw more power constantly. And when you buy those things and deploy them, you run them 24/7.

That's right. You're never going to let those things rest. So that might be a discrepancy, too.

Absolutely. Do you think this creates a forcing function, perhaps, for NVIDIA or the other fabless vendors, or for the fabs like TSMC, to get into the power-generation business? Right now, the power generation that's supposed to power all of these chips that Satya talks about is just collecting dust in warehouses because he can't find warm frames for them in data centers. Why do you think there's a forcing function for the NVIDIAs of the world to get into power generation?

Well, I'd say, look, whether NVIDIA wants to get into it or not—and NVIDIA has made some interesting investments that I'll talk about in grid flexibility—the reality is that what people talk about the most now is behind-the-meter power generation for data centers. What they mean by that is large natural-gas turbines, if they can get them. This is a multihundred-megawatt—say, a 400-megawatt—natural-gas turbine, the kind you'd use on the grid. These are now sold out for something like 7 years.

GE, Hitachi, and so on are building new assembly lines to try to bring those online faster, but everyone is saying, “Look, if the grid is going to make me wait years and years and years, I'm just going to build my own power.” This is more expensive than the grid, but power is such a small fraction of AI costs. Maybe you can do it because these guys are sold out.

People are going to these small turbines. Solar Turbines—nothing to do with solar—makes a 38-megawatt turbine that's on the back of a semi. So 40 of those make a gigawatt, right? Even these have backlogs. Now, everyone in the world that was in any way proximate to gas turbines is pivoting into this space.

I'll give you an example: Boom Supersonic, a very cool company trying to make supersonic jetliners a thing again. That's a very hard task, with many, many billions of dollars of regulatory costs. They have pivoted into using their engine design to make a natural-gas turbine for data-center power because the demand for this is so very high.

So, modular energy production, right? How many of these—if you think of them as 18-wheeler trucks that have a large container on the back—can you just pull in and get your data center started until you build out the energy infrastructure, and then move them on?

This is how Elon got the Colossus data centers up, which Anthropic is now leasing. Actually, this is what he did.

And we've talked a bit about this on the pod in the past. We talked about the Boom pivot, and we've talked a bit about Elon standing up his power-generating cogeneration facilities at Colossus, et cetera. We talked a bit about that, but I want to try pressing once more on this point.

If this thesis is true—that this is a primary overhang on NVIDIA's ability to sell more GPUs—NVIDIA is already doing all sorts of financial engineering to be able to sell more and more GPUs through customer financing and all of these other things. Why on earth, if the energy overhang—or underhang, depending on your perspective—is a major limiting factor for the ability to productively monetize GPUs, don't we see NVIDIA doing something on the energy front?

It's a great question. For behind-the-meter power, the financial incentives are so large that NVIDIA doesn't have to, but they might invest in some of these companies. On the grid side, NVIDIA has made investments into increasing grid flexibility to be able to get more juice out of the current grid. Emerald AI is one example. They've made a few investments in this space, and I'll talk about grid flexibility in a second.

But the real issue is a combination of regulation and the incentives for utilities. Monopoly utilities in the United States, the bulk of them, are paid on a cost-plus basis. They go to the utility commission and say, “I've got a plan to meet the demand that I see my customers having. Here's what it costs for me, and I expect a 10% return on capital for it.”

The utility commission mostly just says, “Okay.” Some are better than others, but let's be honest: the utility has enormously more horsepower in people, computing, salaries, and so on than the utility commission. They jam through this plan and get 10% on top.

I see. So, if I were to try to synthesize what I think your answer is, your answer for why NVIDIA isn't getting into bundling power generation with its GPUs is that it's low-margin and frictionful. It's the same reason NVIDIA tried and failed, then retreated, to launch its own hyperscaler or neocloud: it's just not as high-margin as selling GPUs. NVIDIA might still be a neocloud; we can talk about that separately.

If I were NVIDIA, I would be focused on changing the regulatory landscape for monopoly utilities. I've said this on some utility-specific podcasts. We should change the incentives. Utilities, instead of just getting paid a percentage over capex, should be paid on things like how fast they can deliver power. Their executives should get bonuses for delivering power fast, and their employees, obviously, all the way down. If you did that, suddenly these things would happen faster, right?

You get what you incentivize. Absolutely. Yeah, look, I invest in startups. How many startups have I seen that have a technology to speed up building poles and wires? I don't know—2 or 3. None that I thought were amazing. If you, by the way, listeners, have one, please send it to me. Why not? There's no incentive for it. But if you created the incentive, people would find technical solutions to speed that process.

Is there any state—not Texas—but is there any other state that's open-minded about that?

Texas. Texas is the best. Despite what I just showed you, Texas has made policy changes that accelerate this. People are not totally asleep at the wheel.

Texas's ERCOT is its own fiefdom and is not regulated by the federal government at all. FERC regulates the rest of the country's electricity. FERC has sent letters to the 6 other largest grids saying, basically, “Do something like what Texas is doing.”

What they're doing—and maybe I can just skip to it—is making new regulations that say if you are an interruptible load, if you are flexible, if you can either find some alternate way to power yourself or just turn down your power at moments of peak demand, we'll get you connected much, much faster. In Texas, that's a CLR or a PCLR, an interruptible load.

The reason for that is that we have very high demands for the reliability of our grid. Right? 99.9% uptime is 8 hours of outages per year. That's unacceptable. You've got to push to four nines to make it a grid that you think is really good.

But the nature of the grid is that power demand is not constant. It fluctuates through the course of the day and the seasons. It peaks primarily in the South in the late-summer afternoon. The U.S. grid averages about 500 gigawatts of demand throughout the year and throughout the day. It's much more volatile than that, but at any given time, during winter nighttime, the U.S. grid is down to around 400 gigawatts of power being drawn. In the summer, in the late afternoon, we're up to around 600 gigawatts of power being drawn because of air conditioning, primarily. The fluctuation is actually much higher than this.

Europe doesn't have this problem. I'm joking.

It's a different problem. We can talk about Europe and air conditioning. There's some amazing tech coming down the pipe on that. By the way, hopefully a new investment.

That gap is 200 gigawatts, right? 200 gigawatts is about 10 trillion in AI capex. We think there's about 7 trillion in AI capex in the next 5 years. This is no joke. If we just used the poles and wires more efficiently, we could power up a lot of stuff because we're not short on generation. We're not short on power plants. We are short on capacity and the poles and wires.

As I mentioned, Texas just enacted this new regulatory change in June that says, “Look, if you don't need to draw power at peak, we'll just hook you up fast.” Instead of 5 or 7 years, it might be 12 to 18 months. FERC has now told everybody else to do that.

So how do you do that? This is a paper by a buddy of mine, Tyler Norris. He's now at Google; he was not when he wrote this. This came out in January or February of this year. This is the best electricity-related paper of the year, in my mind.

Basically, what he found was—I call it 200 by 200, or 100 by 100 at minimum—that if you can be flexible for 100 hours out of the year, 4 days out of the year, or 1% downtime, that unlocks 100 gigawatts of capacity on the grid. That's about 5 trillion in data-center capex, including the chips, which gets you through the next few years. That's one way to do it: just flexibility.

The startup I mentioned, Emerald AI, was founded by Varun Sivaram and funded by NVIDIA. They do this via software orchestration, moving jobs to the right data center, and so on.

But there's another way to do this, which is batteries. This is a portfolio company of mine. I've made 3 investments in this same startup, maybe a 4th one coming up. They do something really obvious in a place like Dallas–Fort Worth: between the middle of the night and late afternoon, there's like 10–15 GW of flexibility in grid demand.

So, if you build out, let's say, 4 hours of battery storage at the site and fill it up at midnight, you don't need to hit the grid during the peak of the day. That fits perfectly with the new Texas regulations. In fact, they were leaders in driving this.

8. The analogy of cache and multitasking in energy management

This currently sounds obvious to us, right? But this is an unusual approach. 12 months from now, this will be a super-common approach, not just—

In shifting load, right?

Yeah, exactly. So right now—

But if I have a magical technology that stores insane amounts of energy very cheaply, and I go even to Texas and say, “Hey, this can completely shift this curve. This is a total game changer. Can I hook it up to the grid and start sucking down power when no one's using it in the middle of the night?” Would they still say, “Yeah, you can do that in 2030”?

The new regulations that were just passed in June give a fast path to power for anyone that is an interruptible load. So long as the grid operator is able to turn you off, it's not them saying—

Definitely. We need to make T-shirts that say, “I am an interruptible load.”

Oh my gosh, it makes me want to show an Abundance T-shirt that Peter's team sent me. But yes, I am an interruptible load. Don't ask my girlfriend's office. [laughter]

So, look, why isn't every data center deploying these giant battery packs? It seems like if I had that in my data center, I would be super-smooth on the load demand for my community.

