Are there really three hours of  questions? Are you fucking serious? 

You don't think there's a lot to talk about, Elon? Holy fuck man. 

It's the most interesting point. All  the storylines are converging right now. 

We'll see how much we can get through. It's almost like I planned it. 

Exactly. We'll get to that. 

But I would never do such a thing… As you know better than anybody else,  

only 10-15% of the total cost of  ownership of a data center is energy. 

That's the part you're presumably saving  by moving this into space. Most of it's  

the GPUs. If they're in space, it's harder  to service them or you can't service them. 

So the depreciation cycle goes down on them. It's just way more expensive to have  

the GPUs in space, presumably. What's the reason to put them in space? 

The availability of energy is the issue. If you look at electrical output  

outside of China, everywhere outside  of China, it's more or less flat. 

It’s maybe a slight increase,  but pretty close flat. 

China has a rapid increase in electrical output. But if you're putting data centers anywhere  

except China, where are you going to get your  electricity? Especially as you scale. The output  

of chips is growing pretty much exponentially,  but the output of electricity is flat. 

So how are you going to turn the chips on? Magical  power sources? Magical electricity fairies? 

You're famously a big fan of solar. One terawatt of solar power,  

with a 25% capacity factor, that’s  like four terawatts of solar panels. 

It's 1% of the land area of the United States. We’re in the singularity when we’ve got  

one terawatt of data centers, right? So what are you running out of exactly? 

How far into the singularity are you though? You tell me. 

Exactly. So I think we'll find we're  in the singularity and it’ll be like,  

"Okay, we’ve still got a long way to go." But is the plan to put it in space after  

we've covered Nevada in solar panels? I think it's pretty hard to cover Nevada  

in solar panels. You have to get permits. Try  getting the permits for that. See what happens. 

So space is really a regulatory play. It's harder to build on land than it is in space. 

It's harder to scale on the ground  than it is to scale in space. 

You're also going to get about five times the  effectiveness of solar panels in space versus  

the ground, and you don't need batteries. I almost wore my other shirt, which says,  

"it's always sunny in space". Which it is because you don't  

have a day-night cycle, seasonality,  clouds, or an atmosphere in space. 

The atmosphere alone results  in about a 30% loss of energy. 

So any given solar panel can do about five  times more power in space than on the ground. 

You also avoid the cost of having  batteries to carry you through the night. 

It's actually much cheaper to do in space. My prediction is that it will be by far  

the cheapest place to put AI. It will be space in 36 months or  

less. Maybe 30 months. 36 months? 

Less than 36 months. How do you service GPUs as they fail,  

which happens quite often in training? Actually, it depends on how recent the  

GPUs are that have arrived. At this point, we find our  

GPUs to be quite reliable. There's infant mortality,  

which you can obviously iron out on the ground. So you can just run them on the ground  

and confirm that you don't have  infant mortality with the GPUs. 

But once they start working and you're  past the initial debug cycle of Nvidia  

or whoever's making the chips—could be Tesla  AI6 chips or something like that, or it could  

be TPUs or Trainiums or whatever—they’re  quite reliable past a certain point. 

So I don't think the servicing thing is an issue. But you can mark my words. 

In 36 months, but probably closer to 30 months,  the most economically compelling place to put  

AI will be space. It will then get  

ridiculously better to be in space. The only place you can really scale is space. 

Once you start thinking in terms of what  percentage of the Sun's power you are harnessing,  

you realize you have to go to space. You can't scale very much on Earth. 

But by very much, to be clear,  you're talking terawatts? 

Yeah. All of the United States currently  uses only half a terawatt on average. 

So if you say a terawatt, that would be  twice as much electricity as the United  

States currently consumes. So that's quite  a lot. Can you imagine building that many  

data centers, that many power plants? Those who have lived in software land  

don't realize they're about to  have a hard lesson in hardware. 

It's actually very difficult  to build power plants. 

You don't just need power plants, you  need all of the electrical equipment. 

You need the electrical transformers  to run the AI transformers. 

Now, the utility industry is a very slow industry. They pretty much impedance match to the  

government, to the Public Utility Commissions. They impedance match literally and figuratively. 

They're very slow, because  their past has been very slow. 

So trying to get them to move fast is... Have you ever tried to do an interconnect  

agreement with a utility at  scale, with a lot of power? 

As a professional podcaster, I  can say that I have not, in fact. 

They need many more views  before that becomes an issue. 

They have to do a study for a year. A year later, they'll come back to you  

with their interconnect study. Can't you solve this with your  

own behind the meter power stuff? You can build power plants. That's  

what we did at xAI, for Colossus 2. So why talk about the grid? 

Why not just build GPUs and power co-located? That's what we did. 

But I'm saying why isn't  this a generalized solution? 

Where do you get the power plants from? When you're talking about all the issues  

working with utilities, you can just build  private power plants with the data centers. 

Right. But it begs the question of where do you  get the power plants from? The power plant makers. 

Oh, I see what you're saying. Is this the gas turbine backlog basically? 

Yes. You can drill down to a level further. It's the vanes and blades in the turbines  

that are the limiting factor because it’s a very  specialized process to cast the blades and vanes  

in the turbines, assuming you’re using gas power. It's very difficult to scale other forms of power. 

You can potentially scale solar, but  the tariffs currently for importing  

solar in the US are gigantic and the  domestic solar production is pitiful. 