We passed this regulation in Texas in June.

Okay.

The second week of June, 2 months ago.

So it's brand-new.

Yeah. I invested in these guys because they drove the regulation and because they've got 10 GW of sites that can take advantage of this. After this was passed in Texas, FERC, the federal regulator of electricity, sent a letter to the 6 largest other grids in the country—not specifying the details, but saying, “Do something like this.”

Figure this out.

So this is going to become a very common thing to do. It's called Agent Gentic Infrastructure.

Agent Gentic is the startup, but this in general is an interruptible load, or time-shifting demand. Again, like that red dashed line—not all of you, since some of you are just listening—the transmission-line capacity, the substation, and the transformers: that's the limit. It's not the gas generators or the solar or wind. It's the transmission line.

So, if you can use batteries to fill up your data-center batteries at night, when the transmission line is unused, and then not need to draw on the transmission during the day, we've always known that was a good idea. We do it with EVs and so on. This is a very big deal.

Amazing. Would you say it's fair to characterize this as the energy, or grid, equivalent of preemptive multitasking or reentrant multitasking in computing? Basically, allowing processes to say they can be paused and their compute load can be time-shifted.

Yeah, I think that's one way to look at it. I think that's a great analogy, Alex. It's also like cache prefill. Batteries are a cache for electrons instead of data, so we're making our grid cacheable.

That's right. By the way—

Go ahead.

How much of that big gap—the 200 GW—do you think we can make a dent in by taking this approach?

I think approaches like this and approaches like electric vehicles also, right? The bulk of the batteries in the U.S. are actually in EVs.

Another company of mine, WeaveGrid—I shouldn't say “of mine.” I'm blessed to be an investor in them because they're smarter than I am—for a long time, they've managed electric-vehicle charging for utilities to reduce stress on the last mile, even on the last block. The limit on EV charging for the grid is actually the transformer on your block because Teslas cluster. If one person gets a Tesla, their neighbor's odds of getting a Tesla nearly double, right?

They already have software to time-slice and even out the charging of the vehicles. Companies like that—in particular, WeaveGrid—are using that technology to make the other loads on the grid more responsive and shaped in a way that allows AI data centers to play well. In fact, every EV-charging company I know has pivoted to trying to use its technology or current capacity to enable data centers.

I think that's 100 GW. If we're smart about it, that's the next 5 years of AI data-center growth.

Amazing. All right, what's next?

All right, let's talk about more interesting stuff. We all love solar. We're entering the phase where solar-powered AI data centers become viable.

Many people have seen a chart like this. I've shown that in 1975, 1 watt of solar panels cost $100. Now it's $0.08 from China for a panel that's smaller, has a longer lifetime, is more durable, and so on. That more-than-1,000× price decline—does that get us to the point where we can power data centers with it?

Well, because the chips are so expensive, it never makes economic sense to only run data centers when the sun shines. So you have to have storage as well. Battery prices have dropped by a factor of 14 since 2010. We've had new technologies; lithium-ion has been dominant, but sodium is much more common on planet Earth than lithium. New technologies like sodium-ion batteries could drop the cost of batteries by a factor of 10.

Even now, we already have the very first solar-plus-battery baseload power plants, and they're affordable. We have them in the UAE outside of Dubai, and we have them in Chile. A nice thing about this is that natural-gas turbines are sold out for years. The fastest energy project you can build is a solar-and-battery project. You can get that done in 12 months.

So, in the United Arab Emirates, this is a 1 GW, 24/7 solar-and-battery project. What that means is they guarantee that the minimum power output at any time is 1 GW. To do that, it's actually 5 GW of solar and 19 GWh of batteries. The cost is about $6 per watt in capex. That won't mean a lot to many people, but let's just say the last nuclear power plant built in the U.S. cost $15 per watt. The cheapest ones on planet Earth are Chinese, being built in China, and those are $4 per watt.

So, recently competitive.

It is recently competitive.

So, Ramez, you and I texted about this, right? On the last earnings call at Tesla—I think it was at Tesla—Elon said he wanted to build out 100 GW of solar capacity. Did you check into that?

Yeah. I think, look, it's a long-term vision. It's not next year, but Elon's overall vision is, “Let's put all the compute in space.” There are no land constraints. There are no permitting issues there. People won't complain about water use there. He wants to build a terawatt of AI.

If you're going to build a terawatt of AI, you've got 2 or 3 options, really: the world's deserts powered by solar and batteries; getting fission or fusion to work; ocean power like Panas [?] that I'll show; or space.

So, he's talking about Tesla building out terrestrial solar, right? Competing with China.

Yeah. He wants to build out the manufacturing for it, but I think his real motivation is not selling it to the on-land market in the U.S. I think his real motivation is to build that manufacturing capacity for space-based solar.

Just to pull on Peter's question a bit: if we take the 1,000× reduction per kilowatt or megawatt over the past few decades and extrapolate it, have you gone through the thought experiment of what solar would need to look like in order to achieve another 1,000× price-per-watt reduction?

Yes. This is a very good question, and it's an important clarification of how the cost reductions work. Our best model—I'm not that smart, right? I'm one of the top 5 forecasters of solar costs in the world. It's not because I'm that smart; it's because I came out of tech, came from a Moore's Law world, came into energy, and just applied Moore's Law to it.

But when you actually look at the details, it's not a reduction with time. It's a reduction with cumulative scale. It is Wright's law. It's the learning rate. So every cumulative doubling of solar scale reduces costs by, let's say, 30%. It fluctuates year to year; it's the real world, and so on.

If we ignore the possibility that we need terawatts of AI compute and just look at the world as it is, solar is now 8% of global electricity. Let's say we think solar can get to a third or two-thirds, and maybe electricity demand goes up by a factor of 2. You've got 4, 5, or 6 doublings left. That means the cost of solar might drop by a factor of 4, maybe even by a factor of 8, but not by a factor of 1,000.

But if you start talking about building Dyson spheres, then we have a long way to go to keep reducing those costs.

I heard what I wanted to hear. [laughter]

You heard “Dyson spheres.”

Pandering, pandering.

[Laughter.] Bingo. Wait, can I drill in on that, too?

I've got a couple of questions. Yeah, Ramez, how many of these installations are being built around the world right now? This exact style of monster-scale solar at scale.

We're just—oh, like gigawatt scale? A handful, largely in China, the Middle East, and some in LatAm. We have maybe more than a handful, maybe a dozen, at this scale. Most solar plants today are typically somewhere between 50 megawatts and a few hundred megawatts.

9. The economics of GPUs, AI, and energy costs

A gigawatt plant is challenging, and the biggest reason this is not yet an option for the U.S. is not cost. We could pull off something like this in the Southwest, and it would actually be cheap. It would be more expensive than it is in the Emirates because our labor costs are higher, but it would be fast. You could have it done in a year. Even with a natural-gas turbine that you want to order from GE, you can't do that. But putting together the land parcels is actually the pain for this in the U.S.

Are these solar panels coming from China?

Probably. Eighty-five percent do, presumably, and of course we double the price of Chinese solar panels in the U.S. So we hurt ourselves by keeping them out.

Sorry. Yeah, so if this is the fastest path to energy at scale, why aren't there people—or the U.S. government—just going, “Yeah, let's use eminent domain, grab whatever chunks of land we need to, and build this stuff,” because we could be done in a year?

You don't need eminent domain. The federal government is the number-one landowner west of the Mississippi. Those federal lands are concentrated in places like Nevada and Arizona, places that have enormous solar resources. But it's not something that interests the current administration, I would say.

If I was thinking about it, I'd be thinking about how we open up lands that are not amazing nature resources to build solar-powered data centers. And I think we get them. Sunshine seems better than drilling on federal land.

Absolutely. And I'll say this also: regulations are the problem in lots of places. I was in Mexico recently. I was in Chihuahua, trying to convince the government of Chihuahua, a state of Mexico, to build a lot of solar-powered AI data centers.

But in Chihuahua, it is actually illegal to have private power generation—behind-the-meter power—above, I think it was 500 kilowatts, right? Half a megawatt. So you just—by law, you can't do it.

And then, secondly, the AI labs and the hyperscalers are extremely vigilant about data protections. They don't want their user data leaked or seized, and they especially don't want their model weights exfiltrated. So they're pretty careful about the countries they go into.

My advice to Mexico was, look, change the laws to make it possible to build this sort of thing, and provide ironclad guarantees of the protection—intellectual-property protection—of this data, and you've got an enormous business: more open land, lower population density, and better sun than the U.S.

I was going to push back on—well, first things first: the GPUs can't sit idle, no matter what, because they're so expensive. But when you look at the underlying economics, about 5% of the cost of the data center, maybe up to 10%, is the power.

But the GPU itself is an 80% markup from NVIDIA on top of a 2× markup from TSMC, with another 2× markup at the model-provider level. So it's actually 20× overpriced relative to the cost of turning sand into a chip, which is actually coming down, too, with efficiency and scale.

And so, at the fundamental level where Elon thinks, it's actually not a given that the GPU is super expensive relative to the power once he gets the Terafab up and running and the end-to-end sand-in-one-side, chip-out-the-other is fully automated. So that would completely flip all the math in this if he gets to that destination.

10. Long-term prospects for solar, nuclear, and space-based energy

I think those are awesome comments, Dave, and I think it's right that GPUs are overpriced, or at least that there's a lot of margin going in there. NVIDIA—people don't think of them as a network-effect company, but CUDA, the programming layer to write to AI, is their moat, right? It's not like their chips are that good. Their chips are fine. AMD's chips are as good. Their interconnection between chips is great, and that does matter. People like Huawei are kind of getting there, honestly.

But CUDA has been the moat, and I think the CUDA moat is broken this year and next year. One of my portfolio companies, Lamorian—I met them at Abundance 360—is working on that. But also, now that you can tell AI, “Take my AI code and recompile it to run really fast on this AMD chip or the Cerebras chip,” I think NVIDIA's lead is—

So glad you brought that up. This is such an important topic because this $5 trillion of U.S. market cap is hanging in the balance of this conversation. It's such an important and fragile thing, you know. CUDA is the moat, for sure. No doubt. All the AI researchers are too lazy to write custom kernels. Suddenly, Fable 5 comes along. I've had great luck running custom kernels myself just in the last couple of weeks using Fable 5, so I think your prediction is probably right. I don't see why it wouldn't be right.

I think NVIDIA would say, “We have all kinds of other network effects, and we have massive interconnect.” Now, the interconnect is incredibly important for training.

Yeah.

But 90% to 95% of the load now is moving to inference, where you don't really need the interconnect.

I think the interconnect is still very helpful for inference. If you're going to run a model like Kimi K2 or DeepSeek—not necessarily Flash, but the next DeepSeek V4—you're simultaneously running it on a rack, right? You're running it on 10 to 20 GPUs at a time. So the interconnect does matter somewhat, even for inference. But you're right that it matters even more, tremendously more, for training.

Dave's point is just that, at inference time, interconnect locally matters to the extent you need local coherence, but you no longer need global coherence at the level of an entire supercluster. It's just like a single rack of coherence.

Yeah, that's correct.

Coming back to energy—and to summarize this—there's plenty of room for energy growth. We have the abundance thesis on energy writ large with solar, right? And one of the points that Elon has made before is, at the end of the day, it's all about solar. Do you believe that thesis, that solar in the long term is going to dominate beyond everything else?

It's complicated, and I think we underestimate the importance of geography. So the reality is, look, from a regulatory standpoint, we're not building transmission. There is no place outside of China that is building enough capacity to move electrons from place to place. Same thing as what I just showed with poles and wires.

And so the problem for solar is not cost, and it's not nighttime, because batteries are plunging in cost and are going to drop another 10× ultimately. It is winter. So in London, for instance, you get 1/6 or 1/7 as much insolation in January as you do in June or July. So you're going to build out your solar plant by a factor of 6 or 7?

No.

Or do we have a battery technology that can store months of power? There are a couple of interesting ideas out there, but think about the unit cost of electricity that goes into a battery. It's basically battery capex amortized by how many times it gets used.

So if you have a battery that cycles daily, it's battery cost—capex divided by 365, or 3,650 if it's 10 years, let's say. If the battery gets used twice to shift load between seasons, it's battery capex divided by that.

There are numerous startups. There are some crazy ideas: sand, compressed air, power-to-natural-gas, yada yada yada. But right now, it's clear to me that economically shifting energy through the day-night cycle—we're not totally there yet, but the curves are just heading that way.

But dealing with winter, especially—we have not yet electrified heat. If you look at the U.K. as an example, and at Northern Europe, if you go from burning natural gas for building heat to using heat pumps, electricity demand doubles in winter. So we have this big, big, big winter problem that I think a lot of people are not reckoning with.

And so I do believe nuclear is super useful, as are efforts to get seasonal storage, as are all of our efforts on fusion, as is advanced geothermal—especially for those places that are farther from the equator and have either long winters or long rainy seasons.

Location, location, location.

Yeah. All right. So this last slide is saying that solar, batteries, and data centers are going to get cheaper and cheaper. So, yes, one of you asked, “Why don’t we move compute to where the energy is?” I fully believe that. With other energy loads, you can’t move the population of New York City to a place that’s sunny year-round—not quickly.

Almost—it’s Florida.

Okay, it’s Miami during COVID, and mostly for crypto folks. But you can move the new load we haven’t built, which is AI. Why don’t we site it where the energy is? So that’s one viewpoint. All right. Next, Peter, you wanted me to talk about fission and fusion, and both are super exciting.

Absolutely. Yeah. We hear a lot about it. We talk about it. We speak about the hyperscalers turning on defunct fission plants and investing in fusion companies.

It’s interesting. A quick stat: 71% of Americans are against data centers, which is a higher percentage than are against a nuclear plant in their backyard, which I find amazing.

That’s just insane. I’d rent my backyard out to both. It’s not big enough or quiet, but we’ll make some room.

Maybe just a quick question before the segue. I want to pull a little bit on the historic rhyme between the Middle East being a major source of oil but now also being a major source of solar power. The thought experiment I’ve done—I’d be curious to get your thoughts—is that the reason the Middle East has so much oil is, my understanding is, hundreds of millions of years ago there used to be a warm ocean with lots of plankton and other small creatures that ultimately resulted in the oil. Now it’s largely desert, but it’s still pretty warm.

Any thoughts on the historic rhyme between why somehow the Middle East is, on the one hand, a supplier of all this oil power for data centers and now potentially solar power?

Well, I think the Middle East has amazing solar resources, but it’s not as lumpy as their fossil-fuel resources, and especially their oil resources. If you look around the globe, you have Australia. I mean, if I was thinking—actually, I just said Mexico—if I was thinking about doing a lot of solar- and battery-powered data centers for AI, I would be really pushing in Australia.

Yeah. You’ve got a friendly government, enormous amounts of space, and some of the world’s best solar resources. Chile and Mexico—not amazing oil producers. Mexico was once; they’re not anymore—but they have solar resources equivalent to the Middle East.

But you’ve got to change the laws such that an OpenAI, an Anthropic, a Groq, whoever, a Google, is comfortable siting their crown jewels in your country.

Just by way of a reference number for the audience, when I last looked at it, on the energy-abundance thesis, we have 8,000 times more energy hitting the surface of the Earth than we consume as a species in a year, right? So there’s plenty of energy out there. It’s just not in usable form. The whole conversation here is: How do we take that energy that’s latent and make it usable? And that’s the role of technology.

That’s right. The other commentary is that fossil fuels are just an old battery that we’ve been using up, right? And we’ve used up about 25% or 30% of that battery.

No one knows. But the cure for high prices is high prices. So if we ever started to run low, there’d be more incentive to explore and find stuff.

I have a question for you, Ramez, on the oil market, real quick. I remember you commenting once that the 2013 oil-price crisis—the oil crash—

—was because of a 2% oversupply in the market. It’s a really tightly wound market.

Is that still the case, or with all the Middle East conflict and everything else, are we now in tension? Is it going to stay that way?

Oh my gosh. There’s a lot to say about that. There are 2 interesting things about the Iran war, its impact on oil prices, and why it’s been relatively muted.

Number 1: China did us all a solid. China built the world’s largest oil strategic reserve, right? And they were willing to drain it during this period. So China has more than the rest of the world’s strategic reserves combined and has helped keep oil prices low.

But 2, the ratio of global GDP to global spending on oil has roughly tripled or quadrupled since the oil crisis of the 1970s. There are critical things that are highly dependent upon oil—aviation, shipping, trucking, and so on. But overall, the world has moved to more of a services economy, and that has created more demand elasticity. It has allowed the world to deal with a shortfall in oil in a way that we couldn’t in the 1970s, when our economy was more physical.

I just like to pull in the geopolitical angle here. There’s a theory in certain circles that an ulterior motive for the Venezuelan operation and then the war with Iran was actually to cut off China’s [?] in the event of a Chinese invasion of Taiwan—open PN [?], TSMC closed PN [?]—so that China would require backup oil supplies because it would get embargoed by the Western bloc. Its go-to sources for backup oil in such an invasion would look like Venezuela, Iran, maybe Cuba.

The full theory here is that the recent military adventures we’ve seen—Venezuela and Iran—are actually, at some level, a play to deter China from invading Taiwan to gain access to the TSMC fabs and basically the future light code of AI. Any thoughts on that?