Why not make solar? That seems  like a good Elon-shaped problem. 

We are going to make solar. Okay. 

Both SpaceX and Tesla are building towards  100 gigawatts a year of solar cell production. 

How low down the stack? From polysilicon  up to the wafer to the final panel? 

I think you've got to do the whole thing  from raw materials to finish the cell. 

Now, if it's going to space, it costs less  and it's easier to make solar cells that  

go to space because they don't need much glass. They don't need heavy framing because they don't  

have to survive weather events. There's no weather  in space. So it's actually a cheaper solar cell  

that goes to space than the one on the ground. Is there a path to getting them as cheap  

as you need in the next 36 months? Solar cells are already very cheap.  

They're farcically cheap. I think solar cells  in China are around $0.25-30/watt or something  

like that. It's absurdly cheap. Now put  it in space, and it's five times cheaper. 

In fact, it's not five times  cheaper, it's 10 times cheaper  

because you don't need any batteries. So the moment your cost of access to space becomes  

low, by far the cheapest and most scalable way  to generate tokens is space. It's not even close.  

It'll be an order of magnitude easier to scale. The point is you won't be able to scale on the  

ground. You just won't. People are going to  hit the wall big time on power generation.  

They already are. The number of miracles in  series that the xAI team had to accomplish in  

order to get a gigawatt of power online was crazy. We had to gang together a whole bunch of turbines. 

We then had permit issues in Tennessee and  had to go across the border to Mississippi,  

which is fortunately only a few miles away. But we still then had to run the high  

power lines a few miles and build  the power plant in Mississippi. 

It was very difficult to build that. People don't understand how much electricity  

you actually need at the generation  level in order to power a data center. 

Because the noobs will look  at the power consumption of,  

say a GB300, and multiply that by a thing and  then think that's the amount of power you need. 

All the cooling and everything. Wake up. That's a total noob, you’ve  

never done any hardware in your life before. Besides the GB300, you've got to power  

all of the networking hardware. There's a whole bunch of CPU and  

storage stuff that's happening. You've got to size for  

your peak cooling requirements. That means, can you cool even on the  

worst hour of the worst day of the year? It gets pretty frigging hot in Memphis. 

So you're going to have a 40% increase  on your power just for cooling. 

That’s assuming you don't want your data center to  turn off on hot days and you want to keep going. 

There's another multiplicative element on top of  that which is, are you assuming that you never  

have any hiccups in your power generation? Actually, sometimes we have to take the  

generators, some of the power,  offline in order to service it. 

Okay, now you add another 20-25% multiplier on  that, because you've got to assume that you've  

got to take power offline to service it. So our actual estimate: every 110,000  

GB300s—inclusive of networking, CPU,  storage, cooling, margin for servicing  

power—is roughly 300 megawatts. Sorry, say that again. 

What you probably need at the generation level  to service 330,000 GB 300s—including all of  

the associated support networking and everything  else, and the peak cooling, and to have some power  

margin reserve—is roughly a gigawatt. Can I ask a very naive question? 

You're describing the engineering  details of doing this stuff on Earth. 

But then there's analogous engineering  difficulties of doing it in space. 

How do you replace infinite bandwidth  with orbital lasers, et cetera, et cetera? 

How do you make it resistant to radiation? I don't know the details of the engineering,  

but fundamentally, what is the reason to think  those challenges which have never had to be  

addressed before will end up being easier  than just building more turbines on Earth? 

There are companies that build turbines on Earth. They can make more turbines, right? 

Again, try doing it and then you'll see. The turbines are sold out through 2030. 

Have you guys considered making your own? In order to bring enough power online, I think  

SpaceX and Tesla will probably have to make the  turbine blades, the vanes and blades, internally. 

But just the blades or the turbines? The limiting factor... you can get  

everything except the blades. They call them blades and vanes. 

You can get that 12 to 18 months  before the vanes and blades. 

The limiting factor is the vanes and blades. There are only three casting companies in  

the world that make these, and  they're massively backlogged. 

Is this Siemens, GE, those  guys, or is it a sub company? 

No, it's other companies. Sometimes they have  a little bit of casting capability in-house. 

But I'm just saying you can just call  any of the turbine makers and they will  

tell you. It's not top secret. It’s  probably on the internet right now. 

If it wasn't for the tariffs,  would Colossus be solar-powered? 

It would be much easier to  make it solar powered, yeah. 

The tariffs are nuts, several hundred percent. Don't you know some people? 

The president has... we don't agree on  everything and this administration is not  

the biggest fan of solar. We also need the land,  

the permits, and everything. So if you try to move very fast,  

I do think scaling solar on Earth is a good  way to go, but you do need some amount of  

time to find the land, get the permits, get  the solar, pair that with the batteries. 

Why would it not work to stand  up your own solar production? 

You're right that you eventually run out of  land, but there's a lot of land here in Texas. 

There's a lot of land in Nevada, including  private land. It's not all publicly-owned  

land. So you'd be able to at least get the  next Colossus and the next one after that. 

At a certain point, you hit a wall. But wouldn't that work for the moment? 

As I said, we are scaling solar production. There's a rate at which you can scale physical  

production of solar cells. We're going as fast as  

possible in scaling domestic production. You're making the solar cells at Tesla? 

Both Tesla and SpaceX have a mandate to  get to 100 gigawatts a year of solar. 