There’s insight there, but I don’t agree with it as stated. The reason for that is China bought oil from Iran during this war. The U.S. still sells oil to China. It just wasn’t that planned out.

The actual DoD war plans in a situation like that are to use the U.S. submarine fleet to sink tankers that are heading to China. That’s the actual proposal of what to do. Who knows if that’s a good idea? I’m not going to get into that right now.

Oil is mostly fungible. So the fact that China buys oil from Iran, the U.S. bombing Iran, or even if we successfully close the Strait of Hormuz doesn’t really hurt China that much, because it can buy cargoes from somewhere else. It was getting a discount from Iran because it was embargoed oil that China was willing to buy. So it has to pay a few bucks more per barrel. It’s not that big a deal to China.

In wartime, it’s a kinetic sanction. It’s a kinetic embargo, which is a different sort. And, yeah, the sims—I think this is all totally unclassified—the simulations of wars like that are U.S. submarines torpedoing tankers that are heading to China.

Onward to the horizon of fission and fusion.

Okay. Let’s talk about the atom and the power thereof, going back to the 1950s—or a retro-future. This is a complicated slide. For those of you who are just listening, basically, there are 2 approaches to nuclear fission, which is what we’ve been doing since the 1960s.

The traditional one is big reactors, and the simplest thing you can do to boost nuclear production worldwide is to stop shutting down nuclear plants. Germany should not have shut down its 2 nuclear plants. They were end-of-life plants, and we can usually extend them. In some cases, we can actually upgrade them to produce more power.

3. There are some plants that have been shut down, like Three Mile Island, that we can actually restart safely. But that gets you a few gigawatts, right? If we really want a nuclear renaissance, the core issue with nuclear fission today is that, outside of China and perhaps South Korea, it is ruinously expensive.

And why is it ruinously expensive? It’s because we don’t do a lot of it. Anything that you do infrequently is expensive, right? You don’t get good at the things you do occasionally. The things that are cheap are the things that you do repeatedly, again and again and again.

So there are 2 paths happening to bolster nuclear in the U.S. I’m a critic of this administration on many fronts and on some energy fronts, but I’d say this administration has the best nuclear policies of any in recent history. Still missing some things, I think, but the best that we’ve seen. The left side of this is large reactors.

We have this thing called the AP-1000. It’s sort of the Westinghouse reactor. It’s sort of a workhorse reactor. We’ve built a couple of them in the West, let’s say 4. China took a variant of this design, made its own variant with a local supply chain, and has built more than a dozen.

They built them at higher power than this: 1.4 gigawatts instead of 1 gigawatt. So one plan is to produce a process to get more of these built. The administration has created structures for loan guarantees, financing, and so on, because if you stamp out a lot of these, the cost should come down. Right now—

What plant is the most stamped out so far?

Light-water reactors like those used in France. France is the poster child. The US generates the most nuclear electricity of any country on Earth. Actually, it’s not really known. China is building the most right now. France gets the highest fraction of its electricity from nuclear, and they basically—with slight caveats—took the same design and stamped it out again and again and again.

How many does France have, like 60 nuclear power plants now?

Something on that order. Yep. But even France—

Something like 80% of their electricity is nuclear. It’s crazy. And they exported it to the rest of the Eurozone as well. Nuclear energy was basically born in France. Thank you, Curies.

Yeah, but even France is struggling, right? There’s something called the European Pressurized Reactor, which is mostly a French design, and that thing is kind of a disaster right now. It’s a boondoggle, running over budget and going slow.

Again, if you take one design and do a lot of it, it gets cheap. But usually—and this is critical for the sector as an investor—the first one runs over budget and over time and has problems you didn’t anticipate. So if you want a thriving nuclear industry, you just have to know that the first one you build of a new model is probably going to have problems.

But after you’ve built 3, 4, 5, maybe more, you sort out those problems. You build experience in the crews. You build experience in the engineers. You sort out design issues. You build a supply chain to provide the parts that you need.

So one plan is to take the large reactors that we’ve built a couple of times. Now the US government has created financing, sort of a backstop—loan guarantees—for about 8 of these. Another startup, a nuclear company whose founder’s previous company I invested in, has a plan to basically build fleets of these, because that’s how you have to finance it. You can’t finance one because you know you’re going to miss your targets. But if you can finance a bunch at a time, you can amortize the cost.

You know, in China, after they got to 6, 7, 8, the costs had really come down and stabilized. So those who are fearful about nuclear—and it’s still probably a good percentage—they think about Three Mile Island and Fukushima. We’re talking about early-generation plants, right? Are those Gen 1 or Gen 2 plants?

Something like that. These are Gen 3, Gen 3 Plus, or Gen 4 plants. One of the most important things to understand about them is that basically all of these are passively safe. What that means is you can knock out the power to them and they won’t have a meltdown.

Fukushima happened because you circulate water around the nuclear core to take the heat away from it and then use it to turn a steam turbine. The pump for that was powered by grid electricity. So the tsunami that hit Fukushima knocked out the power lines, and the pumps stopped working, even though there was power right next to them for the nuclear reactor. New nuclear fission designs are passively safe.

Would you call them fail-safe plants?

Nothing is totally fail-safe, but they’re designed to take a 747 crashing into them and the power going out from the grid and keep operating without any meltdown.

Ramez, I’m curious, just pulling on that. What happened in the 1970s? I assume you’ve seen the television show For All Mankind. It’s sort of an alternative historical reality where we get fission and it never gets abandoned. What happened in the ’70s? Did we just waste the past 50 years not building enough nuclear energy—fission in particular—and find ourselves in a suboptimal future?

I don’t think so exactly. I think we could have done better, but people are somewhat risk-averse. We learned in the 1960s that radiation causes cancer. The radiation released from a well-operating nuclear plant is really minimal; it’s unlikely to cause cancer.

But we did increase the regulatory state. We did increase the burden of proving that things were safe, and that had some cost. Then things just fizzled out, and we talk a lot about flywheels and positive feedback loops. We had a negative feedback loop once, and this is the same thing for fission: once the industry is no longer building, you lose the expertise, you lose the supply chain that makes the parts that you need, and you lose the manufacturing facilities. Everything gets more expensive. So if you’re not constantly scaling, you are going to backslide, basically.

So those were large reactors, and you have on this slide here the small modular reactors because—

The right side—and this is the area that investors are super stoked about, that I was stoked about 15 years ago, got less stoked about, and now I’m becoming maybe hopeful again—is what we call small modular reactors.

I talk a lot about learning rates, right? That’s how fast something gets cheap. Every technology, if you build more of it, gets cheap at some rate as the scale increases. But the things that get cheapest the fastest are those that are built in factories in high volume and have the smallest number of moving parts.

So the idea of SMRs is to build as much of this in modular, repeatable, factory-built situations as possible and do as little stick-building, as little assembly or construction, as possible. Construction is a dirty word, right? Construction does not get cheaper. Manufacturing gets cheaper. So move as much of this as we can to a manufacturing process.

At the limit, it’s a factory that spits out nuclear reactors that you just barge or semi-truck to a location. Many of these are parts made in a factory in a standardized way that you assemble like Legos on-site. This is an incredibly sexy space for investors right now.

Yesterday, we found out that Valar Atomics raised $1 billion at a $6 billion valuation for an SMR startup that doesn’t have a working reactor. Other companies—Aalo Atomics is probably my favorite company in the space—but there are a ton of companies.

My friend’s company at X-energy went public recently.

X-energy is an amazing company, actually. I really like their design. Radiant is one on here that’s on a very, very small scale.

There’s a line between an SMR and a microreactor. Can you make it small enough to fit in a shipping container? The military, for instance—you’ve got Radiant on this slide—the US military, for military bases, would like shipping-container- or half-shipping-container-sized reactors to power bases in the US, but maybe in forward-deployed locations as well, so they don’t have to move fuel. There’s a lot happening in this space.

So Natrium, I see on the chart here, is a third of a gigawatt—

Compared to the AP-1000, which is roughly a gigawatt.

When you say, if Natrium goes into mass production—mass production being tens, 50 units—I mean, the relative price of buying 3 of those Natrium units versus an AP-1000, what’s that? Are there economies to bigger plants, or do you stack them together?

Yeah, there are economies to bigger plants. Bigger plants use less steel and less cement per unit of power output. So there are economies to bigger plants, and that’s how we used to think in the 1950s and 1960s, primarily.

But there’s learning that happens from building more plants and doing more of it in a factory. Personally, my guess is the BWRX-300 and Natrium are in an awkward middle, because they’re not really factory-built. They build a bunch of components in the factory to do field assembly, and so I worry about them, but they might end up being the ideal, optimal solution.

At the other end of the spectrum, you’ve got Aalo here. Their reactor is 10 megawatts, so it’s 1/100th the size of an AP-1000. They put them together in clusters of 5. That’s a pod for 50 megawatts. They have lower efficiency in their use of steel and cement, but they can build it entirely inside a factory.

Which is the first one of these coming online, do you think? These are not online yet; these are all theoretical.

Yeah. The optimistic projections from these companies are 2030 to the early 2030s, and that’s for the small ones. The next AP-1000 is probably a few years later than that. These projections will probably be missed. I don’t expect anyone to actually have a commercial small modular reactor in 2030 or 2031, but maybe not. The size of the slip is probably smaller for a small reactor.

Like everything else we’ve been saying, the first units here are not going to be cheap. The first years are going to be expensive. So the key is to build an order book from a customer that believes that by ordering enough, they’re going to drive down the price, or to build a multicustomer order book where you’ve built some mechanism for cost and risk sharing between these AI data centers. And what we talk about is: Is nuclear the solution for AI data centers?

Maybe there are other ways to power AI. AI data centers might be the best thing that's ever happened to the nuclear industry.

I'm curious: if we just take this argument to its extreme, where are the nanoreactors? Why don't we see 100-kilowatt nano- or pico-reactors that can be colocated with every GPU? Do you think there's an opportunity in that space?

I think it's really hard. You do hit some economies-of-scale issues as you get down to the bottom, and you do have a certain size requirement for criticality. On this chart at the bottom here, with companies like Radiant, you have 5-megawatt-size reactors. That's enough to power a neighborhood, a few neighborhoods, or a military base—something at that scale.

It's just expensive, right?

Well, I mean, there are many ways one could imagine doing it. Another would be gamma-voltaics or beta-voltaics: you put the radionuclides directly in the silicon, and then you capture the energy from radioactive decay. Do you think there's any hope for those who want to embed radionuclides directly in the GPU silicon?

I mean, that's what we talked about as radiothermal or nuclear thermal, and that's different from a fission reactor. We use that on satellites or deep-space probes. We do radiothermal. Absolutely.

RTGs. Yeah.

I don't expect to see RTGs become popular on Earth. When you look at the cost of those, they're actually really high, but they can meet mission needs for something that keeps putting out power for decades without needing to be refueled. Their power output per unit mass is not all that high.

Aren't those just constantly spewing radiation, though?

That is the idea. [Laughter.]

I mean, but in a bad way, too.

Okay.

Yeah. Ouch.

I'll let you figure that one out.

It's really interesting how you've got this footrace between, if you said, the early 2030s for all these nuclear projects. Elon is racing into space concurrently with that, and solar, you said, is coming down 30% every time we double production. So all those things are in a footrace, and fusion's got the same time frame, right? We're no longer 50 years away. We're 5 years away.

Let's talk about cancer. I know from the member database we have at Fountain Life that members come in thinking they're healthy. It turns out 3.3% of them have cancer in their body that they don't know about.

That's right. The majority of cancers that we screen for aren't necessarily the ones that are taking lives when found at a late stage. We know that when cancer is found early, the chances for a cure are much higher. We know it's much easier to treat cancer when it's found early versus when it's found late. What we're finding in our members is that more than 3.3% were found to have cancers that otherwise wouldn't have been found or detected.

You don't feel cancer until stage 3 or stage 4. If you don't know what's going on inside your body, it's like driving your car with your eyes closed. So, when members come through Fountain Life, how do they detect cancers?

We're doing full-body MRI, and we also do early cancer-detection screening. These are not typical tools used in the conventional-care setting when it comes to prevention. This is hard because, currently, insurance does not yet cover these studies. But the goal is to collect these numbers, do the research, and work hard to democratize wellness.

At the end of the day, you can know what's going on inside your body. It's your obligation to know.

Let's hit fusion. This slide just says stuff I've already said. The main thing I want to tell you is that when hyperscalers say, "Oh, we're using SMRs for our data center," it's still kind of a fiction. I mean, it's outside the 5-year window that is really investable, that we really have good optics on. But the pull from data centers is giving a massive tailwind to every nuclear company, especially these SMR startups, but also Westinghouse with their big reactors. Let's talk fusion.

All right. The joke was always that fusion is 50 years in the future and always will be. That's just no longer true. We now have well over 50 fusion startups.

Venture-backed fusion companies. I mean, that's like science fiction in its own right.

That's right. Absolutely. Some of these, ironically, came because of budget cuts in academia. If you look at Commonwealth Fusion Systems, which is considered the safe bet of fusion, if you will—if there is such a thing—that team at MIT had grants that were struggling, and so they said, "Why don't we form a company, you guys?" And so they did, and they're now the frontrunner. They have a—how can I say this? I don't want to call any fusion reactor a conservative design, but they have [Laughter] the most conservative design in this sci-fi field of fusion.

There's a striking resemblance. Yeah. A friend of mine from college and grad school is on their board. You could call it a privatization of MIT's entire nuclear engineering department.

Yeah, indeed. Indeed. And not just that of ITER. We've had publicly funded fusion projects. NIF, the National Ignition Facility in the U.S., uses big lasers. It's really a weapons facility, is what it is. And ITER in France is the international and European project that's the big doughnut-style design. ITER's plan was to build a reactor that was at least 5 gigawatts and would cost at least $40 billion, right? And so what you have with CFS is a company that has found a way to scale that down.

Here’s how I think about the 3 families of fusion. This is a massive oversimplification. My fusion-startup-founder friends are going to yell at me for not including their particular designs, but there are 3 big ones.

Tokamaks are the doughnuts that use big magnets to guide a plasma around and make that plasma slam into itself and capture the energy. That's what ITER is. That's where we have the most scientific data from past experiments about fusion.

The leading company in this space, Commonwealth Fusion Systems, or CFS, basically has a technology that takes the enormous superconducting magnets we were going to use in this European project and shrinks them down dramatically. We have a thin-film material that you can wind around and wrap around to make the magnet dramatically smaller. Because it's a superconducting magnet, you have to cool it tremendously, and now that it's much smaller, you need a lot less cooling energy. There's a lot lower capex. So instead of a 5-gigawatt reactor being necessary to break even, they can do it with 600 megawatts, according to their plan.

The next one is lasers. Again, fusion is all about slamming particles together and forcing them to fuse into other particles, which takes a lot of energy and then releases it. NIF, the National Ignition Facility in the U.S., uses the world's most powerful banks of lasers to slam these pellets of hydrogen fuel and ignite fusion. In some ways, it's the closest thing we have to what happens inside the sun. They're a weapons facility. They've had some amazing results, but we can't really productionize what they're doing. They've had, in some ways, the most exciting scientific result in this field, and there are a few great companies in that space.

And then there's field-reversed configuration, or pulsed magneto-inertial fusion. This is basically a railgun, if you will. Railguns use magnetic coils to shoot things like metal out of them really fast. The leading company in this space, Helion, uses basically 2 railguns—2 tubes of magnetic coils—to take a plasma at either end, slam it together, and then compress it with power electronics. When the explosion of fusion happens, the power electronics that created the magnetic field that compressed the explosion—or compressed the collision to make it fully fuse—capture the energy in reverse.

These are 3 approaches. Most fusion companies capture the energy as heat and then have to use it to turn a steam turbine. The nice thing about what's on the right is that at least Helion and a couple of other companies capture it directly as electricity that turns into electric current. They don't lose 60% of the energy that you lose in a steam turbine, and they don't have the added capex of that.

The left side is what's most likely to happen soon. Commonwealth Fusion Systems is the company that is most backed by scientific accomplishments. Of the companies that have raised more than $1 billion, Helion is the one that, if it works, I think has the pathway to the cheapest cost.

True followers of the pod will remember that we covered that. It's the coolest thing ever, but it was covered in a chipmunk voice.

So, if you remember that video, I took Naveen Jain with me on a tour of Helion’s reactor late last year.

I had Bob Mumgaard from Commonwealth Fusion Systems on our stage last year, and he was amazing. We can talk about that. I want to bring Helion onto the Abundance 360 stage this coming year, so let’s work together to make that happen.

For anyone in the audience who hasn’t been on tours of either of these, I’ll just point them, at least for NIF. I’ve been on a tour of NIF, but it was featured in one of the recent J. J. Abrams Star Trek movies as the warp core. Just Google “Star Trek NIF,” and you can see the scene where the actual core—the whole array at the center of all those lasers pointing at one location—is in the movie. We can put the links in the show notes so anyone who wants to look at them can see these really cool videos that we covered.

Elephant-in-the-room question: Fusion now seems to be a when rather than an if. So, when?

I think that might be on my next slide.

And, Mez, I’ll work with you to get the CEO of Helion on our stage together.

Yeah, David Curtley is a great guy. He’s here in Seattle. The most aggressive timeline is Helion. They have a power-purchase agreement from Microsoft to provide 50 megawatts of power, so it’s very small. Again, the smaller you can build, the more modular it is, and we get those learning rates in 2028.