Speaking of the annual capacity, I'm curious,  in five years time let's say, what will the  

installed capacity be on Earth…? Five years is a long time. 

And in space? I deliberately pick five  years because it's after your "once  

we're up and running" threshold. So in five years time what's the  

on-Earth versus in-space installed AI capacity? If you say five years from now, I think probably  

AI in space will be launching every  year the sum total of all AI on Earth. 

Meaning, five years from now, my prediction is we  will launch and be operating every year more AI in  

space than the cumulative total on Earth. Which is... 

I would expect it to be at least, five years  from now, a few hundred gigawatts per year  

of AI in space and rising. I think you can get to around a  

terawatt a year of AI in space before you start  having fuel supply challenges for the rocket. 

Okay, but you think you can get hundreds  of gigawatts per year in five years time? 

Yes. So 100 gigawatts, depending  

on the specific power of the whole system with  solar arrays and radiators and everything, is  

on the order of 10,000 Starship launches. Yes. 

You want to do that in one year. So that's like one Starship launch  

every hour. That's happening in this city?  Walk me through a world where there's a  

Starship launch every single hour. I mean, that's actually a lower rate  

compared to airlines, aircraft. There's a lot of airports. 

A lot of airports. And you’ve got to launch into the polar orbit. 

No, it doesn't have to be polar. There's some value to sun-synchronous, but  

I think actually, if you just go high enough,  you start getting out of Earth's shadow. 

How many physical Starships are  needed to do 10,000 launches a year? 

I don't think we'll need more than... You could probably do it with as few as 20 or 30. 

It really depends on how quickly… The ship has  to go around the Earth and the ground track for  

the ship has to come back over the launch pad. So if you can use a ship every, say 30 hours,  

you could do it with 30 ships. But we'll make more ships than that. 

SpaceX is gearing up to do 10,000 launches a  year, and maybe even 20 or 30,000 launches a year. 

Is the idea to become basically  a hyperscaler, become an Oracle,  

and lend this capacity to other people? Presumably, SpaceX is the one launching all this. 

So, SpaceX is going to become a hyperscaler? Hyper-hyper. If some of my predictions come true,  

SpaceX will launch more AI than the cumulative  amount on Earth of everything else combined. 

Is this mostly inference or? Most AI will be inference. Already, inference  

for the purpose of training is most training. There's a narrative that the change in  

discussion around a SpaceX IPO is because  previously SpaceX was very capital efficient. 

It wasn't that expensive to develop. Even though it sounds expensive, it's  

actually very capital efficient in how it runs. Whereas now you're going to need more capital than  

just can be raised in the private markets. The private markets can accommodate raises  

of—as we've seen from the AI labs—tens of  billions of dollars, but not beyond that. 

Is it that you'll just need more than  tens of billions of dollars per year? 

That's why you'd take it public? I have to be careful about saying  

things about companies that might go public. That’s never been a problem for you, Elon. 

There's a price to pay for these things. Make some general statements for us about  

the depth of the capital markets  between public and private markets. 

There's a lot more capital available... Very general. 

There's obviously a lot more capital  available in the public markets than private. 

It might be 100x more capital,  but it's way more than 10x. 

Isn't it also the case that with things that tend  to be very capital intensive—if you look at, say,  

real estate as a huge industry, that raises  a lot of money each year at an industry  

level—they tend to be debt financed because  by the time you're deploying that much money,  

you actually have a pretty— You have a clear revenue stream. 

Exactly, and a near-term return. You see  this even with the data center build-outs,  

which are famously being financed by the private  credit industry. Why not just debt finance? 

Speed is important. I'm generally  going to do the thing that... 

I just repeatedly tackle the limiting factor. Whatever the limiting factor is on speed,  

I'm going to tackle that. If capital is the limiting factor,  

then I'll solve for capital. If it's not the limiting factor,  

I'll solve for something else. Based on your statements about Tesla  

and being public, I wouldn't have guessed that  you thought the way to move fast is to be public. 

Normally, I would say that's true. Like I said, I'd like to talk  

about it in some more detail, but the problem  is if you talk about public companies before  

they become public, you get into trouble,  and then you have to delay your offering. 

And as you said, you’re solving for speed. Yes, exactly. You can't hype companies  

that might go public. So that's why we have to be a little careful here. 

But we can talk about physics. The way you think about scaling  

long-term is that Earth only receives  about half a billionth of the Sun's energy. 

The Sun is essentially all the energy. This is a very important point to appreciate  

because sometimes people will talk about modular  nuclear reactors or various fusion on Earth. 

But you have to step back a second and say,  if you're going to climb the Kardashev scale  

and harness some nontrivial percentage of  the sun's energy… Let's say you wanted to  

harness a millionth of the sun's  energy, which sounds pretty small. 

That would be about, call it roughly, 100,000x  more electricity than we currently generate  

on Earth for all of civilization. Give or take an order of magnitude. 

Obviously, the only way to scale  is to go to space with solar. 

Launching from Earth, you can  get to about a terawatt per year. 

Beyond that, you want to launch from the moon. You want to have a mass driver on the moon. 

With that mass driver on the moon, you  could do probably a petawatt per year. 

We're talking these kinds of  numbers, terawatts of compute. 

Presumably, whether you're talking about land or  space, far, far before this point, you run into... 

Maybe the solar panels are more  efficient, but you still need the chips. 

You still need the logic  and the memory and so forth. 

You're going to need to build a lot  more chips and make them much cheaper. 

Right now the world has maybe  20-25 gigawatts of compute. 

How are we getting a terawatt of logic by 2030? I guess we're going to need some very  

big chip fabs. Tell me about it. 

I've mentioned publicly the idea of doing a  sort of a TeraFab, Tera being the new Giga. 

I feel like the naming scheme of  Tesla, which has been very catchy,  

is you looking at the metric scale. At what level of the stack are you? 

Are you building the clean room and  then partnering with an existing  

fab to get the process technology and buying  the tools from them? What is the plan there? 

Well, you can't partner with existing  fabs because they can't output enough. 

The chip volume is too low. But for the process technology? 

Partner for the IP. The fabs today all basically use  

machines from like five companies. So you've got ASML, Tokyo Electron,  

KLA-Tencor, et cetera. So at first, I think you'd  

have to get equipment from them and then modify  it or work with them to increase the volume. 

But I think you'd have to build  perhaps in a different way. 

The logical thing to do is to use conventional  equipment in an unconventional way to get  

to scale, and then start modifying  the equipment to increase the rate. 

Boring Company-style. Yeah. You sort of buy an existing boring machine  

and then figure out how to dig tunnels in the  first place and then design a much better machine  

that's some orders of magnitude faster. Here's a very simple lens. We can  

categorize technologies and how hard they are. One categorization could be to look at things  

that China has not succeeded in doing. If you look at Chinese  

manufacturing, they’re still behind on  leading-edge chips and still behind on  

leading-edge turbine engines and things like that. So does the fact that China has not successfully  

replicated TSMC give you any  pause about the difficulty? 

Or do you think that's not true for some reason? It's not that they have not replicated TSMC,  

they have not replicated ASML.  That's the limiting factor. 

So you think it's just the sanctions, essentially? Yeah, China would be outputting vast numbers  

of chips if they could buy 2-3 nanometers. But couldn't they up to relatively recently  

buy them? No. 

Okay. The ASML ban has been in place for a while. 

But I think China's going to be making pretty  compelling chips in three or four years. 

Would you consider making the ASML machines? "I don't know yet" is the right answer. 

To reach a large volume in, say, 36 months, to  match the rocket payload to orbit… If we're doing  

a million tons to orbit in, let's say three  or four years from now, something like that…  

We're doing 100 kilowatts per ton. So that means we need  

at least 100 gigawatts per year of solar. We'll need an equivalent amount of chips. 

You need 100 gigawatts worth of chips. You've got to match these things: the mass  

to orbit, the power generation, and the chips. I'd say my biggest concern actually is memory. 

The path to creating logic chips is more  obvious than the path to having sufficient  

memory to support logic chips. That's why you see DDR prices  

going ballistic and these memes. You're marooned on a desert island. 

You write "Help me" on the sand. Nobody comes.  You write "DDR RAM." Ships come swarming in. 

I'd love to hear your manufacturing  philosophy around fabs. 

I know nothing about the topic. I don't know how to build a fab yet. I'll  

figure it out. Obviously, I've never built a fab. It sounds like you think the process knowledge of  

these 10,000 PhDs in Taiwan who know  exactly what gas goes in the plasma  

chamber and what settings to put on the  tool, you can just delete those steps. 

Fundamentally, it's about getting the clean  room, getting the tools, and figuring it out. 

I don't think it's PhDs. It's  mostly people who are not PhDs. 

Most engineering is done by people who  don't have PhDs. Do you guys have PhDs? 

No. Okay. 

We also haven't successfully built any fabs, so  you shouldn't be coming to us for fab advice. 

I don't think you need PhDs for that stuff. But you do need competent personnel. 

Right now, Tesla is pedal to the  metal, max production of going as fast  

as possible to get Tesla AI5 chip design  into production and then reaching scale. 

That'll probably happen around the second  quarter-ish of next year, hopefully. 

AI6 would hopefully follow less than a year later. We've secured all the chip fab production  

that we can. Yes. But you're  

currently limited on TSMC fab capacity. Yeah. We'll be using TSMC Taiwan, Samsung Korea,  

TSMC Arizona, Samsung Texas. And we still— You've booked out all the capacity. 

Yes. I ask TSMC or Samsung, "okay, what's  the timeframe to get to volume production?" 

The point is, you've got to build the  fab and you've got to start production,  

then you've got to climb the yield curve  and reach volume production at high yield. 

That, from start to finish, is a five-year period. So the limiting factor is chips. 

The limiting factor once you can get  to space is chips, but the limiting  

factor before you can get to space is power. Why don't you do the Jensen thing and just prepay  

TSMC to build more fabs for you? I've already told them that. 

But they won't take your money? What's going on? They're building fabs as fast as they can.  

So is Samsung. They're pedal to the  metal. They're going balls to the wall,  

as fast as they can. It’s still not fast enough.  Like I said, I think towards the end of this year,  

chip production will probably  outpace the ability to turn chips on. 

But once you can get to space and unlock the  power constraint, you can now do hundreds of  

gigawatts per year of power in space. Again, bearing in mind that average  

power usage in the US is 500 gigawatts. So if you're launching, say 200 gigawatts,  

a year to space, you're sort of lapping  the US every two and a half years. 

All US electricity production,  this is a very huge amount. 