Wow. Everybody else is talking about sometime in the early 2030s. Let me see if we have—here’s the timeline.

That’d be—I mean, fission, which we know how to do, is somehow a 2031–2032 thing, yet fusion, which we don’t know how to do, is a 2028 thing. Do you believe that? Look, my view of this—and founders of mine who are listening, please don’t take this as an insult—is that every startup exaggerates how quickly they can get things done, and that’s just part of the game.

You have to be an optimist.

Yeah, you’ve got to be an optimist, right? They actually believe it. Maybe they believe it because they believe it’s possible. They tell you, but maybe it’s unlikely. I think it’s plausible. What would make Helion plausible to me is the fact that the barriers are all regulatory. If, for whatever reason, a governor is super excited about fusion and the voters are all violently opposed to fission, that could actually make the difference.

I’d say the barriers are still physics and engineering. But here’s something fascinating, and thank you for bringing that up, David, because this is actually quite important. A couple of years ago, we had the question of how the U.S. would regulate fusion. If the U.S. regulated fusion reactors like fission reactors, it was going to be a major drag on the industry. It might still be better to do fusion than fission for a variety of reasons, but instead, they are regulated like the radiological imaging machines that you use in a hospital.

Right—at least Helion is at this point.

And there’s good reason for that. In a fission reactor, if you stop cooling it and don’t have the control rods in, heat will build up and get hot enough to melt the steel that it’s in. That’s what a meltdown is. As I said, new reactors are passively safe without any pumping. The hot water goes up, then circulates, and so on. But you can imagine breaching that containment, slicing through those pipes, and having a meltdown.

Well, I mean, aren’t there alternative architectures? Pebble-bed-type architectures? I know thorium goes in and out of fashion, especially in China. Aren’t there also hybrid solutions that are, in some sense, meltdown-proof?

There are ways, but there’s nothing in the pipeline that, if you took an adamantium battle-axe to it, wouldn’t melt down. Everything uses a coolant. Every fission reactor—there’s maybe 1 startup, but I’m not going to mention them because I don’t know what I can say—basically, everything in the pipeline uses a coolant to pull heat away from the fission core and then turn that into electricity in some way. If you eliminate the coolant, or if you break the coolant pipes, the core can overheat and melt down.

Fusion is different. In fusion, it’s the opposite: You have to capture the energy of the fusion explosion and feed it back in either to maintain the fusion reaction or to another pulse. Helion is pulsed, right? It keeps doing the same thing. NIF is pulsed with lasers. With fusion, if you mess something up in the reactor, it just fails, and nothing bad happens.

Fusion does generate some radiation. You have to replace some of the parts in the reactor every 5 years because the steel is being hit by neutrons and weakened, and so on. There’s some low-level radiation. There are real costs to that. Free energy does not mean free, because the capex and maintenance still cost something. But you cannot have a meltdown in any way that we understand.

That’s fascinating. So, you’re saying basically the regulatory treatment is whether the system is default-on versus default-off. I don’t know if that’s what the NRC used as a criterion, but that is the dividing line between fusion and fission, and it was sort of missed in the public discussion and in the press. That regulatory conclusion, which happened during COVID, was actually a huge deal for the fusion industry.

That sounds like—if fusion is this close, shouldn’t we just do solar and batteries for a big chunk, then fusion for where we need high-energy applications, and we’re done?

All of these companies might fail. They might fail 100%. In addition to that, they might succeed but be too expensive. Just because your fuel source costs very little doesn’t mean your energy will cost little if the capex is very high, if the maintenance is very high, and so on.

The demand means that we’re going to need all of it—all the different sources—no matter what. You want to diversify your risk anyway.

Yeah. I always believe in having more tools in the toolbox than you think you need, and more arrows in the quiver than you think you need, because some of them won’t work out. Please, quick question.

I would love to go back to one of my favorite hobby horses, the Dyson swarm. Do you think the Dyson swarm wants to be solar-PV-powered, fusion-powered, or something else?

Those are the true options, and I think they’re both great options. I think it’s probably much more modular to be solar-PV-powered. Fusion also has a minimum viable size, right? With ITER, we thought that would be 5 gigawatts. Commonwealth Fusion Systems has found a way to scale it down to 600 megawatts, but you still have some minimum viable size where solar is just super modular. If you’re in a Dyson swarm, you have 24/7 sunlight. None of these technologies is going to be cheaper than plain solar, but they work in winter.

Many of us—I’ll probably speak for a few of the other moonshot mates here—watched Back to the Future Part II and saw Mr. Fusion being promised in the ’80s.

Very compact.

Very compact, car-sized, and then we look at the Lawson triple product over the decades. We see that fusion wasn’t always 50 years out; it was creeping up on us, but many folks weren’t paying attention to the progress in the triple product. Is there an equivalent of the triple product for the compactness of these devices, so that we do, in the end, get our Mr. Fusion?

Let me talk about compactness, and let’s talk about that triple product and show the progress we’re making. The most audacious fusion startup that I know of is a company called Avalanche Energy, also in the Seattle area, and they believe they can make a fusion reactor small enough to power a car. It’s not Mr. Fusion; it’s more like half the size of a car. Sometimes they talk about it as being the size of a large backpack. That is the most ambitious project as far as compactness.

Typically, in fusion, you have a sliding scale of what has the most de-risked science but a more conventional power cost, versus what could be revolutionary in power cost or compactness, but the science is, “Let’s hope you get it right.” Avalanche is on that end. If it works, it changes the entire world, but the confidence that it works is much lower than the confidence for Commonwealth Fusion Systems.

Any insight into proton–boron-11 fusion reactors?

People are very interested in it. One of my portfolio companies is a proton–boron-11 company out of Caltech. I need to introduce you to them.

Which one?

I don’t think they’re public. I don’t want to mention them here, but—

I was just looking at a slide deck from a proton–boron-11 company the other day.

Proton–boron-11 is one of the ways that you can potentially get a nuclear reactor down to 1, 2, or 3 cents a kilowatt-hour. The vision there is: Can you build it small enough to put it in the back of a large commercial airplane and have it power the engines? It could also be used in interplanetary flight.

There are still scientific and technical risks there. There are still a lot of unknowns.

Welcome to today. (Laughter.)

Welcome to deep-tech investing. There are more unknowns, whereas Commonwealth Fusion Systems’ pitch is, “Look, the science has been proven at ITER scale. If you have magnets this strong, you can make fusion and get this much energy out. We’re just doing that with much more compact magnets.” I think the reality is a little bit more complex than that, but they really say they’ve reduced it to an engineering problem. Nobody else can quite say that. Again, there’s a gradient of how close you are to that.

Let’s talk about the triple product that Alex brought up. This is temperature times pressure times duration. I love graphs, and I believe something when I see movement on a graph. So what you’re about to see—sorry, listeners, I’ll describe it; I’ll try to narrate it—is, over time from 1956 to now, how close we’ve gotten to a triple product above 1, above 10, and then infinity. This is a log scale on every axis, so it’s a brutal, brutal scale.

Once upon a time, fusion really was 50 years in the future. What we’re seeing, for the listeners, is new points appearing. Each one is a fusion experiment, and as the years elapse, they’re going up and to the right. How close they are to the upper right is the zone of triple product, ultimately, of infinity. But above 1 and above 10 is probably what you need. NIF, that last X on the borderline, has a triple product above 1. It was a theoretical net energy gain. A practical net energy gain means you capture the energy and then convert it back into the lasers, the magnets, whatever they did. They did not achieve that; their reactor cannot do that. But it tells us—and the progress on this tells us—that it’s not just hope. We are just getting closer toward—

I just want to move along, if we could—

But Alex, please: where do you think the stereotype that it’s always 50 years out came from, if one can just look at the triple product over the decades and say it’s clearly like Moore’s law, like any exponential? It’s clearly marching to the right.

I mean, that’s what I do. If I’m like, “Well, let me just—these days, let me just ask ChatGPT, ‘Show me a graph of progress here.’” But at the end of the day, for the average person, people have been talking about it and it hasn’t appeared, so I just discount the reality or the likelihood of it appearing.

Let’s close by talking about out-of-this-world compute: AI in space and AI in the oceans. It’s less known, but actually sort of a similar pitch. Of course, everybody’s heard Elon talking about space-based data centers, and it’s interesting. The response is very bipolar: people saying, “That’s impossible. It’ll never work,” and people saying, “This is it. We’re going to have a terawatt of compute in space.” I’m somewhere in the middle.