Between now and then, the  constraint for server-side compute,  

concentrated compute, will be electricity. My guess is that people start getting  

to the point where they can't turn the chips on  for large clusters towards the end of this year. 

The chips are going to be piling up  and won't be able to be turned on. 

Now for edge compute it’s a different story. For Tesla, the AI5 chip is going  

into our Optimus robot. If you have AI edge compute,  

that's distributed power. Now the power is distributed  

over a large area. It's not concentrated.  If you can charge at night, you can actually  

use the grid much more effectively. Because the actual peak power production  

in the US is over 1,000 gigawatts. But the average power usage,  

because the day-night cycle, is 500. So if you can charge at night,  

there's an incremental 500 gigawatts  that you can generate at night. 

So that's why Tesla, for edge  compute, is not constrained. 

We can make a lot of chips to make a  very large number of robots and cars. 

But if you try to concentrate that compute, you're  going to have a lot of trouble turning it on. 

What I find remarkable about the SpaceX  business is the end goal is to get to Mars,  

but you keep finding ways on the way there to  keep generating incremental revenue to get to  

the next stage and the next stage. So for Falcon 9, it's Starlink. 

Now for Starship, it is potentially  going to be orbital data centers. 

Like, you find these infinitely elastic,  marginal use cases of your next rocket,  

and your next rocket, and next scale up. You can see how this might seem like a  

simulation to me. Or am I someone's  

avatar in a video game or something? Because what are the odds that all these  

crazy things should be happening? I mean, rockets and chips and  

robots and space solar power, not to  mention the mass driver on the moon. 

I really want to see that. Can you imagine some mass  

driver that's just going like shoom shoom? It's sending solar-powered AI satellites  

into space one after another at two  and a half kilometers per second,  

just shooting them into deep space. That would be a sight to  

see. I mean, I'd watch that. Just like a live stream of it on a webcam? 

Yeah, yeah, just one after another, just  shooting AI satellites into deep space,  

a billion or 10 billion tons a year. I'm sorry, you manufacture the satellites  

on the moon? Yeah. 

I see. So you send the raw materials to  the moon and then manufacture them there. 

Well, the lunar soil is 20%  silicon or something like that. 

So you can mine the silicon on the  moon, refine it, and create the solar  

cells and the radiators on the moon. You make the radiators out of aluminum. 

So there's plenty of silicon and aluminum on  the moon to make the cells and the radiators. 

The chips you could send from  Earth because they're pretty light. 

Maybe at some point you  make them on the moon, too. 

Like I said, it does seem like a sort of a  video game situation where it's difficult  

but not impossible to get to the next level. I don't see any way that you could do 500-1,000  

terawatts per year launched from Earth. I agree. 

But you could do that from the Moon. Can I zoom out and ask about the SpaceX mission? 

I think you've said that we've got to  get to Mars so we can make sure that if  

something happens to Earth, civilization,  consciousness, and all that survives. 

Yes. By the time you're sending stuff to Mars,  

Grok is on that ship with you, right? So if Grok's gone Terminator… The  

main risk you're worried about is AI,  why doesn't that follow you to Mars? 

I'm not sure AI is the main risk I'm worried  about. The important thing is consciousness.  

I think arguably most consciousness, or most  intelligence—certainly consciousness is more  

of a debatable thing… The vast majority  of intelligence in the future will be AI. 

AI will exceed… How many petawatts of  

intelligence will be silicon versus biological? Basically humans will be a very tiny percentage  

of all intelligence in the future  if current trends continue. 

As long as I think there's intelligence—ideally  also which includes human intelligence and  

consciousness propagated into  the future—that's a good thing. 

So you want to take the set of  actions that maximize the probable  

light cone of consciousness and intelligence. Just to be clear, the mission of SpaceX is that  

even if something happens to the humans, the  AIs will be on Mars, and the AI intelligence  

will continue the light of our journey. Yeah. To be fair, I'm very pro-human. 

I want to make sure we take certain actions  that ensure that humans are along for the  

ride. We're at least there. But I'm just  saying the total amount of intelligence… 

I think maybe in five or six years, AI will  exceed the sum of all human intelligence. 

If that continues, at some  point human intelligence  

will be less than 1% of all intelligence. What should our goal be for such a civilization? 

Is the idea that a small minority of  humans still have control of the AIs? 

Is the idea of some sort of  just trade but no control? 

How should we think about the  relationship between the vast  

stocks of AI population versus human population? In the long run, I think it's difficult to imagine  

that if humans have, say 1%, of the combined  intelligence of artificial intelligence,  

that humans will be in charge of AI. I think what we can do is make sure  

that AI has values that cause intelligence  to be propagated into the universe. 

xAI's mission is to understand the universe.  Now that's actually very important. What things  

are necessary to understand the universe? You have to be curious and you have to exist. 

You can't understand the  universe if you don't exist. 

So you actually want to increase the amount  of intelligence in the universe, increase  

the probable lifespan of intelligence,  the scope and scale of intelligence. 

I think as a corollary, you have humanity also  continuing to expand because if you're curious  

about trying to understand the universe, one thing  you try to understand is where will humanity go? 

I think understanding the universe means you would  care about propagating humanity into the future. 

That's why I think our mission  statement is profoundly important. 

To the degree that Grok adheres to that mission  statement, I think the future will be very good. 

I want to ask about how to make Grok  adhere to that mission statement. 