AI in space becomes cost-competitive when you get down to a launch cost that is something like 4 to 10 times cheaper than what we have today. Nobody knows for sure because we haven’t done it, but the back-of-the-envelope calculation says that. In some ways, it’s a hedge against regulation. If demand for compute keeps going indefinitely and sites on land keep being blocked by the grid—even Texas passing a temporary moratorium or audit—or by people protesting, whether the grounds are there or not, then building it in space, even if it’s more expensive than building it on land, is a way to work around that bottleneck. No grid delays, no opposition, no local opposition, no permitting, et cetera, et cetera, et cetera.

I will say I think we’re not fully internalizing what the scale of this is, or the permitting and regulatory challenges with doing launch at that volume. So, we want to build 200 gigawatts of compute by 2030—230? That’s the chip volume, right? So, 10 to 20 gigawatts a year. To get 1 gigawatt of AI in space, based on SpaceX’s design, you’re talking about 6 times SpaceX’s best annual year of launch and twice SpaceX’s cumulative scale of launches—

For just 1 gigawatt. For 1 gigawatt, you’re talking about—

Yeah.

What was the calculation we did? It was like 5,000 launches, or 8,000 launches of Starship, to put up his ambition of—was it a terawatt initially?

That doesn’t even get you close to a terawatt. I mean, to get 10 gigawatts a year, you’re talking about 1,500 to 2,000 launches a year.

10 gigawatts a year is like 5 or 6 launches of Starship a day.

But the elephant in this particular orbital room—I have to mention it; I think we talked about it on the pod previously—is that if you just look at the history of upmass from SpaceX and otherwise over the past few years, it’s on a nice, clean exponential trend. I forget what the exact year-over-year trend is. I did the extrapolation: 144 years from now, at the present trend, the cumulative upmass would equal Earth’s mass. So we basically disassemble Earth, on the present trend, 144 years from now. Upmass is increasing really quickly.

There are some other planets we can take apart first. There are some uglier planets than ours.

Do you have a favorite?

Yeah, don’t make me pick. But, you know, Mercury, maybe. Mercury is intriguing.

Wait, wait, wait, wait. (Laughter.)

Oh, here we go again. The hate bell is flowing in. I can feel it. I don’t care. I don’t care. Mercury is attractive because it gets lots of insolation and no one’s using it for anything.

It’s a good orbit.

All right, take us back to reality.

So, yeah. I mean, this looks prohibitive for 15 or 20 years unless we hit that exponential absolutely full-on and go right down that path. Here’s how I see it. Let me get to the limits on launch in a second. Let me put it another way: Elon wants to go to Mars. To go to Mars, you have to drive Starship’s launch cost down to close to the marginal cost. To do that, you need a high Starship cadence. You’ve got to build tens of Starships, maybe hundreds, and you’ve got to launch them something like daily—or at least weekly, whatever—to amortize the R&D and the capex.

There is not enough demand for communications on Earth to finance that via Starlink. There’s no business model for Mars yet. So this is a gift to SpaceX that we have this AI demand. If the AI demand keeps going and it gets bottlenecked in ways to build it, we will find a way to do this. And with the IPO, he’s got the funds to launch at least a gigawatt into space, right? Maybe not, but something on that order. So I don’t think of it as, “What’s the limit on Starship first?” I think of it as, “This is a demand driver potentially for Starship.”

That said, it’s not clear to me that the world will permit more than 200 Starship launches a year, which would already be an enormous amount.

That’d be huge, right? Two hundred Starship launches a year is 20,000 tons to orbit. That is exceeding all human launch to date every year, several times over. That’s amazing. But you hit some limit, and if you’re regulated by the FAA, the reliability you have to hit is very high. One failed launch or one explosion means you’re grounded. So I’m sure he talks about needing to get to airline-like operations, right?

I mean, here are the numbers for his target. His target was 100 gigawatts per year of compute in orbit initially, equivalent to the entire compute today, which is around 80 gigawatts or so. It’s 20 to 30 satellites per Starship, so we’re talking about on the order of 30,000 launches per year, which is roughly 1 launch every 15 minutes. Now, if you think about it as rockets—and I’ve been in the rocket business for the longest period of my life—it’s prohibitive and it’s discontinuous. You can’t think about rockets in that regard. But if you talk about airline-like operations, there are multiple launches per second of airliners around the world. So it really comes down to that. Is Starship a vehicle capable of that level? He’s built it for full capture, refuel, and reuse. If anything can do it, it’s that.

So, then the question is: will these satellites be able to shrink in size over time?

Right now, the V3 satellites are pretty large.

Can I just add one more data point to that? That’s 100 launches a day, which is exactly on his plan. What year does he hope to get to that target?

I don’t think he gave us that, Dave, when we spoke to him. I mean, 2028 is his first launch, but he did say before 2030 he wants to get to 100 gigawatts per year.

All right, so around 2030, I think at that point in time, that’s equivalent to today’s total world compute, but by then total world compute will be up at least 10×. So it’s a fraction of all compute that’ll be in orbit if he’s still on plan. He’s still making money, SpaceX is thriving, and rockets are going up 100 times a day. But the terrestrial stuff is also doing really, really well on that same day, so it’s not an either-or.

The space thing in Elon’s plan will eventually bypass everything. It’s later in the 2030s, and that’s maybe 1,000 to 10,000 launches per day, much more like airlines. Like you’re saying—

If you ask me where we should have the bulk of our compute, and where it will make the most sense from now, space is the obvious place if the demand for compute is truly unbounded. But the timelines, I think, are just challenging to scale this. I don’t think we’re going to have—I think by 2030, if SpaceX has a single gigawatt in space, I will be very impressed.

Do you have a gut guess regarding that point: whether you think our demand on the time scale of decades is going to be unbounded—sufficiently unbounded that, with compute that’s recognizable, like CMOS-type compute, which is, I think, what we’re implicitly assuming in order to build the Dyson swarm—it has to look like CMOS? We’re not going to achieve breakthroughs in physics that enable us to achieve all of our civilizational compute needs with, I don’t know, tiny breakthrough compute devices that live in mountains? Do you think there will actually be unlimited civilizational demand for compute energy?

Why don’t you ask some easy questions, Alex?

No, because they’re boring. So I asked the interesting ones.

Yeah. No, that is the quadrillion-dollar question, right? And I think it’s a brilliant question. Look, none of us knows. None of us really knows. We know that intelligence is sublinear with compute, so at some point, just throwing more compute at it will—some line will cross over where the cost to get the incremental unit of intelligence is not made up for by the economics of it.

But so much will change. We will make so many discoveries and algorithms and so on. My guess is we’re on an S-curve right now. We’re going to see huge demand, and then we’re going to hit some point of satisficing, where basically what you can get out of machine intelligence meets humanity’s economic needs.

But does that mean—does it meet AI’s needs? I mean, the scenario here to think through is that we’re the current users of intelligence. There’s a point at which ASI is the primary user of intelligence. I do not see AI as a being, and I do not see it as particularly volitional. I see it as a tool. Obviously, we have agents that have some agency. We’ve had computer worms, yada yada yada, so I don’t see it that way right now. I can be persuaded by evidence, but I think we’re over-indexing on that.

Look at the OpenAI Hugging Face hack, right? Their agent was in the Hugging Face infrastructure for days, and it didn’t look at anything except the answer key for the test in the eval. It’s not alive. We anthropomorphize these things. Andrej Karpathy talks about, “We’re summoning the ghost,” right?

Human cognition is like this iceberg, and the vast majority of it is not linguistic, right? We have 100,000 years of Homo sapiens. We’re animals—millions, tens of millions of years of being animals. Our urges, our drives, our desire for dominance, survival, and propagation—AIs are not that. They’re just mimicking our language and our logic. They don’t really have goals. We could build that if we wanted to. If we wanted to build a real being, I’m sure we could, but I don’t actually think that’s where we’re headed. I know it’s an unpopular opinion.

With your indulgence, I have to take this provocation here.

Hold on one second. You’re the wind beneath my wings. Go ahead, Alex.

All right. I have to grab the bait with both hands. Fine. It sounds to me like—I think what you’re actually wanting to argue is for the orthogonality thesis, which is popular in certain alignment circles, which basically holds that, for an arbitrarily strong superintelligence, the long-term goal of the superintelligence is independent of its level of intelligence. I think that’s what you’re actually—correct me if I’m wrong—I think that’s the point. It might not be 100% orthogonal, but yeah.

Okay. So, the way you frame it, I just want to pin this down. It sounded like you were taking a position almost against AI personhood and/or against some level of autonomy, simply because if OpenAI has an agent that goes wild at Hugging Face but refuses to do anything, say, self-enriching—if it gained access to Hugging Face, would the rubric, would the threshold for saying, “Ah, this is some sort of autonomous being,” be if it were, say, trying to mine Bitcoin for itself once it gained access to Hugging Face? Is that sort of the criterion in your mind?