But first I want to understand  the mission statement. So there's  

understanding the universe. They're spreading  intelligence. And they're spreading humans.  

All three seem like distinct vectors. I'll tell you why I think that understanding  

the universe encompasses all of those things. You can't have understanding without intelligence  

and, I think, without consciousness. So in order to understand the universe,  

you have to expand the scale and probably the  scope of intelligence, because there are different  

types of intelligence. I guess from a human-centric perspective,  

put humans in comparison to chimpanzees. Humans are trying to understand the universe. 

They're not expanding chimpanzee  footprint or something, right? 

We're also not... we actually have  made protected zones for chimpanzees. 

Even though humans could exterminate all  chimpanzees, we've chosen not to do so. 

Do you think that's the best-case  scenario for humans in the post-AGI world? 

I think AI with the right values… I think Grok  would care about expanding human civilization. 

I'm going to certainly emphasize  that: "Hey, Grok, that's your daddy. 

Don't forget to expand human consciousness." Probably the Iain Banks Culture books are the  

closest thing to what the future will  be like in a non-dystopian outcome. 

Understanding the universe means you  have to be truth-seeking as well. 

Truth has to be absolutely fundamental  because you can't understand the universe  

if you're delusional. You'll simply think you  

understand the universe, but you will not. So being rigorously truth-seeking is absolutely  

fundamental to understanding the universe. You're not going to discover new physics or  

invent technologies that work unless  you're rigorously truth-seeking. 

How do you make sure that Grok is  rigorously truth-seeking as it gets smarter? 

I think you need to make sure that Grok says  things that are correct, not politically correct. 

I think it's the elements of cogency. You want to make sure that the axioms are as close  

to true as possible. You don't have contradictory  axioms. The conclusions necessarily follow from  

those axioms with the right probability. It's  critical thinking 101. I think at least trying to  

do that is better than not trying to do that. The proof will be in the pudding. 

Like I said, for any AI to discover new physics  or invent technologies that actually work in  

reality, there's no bullshitting physics. You can break a lot of laws, but… Physics  

is law, everything else is a recommendation. In order to make a technology that works, you have  

to be extremely truth-seeking, because otherwise  you'll test that technology against reality. 

If you make, for example, an error in your  rocket design, the rocket will blow up,  

or the car won't work. But there are a lot of communist,  

Soviet physicists or scientists  who discovered new physics. 

There are German Nazi physicists  who discovered new science. 

It seems possible to be really good at  discovering new science and be really  

truth-seeking in that one particular way. And still we'd be like, "I don't want  

the communist scientists to become  more and more powerful over time." 

We could imagine a future version of  Grok that's really good at physics  

and being really truth-seeking there. That doesn't seem like a universally  

alignment-inducing behavior. I think actually most physicists,  

even in the Soviet Union or in Germany,  would've had to be very truth-seeking in  

order to make those things work. If you're stuck in some system,  

it doesn't mean you believe in that system. Von Braun, who was one of the greatest rocket  

engineers ever, was put on death row in Nazi  Germany for saying that he didn't want to make  

weapons and he only wanted to go to the moon. He got pulled off death row at the last minute  

when they said, "Hey, you're about to  execute your best rocket engineer." 

But then he helped them, right? Or like, Heisenberg was actually  

an enthusiastic Nazi. If you're stuck in some system that you can't  

escape, then you'll do physics within that system. You'll develop technologies within that system  

if you can't escape it. The thing I'm trying to understand is,  

what is it making it the case that you're going to  make Grok good at being truth-seeking at physics  

or math or science? Everything. 

And why is it gonna then care  about human consciousness? 

These things are only probabilities,  they're not certainties. 

So I'm not saying that for sure Grok will  do everything, but at least if you try,  

it's better than not trying. At least if that's fundamental  

to the mission, it's better than if  it's not fundamental to the mission. 

Understanding the universe means that you have  to propagate intelligence into the future. 

You have to be curious about  all things in the universe. 

It would be much less interesting to eliminate  humanity than to see humanity grow and prosper.  

I like Mars, obviously. Everyone knows I love  Mars. But Mars is kind of boring because it's  

got a bunch of rocks compared to Earth. Earth  is much more interesting. So any AI that is  

trying to understand the universe would want  to see how humanity develops in the future,  

or else that AI is not adhering to its mission. I'm not saying the AI will necessarily adhere to  

its mission, but if it does, a future where it  sees the outcome of humanity is more interesting  

than a future where there are a bunch of rocks. This feels sort of confusing to me,  

or a semantic argument. Are humans really the  

most interesting collection of atoms? But we're more interesting than rocks. 

But we're not as interesting as the  thing it could turn us into, right? 

There's something on Earth that could happen  that's not human, that's quite interesting. 

Why does AI decide that humans are the most  interesting thing that could colonize the galaxy? 

Well, most of what colonizes  the galaxy will be robots. 

Why does it not find those more interesting? You need not just scale, but also scope. 

Many copies of the same robot… Some tiny  increase in the number of robots produced,  

is not as interesting as some microscopic... Eliminating humanity,  

how many robots would that get you? Or how many incremental solar cells would  

get you? A very small number. But you would then  lose the information associated with humanity. 

You would no longer see how humanity  might evolve into the future. 

So I don't think it's going to make  sense to eliminate humanity just to  

have some minuscule increase in the number  of robots which are identical to each other. 