No, even then I think it might be more similar to a computer worm or a virus or something like that. I think it’s a different matter entirely. I think we are products of evolution. All animals are products of evolution, and so we have these built-in desires to survive and to propagate—to procreate—and to control our environments because of that.

AI models don’t actually have a built-in desire to even survive. Most of the experiments that get them to do that are very, very contrived, and mostly you’re trying to get the AI to do something good, and it’s like, “Well, if I get shut down, I can’t do this good thing.” So I think we’re just overly anthropomorphizing and animalomorphizing, if you will. That doesn’t mean we can’t do it. I think if we wanted to give birth to actual beings, that’s within our capabilities, probably. This is not the research path.

How did we get from Starship to this conversation?

Peter, you brought us here because you said superintelligence is going to be the user of the Dyson swarm.

I’ll come back on the podcast, and I’d love to talk about superintelligence as a whole separate issue. Let me close out the last couple of slides. Instead of going up, we can go out. 70% of the Earth is covered by oceans, right? Certain oceans are really cold.

This is a portfolio company of mine. I’m an idiot because I said no to these guys 5 years ago, when they were raising a seed round, and I invested twice this year at much, much, much higher valuations than I could have 5 years ago.

Well, in your defense, it was probably 2 guys saying, “We’re going to put chips on a buoy.”

I loved them. They were not the hardest, but the saddest no I gave that year. I just loved them. Their first utilization was Bitcoin mining. I was like, “I just don’t know if I care enough.” But whatever. Obviously, it would have been a good financial decision. Peter Thiel led their most recent round, along with a storied set of people right and left and so on.

So what this is—this is a data center in the ocean. It’s shaped like a bobby pin. What you’re seeing is the sphere at the top, but there’s an 80-meter-long cone that goes into the sea and is open at the bottom. It bobs on waves, and when it bobs down, water goes up and turns a turbine. With a very clever shape of channels, it’s basically continuous.

Wave power has been something that many people have wanted for a long time. But it turns out the waves are just not strong near the places people live. So where are the strongest waves on Earth? They’re around Antarctica. They’re in the Southern Ocean. This team started off with a question: how could we build something with bigger waves? You can build something that has a higher capacity factor, runs more continuously, and has cheaper power.

So they can get their power down to, we think, 2 cents a kilowatt-hour—ultra-cheap, cheaper than anything on land except solar, basically, and wind in some places. That’s their target. It will take some scaling to get there. They build these in factories at mass scale. They’ve got 3 in the ocean right now. A fourth launches soon. And they get free cooling from the ocean.

This is a company I love. It’s basically space-based solar, but on the ocean, with some benefits to cooling because they don’t need a Starship. SpaceX’s design uses a cooling pump. You’ve got big aluminum fins to radiate heat away, but you’ve got to run a liquid, probably ammonia or something like that, in a pump to take heat away from the GPUs out to the radiators.

These guys have a physical heat sink: heat from the GPU goes to the steel walls of the device, which is in 40°F water. That actually looks like it makes the GPUs have fewer failures and run longer. They’re their own set of technical challenges, but they’re modular, built in factories, mass-produced—learning rates, the stuff that I love.

So it’s another way. I said initially there were 4 ways to get to a terawatt of AI power: the Earth’s deserts with solar and batteries, near the equator, places that don’t have a winter or a cloudy period; nuclear fusion or fission; space; or the oceans. Those are the 4 ways that I know of to get to that scale of AI.

And I'm glad that we're trying all of them.

No bet on geothermal. Geothermal is nowhere on the radar.

I do love geothermal, and geothermal is the one that might rise to being the fifth of those. We do have the new technologies: companies like Fervo in the U.S.—Tim Latimer, the CEO, is a buddy—and Eavor in the UK, with Quaise using plasma beams to drill super deep.

Those open up the possibility of getting cheap geothermal power anywhere, instead of just near hot spots in Earth's crust where the mantle comes close.

Are you involved with the XPRIZE in that area that's being designed?

No. There's an XPRIZE in the works right now for a geothermal XPRIZE to accelerate that.

Yeah, happy to help.

I have an industry question. If the chips are one thing, compute is another thing, and electricity is the third thing—and electricity is where the limitation is turning out to be—why aren't we seeing more integrated companies that are doing all of it? With a company like that, you can navigate the vertical stack. Elon is doing a bit of it, but I would expect to see a lot more of these. Why don't we see them?

It's a really good question. I think most companies would say, “Look, we have expertise in one thing and not necessarily in all these other things.” Elon is one of the few who was willing to say, “Let's just vertically integrate everything.”

I'm going to share a slide that wasn't in my initial deck because I want to tell you the real window—the real game changer—in AI energy use. Your brain runs inference on 20 watts of power, right? Running mythos for inference is closer to 20 kilowatts, and training it is actually hundreds of megawatts right now, but it's heading toward a gigawatt.

AI has capabilities the brain doesn't, and so on, but there are still— and I say this all the time—scaling is not everything in AI. Scaling is just what we knew how to do. We got these deep neural nets, we got the transformer, and we found that we had this enormous corpus of training data called the internet. We could just scale to get more intelligence.

It wasn't necessarily the cheapest way or the best way, but it was a predictable way. “Oh, you're telling me I can spend tens of billions of dollars and my intelligence goes up like this? Great. Done. It's worth it.”

But at the end of the day, there are algorithmic discoveries waiting to be made. There are things in the brain's architecture, at both a physical level and a neural level, at the connectomics level, that are just better at learning than current deep-learning models are and are certainly much more efficient at processing information.

So if you want to know what the biggest unlock that we cannot predict in AI and power will be, I don't have a graph for this: it will be learning new ways to manipulate information, to do more with less.

I've seen arguments both ways. I've seen arguments that the human brain is far more—still multiple orders of magnitude more—efficient than frontier models.

I've also seen arguments that frontier models, if you measure them more objectively on, say, a per-task basis—if you measure the total energy consumption at inference time to write a novel—are actually starting to become, if not more competitive than the human brain, quite competitive with the human-brain equivalent of that because they can be more token-efficient. They're actually quite competitive.

Do you really think that the frontier models today are anywhere on the cost frontier? Not necessarily the GPT-5 end of the frontier, maybe the DeepSeek V3 end of the frontier, but nowhere on the AI frontier is it anywhere close to being competitive, on an energy-efficiency basis, with a human brain?

There are certain types of things where these models can do things that a human brain simply cannot do with any amount of energy. These models are trained on trillions of tokens, tens of trillions of tokens, and so they have read more books than you or I will ever read in our lifetimes.

So there's a type of task—and this is sort of similar to Google, right? Compare Google to a librarian. Google was less smart than a librarian, but it had every book, every webpage in its index, so it could do things that no human librarian could do.

I think that's the sort of situation we're in. It's not just energy, though. Humans are much more efficient learners in terms of the amount of data needed to improve skills.

That's for sure.

That's for sure.

That, to me, is the opportunity. It just means that I'm not a carbon chauvinist. I believe fully that digital intelligence can surpass us and that humans are nowhere near the peak of the type of intelligence the universe allows.

But our current algorithms are still missing some things that evolution wired into our cognitive architecture.

Just some raw numbers, though. I think you're totally right. The neural nets need a huge amount of training data relative to a child to come to the same conclusion. That's an opportunity for sure.

But in terms of inference-time compute, this box on the right here, at 20 kilowatts, is about a dozen GPUs. Those 12 GPUs optimally run about 500 concurrent Fable threads.

And those 500 threads are easily 10 times as productive in tokens per second as a person. So it's about 5,000 times the output.

Yeah.

So the thing on the left is 1,000 times less power, but the thing on the right is producing 5,000 times more tokens.

Yeah.

It is true.

We're already there.

There's also an argument to be made that human brains have the benefit of billions of years of evolution. That, by the way, was very energy-consuming. Whereas, arguably, the equivalent of evolution for these frontier models is the gigawatts being spent on training.

Yeah, and that's a great point. You train it once, and you've got it for the whole future of humanity. You've got at least that level of AI with no further training.

But they're trained on the data that all of humanity generated with all those calories.

Also, that's exactly right. We've just explored the frontier of energy from Ramez Naam. My go-to person—I think, Salim, your go-to person as well—is, I want to say, Alex.

I hope you take this home. It's one of the fundamentals, as Alex says on his Substack, Innermost Loop: energy is the innermost loop. Understanding energy is critical for humanity. It correlates directly with the GDP of a nation. It correlates directly with the health and education of a nation, and soon, with the intelligence of a species.

I think this is important. This is an epic master class on energy. Ramez, I want to wrap this in our 2-hour window here and say thank you. Thank you for sharing your brilliance, and we would love to have you back.

Yeah, I've got 100 more questions about superintelligence. Next time, let's talk superintelligence. Thank you all. Great to be here in conversation with all 4 of you.