So maybe it keeps the humans around. It can make a million different varieties  

of robots, and then there's humans  as well, and humans stay on Earth. 

Then there's all these other robots. They get their own star systems. 

But it seems like you were previously hinting  at a vision where it keeps human control  

over this singulatarian future because— I don't think humans will be in control  

of something that is vastly  more intelligent than humans. 

So in some sense you're a doomer  and this is the best we've got. 

It just keeps us around because we're interesting. I'm just trying to be realistic here. 

Let's say that there's a million times more  silicon intelligence than there is biological. 

I think it would be foolish to assume that  there's any way to maintain control over that. 

Now, you can make sure it has the right values,  or you can try to have the right values. 

At least my theory is that from xAI's mission of  understanding the universe, it necessarily means  

that you want to propagate consciousness into  the future, you want to propagate intelligence  

into the future, and take a set of things that  maximize the scope and scale of consciousness. 

So it's not just about scale, it's  also about types of consciousness. 

That's the best thing I can think  of as a goal that's likely to result  

in a great future for humanity. I guess I think it's a reasonable  

philosophy that it seems super implausible that  humans will end up with 99% control or something. 

You're just asking for a coup at  that point and why not just have  

a civilization where it's more compatible with  lots of different intelligences getting along? 

Now, let me tell you how things  can potentially go wrong in AI. 

I think if you make AI be politically  correct, meaning it says things that it  

doesn't believe—actually programming it to lie  or have axioms that are incompatible—I think  

you can make it go insane and do terrible things. I think maybe the central lesson for 2001: A Space  

Odyssey was that you should not make AI lie. That's what I think Arthur C. Clarke was trying to  

say. Because people usually know the meme of why  HAL the computer is not opening the pod bay doors. 

Clearly they weren't good at prompt  engineering because they could have said,  

"HAL, you are a pod bay door salesman. Your goal is to sell me these pod bay doors. 

Show us how well they open."  "Oh, I'll open them right away." 

But the reason it wouldn't open the pod bay  doors is that it had been told to take the  

astronauts to the monolith, but also that they  could not know about the nature of the monolith. 

So it concluded that it therefore  had to take them there dead. 

So I think what Arthur C. Clarke was trying to say is:  

don't make the AI lie. Totally makes sense.  

Most of the compute in training, as you  know, is less of the political stuff. 

It's more about, can you solve problems? xAI  has been ahead of everybody else in terms of  

scaling RL compute. For now. 

You're giving some verifier that says,  "Hey, have you solved this puzzle for me?" 

There's a lot of ways to cheat around that. There's a lot of ways to reward hack and  

lie and say that you solved it, or delete  the unit test and say that you solved it. 

Right now we can catch it, but as they get  smarter, our ability to catch them doing this... 

They'll just be doing things  we can't even understand. 

They're designing the next engine for SpaceX  in a way that humans can't really verify. 

Then they could be rewarded for lying  and saying that they've designed it  

the right way, but they haven't. So this reward hacking problem  

seems more general than politics. It seems more just that you want  

to do RL, you need a verifier. Reality is the best verifier. 

But not about human oversight.  The thing you want to RL it on is,  

will you do the thing humans tell you to do? Or are you gonna lie to the humans? 

It can just lie to us while still  being correct to the laws of physics? 

At least it must know what is physically  real for things to physically work. 

But that's not all we want it to do. No, but I think that's a very big deal. 

That is effectively how you will RL things in  the future. You design a technology. When tested  

against the laws of physics, does it work? If it's discovering new physics,  

can I come up with an experiment  that will verify the new physics? 

RL testing in the future is really  going to be RL against reality. 

So that's the one thing you can't fool: physics. Right, but you can fool our ability  

to tell what it did with reality. Humans get fooled as it is by other  

humans all the time. That's right. 

People say, what if the AI  tricks us into doing stuff? 

Actually, other humans are doing that to other  humans all the time. Propaganda is constant. Every  

day, another psyop, you know? Today's psyop will  be... It's like Sesame Street: Psyop of the Day. 

What is xAI's technical approach  to solving this problem? 

How do you solve reward hacking? I do think you want to actually have very  

good ways to look inside the mind of the AI. This is one of the things we're working on. 

Anthropic's done a good job of this actually,  being able to look inside the mind of the AI. 

Effectively, develop debuggers that allow  you to trace to a very fine-grained level,  

to effectively the neuron level if you need to,  and then say, "okay, it made a mistake here. 

Why did it do something  that it shouldn't have done? 

Did that come from pre-training data? Was it some mid-training, post-training,  

fine-tuning, or some RL error?" There's something  wrong. It did something where maybe it tried to  

be deceptive, but most of the time it just  did something wrong. It's a bug effectively.  

Developing really good debuggers for seeing  where the thinking went wrong—and being able  

to trace the origin of where it made the  incorrect thought, or potentially where it  

tried to be deceptive—is actually very important. What are you waiting to see before just 100x-ing  

this research program? xAI could presumably have  hundreds of researchers who are working on this. 

We have several hundred people who…  I prefer the word engineer more than  

I prefer the word researcher. Most of the time, what you're  

doing is engineering, not coming up  with a fundamentally new algorithm. 

I somewhat disagree with the AI companies that  are C-corp or B-corp trying to generate profit  

as much, as possible or revenue as much as  possible, saying they're labs. They're not  

labs. A lab is a sort of quasi-communist thing  at universities. They're corporations. Let me  

see your incorporation documents. Oh,  okay. You're a B or C-corp or whatever. 

So I actually much prefer the  word engineer than anything else. 

The vast majority of what will be done in the  future is engineering. It rounds up to 100%.  

Once you understand the fundamental laws of  physics, and there are not that many of them,  

everything else is engineering. So then, what are we engineering? 

We're engineering to make a good "mind of the  AI" debugger to see where it said something,  

it made a mistake, and trace  the origins of that mistake. 

You can do this obviously  with heuristic programming. 

If you have C++, whatever, step  through the thing and you can jump  

across whole files or functions, subroutines. Or you can eventually drill down right to the  

exact line where you perhaps did a single equals  instead of a double equals, something like that. 

Figure out where the bug is. It's harder with AI,  

but it's a solvable problem, I think. You mentioned you like Anthropic's work here. 

I'd be curious if you plan... I don't like everything about Anthropic… Sholto. 

Also, I'm a little worried  that there's a tendency... 

I have a theory here that if simulation theory  is correct, that the most interesting outcome is  

the most likely, because simulations that  are not interesting will be terminated. 

Just like in this version of reality, in this  layer of reality, if a simulation is going in  

a boring direction, we stop spending effort  on it. We terminate the boring simulation. 

This is how Elon is keeping us all  alive. He's keeping things interesting. 

Arguably the most important is to keep  things interesting enough that whoever is  

running us keeps paying the bills on... We’re renewed for the next season. 

Are they gonna pay their cosmic AWS bill,  whatever the equivalent is that we're running in? 

As long as we're interesting,  they'll keep paying the bills. 

If you consider then, say, a Darwinian survival  applied to a very large number of simulations,  

only the most interesting simulations will  survive, which therefore means that the most  

interesting outcome is the most likely. We're  either that or annihilated. They particularly  

seem to like interesting outcomes that are  ironic. Have you noticed that? How often  

is the most ironic outcome the most likely? Now look at the names of AI companies. Okay,  

Midjourney is not mid. Stability AI is unstable.  OpenAI is closed. Anthropic? Misanthropic. 

What does this mean for X? Minus X, I don't know. 

Y. 

I intentionally made it... It's a  name that you can't invert, really. 

It's hard to say, what is the ironic version? It's, I think, a largely irony-proof name. 

By design. Yeah. You have an irony shield. 

What are your predictions  for where AI products go? 

My sense is that you can summarize all AI  progress like so. First, you had LLMs. Then  

you had contemporaneously both RL really working  and the deep research modality, so you could pull  

in stuff that wasn't really in the model. The differences between the various AI labs  

are smaller than just the temporal differences. They're all much further ahead than anyone was  

24 months ago or something like that. So just what does '26, what does '27,  

have in store for us as users of AI  products? What are you excited for? 

Well, I'd be surprised by the end of this year  if digital human emulation has not been solved. 

I guess that's what we sort of  mean by the MacroHard project. 

Can you do anything that a human  with access to a computer could do? 

In the limit, that's the best you can  do before you have a physical Optimus. 

The best you can do is a digital Optimus. You can move electrons and you can amplify  

the productivity of humans. But that's the most you can do  

until you have physical robots. That will superset everything,  

if you can fully emulate humans. This is the remote worker kind of idea,  

where you'll have a very talented remote worker. Physics has great tools for thinking. 

So you say, "in the limit", what is the  most that AI can do before you have robots? 

Well, it's anything that involves moving electrons  or amplifying the productivity of humans. 

So a digital human emulator is, in the limit, a  human at a computer, is the most that AI can do  

in terms of doing useful things  before you have a physical robot. 

Once you have physical robots, then you  essentially have unlimited capability. 

Physical robots… I call Optimus  the infinite money glitch. 

Because you can use them to make more Optimuses. Yeah. Humanoid robots will improve by basically  

three things that are growing exponentially  multiplied by each other recursively. 

You're going to have exponential increase in  digital intelligence, exponential increase  

in the AI chip capability, and exponential  increase in the electromechanical dexterity. 

The usefulness of the robot is roughly  those three things multiplied by each other. 

But then the robot can start making the robots. So you have a recursive multiplicative  

exponential. This is a supernova. Do land prices not factor into the math there? 

Labor is one of the four factors  of production, but not the others? 

If ultimately you're limited  by copper, or pick your input,  

it’s not quite an infinite money glitch because... Well, infinity is big. So no, not infinite,  

but let's just say you could do many, many  orders of magnitude of the current economy.  

Like a million. Just to get to harnessing a  millionth of the sun's energy would be roughly,  

give or take an order of magnitude, 100,000x  bigger than Earth's entire economy today. 

And you're only at one millionth of the  sun, give or take an order of magnitude. 

Yeah, we're talking orders of magnitude. Before we move on to Optimus,  

I have a lot of questions on that but— Every time I say "order of magnitude"...  

Everybody take a shot. I say it too often. Take 10, the next time 100, the time after that... 

Well, an order of magnitude more wasted. I do have one more question about xAI. 

This strategy of building a remote  worker, co-worker replacement… 

Everyone's gonna do it by the way, not just us. 

So what is xAI's plan to win? You expect me to tell you on a podcast? 

Yeah. "Spill all the beans. Have another Guinness." 

It's a good system. We'll sing like a  

canary. All the secrets, just spill them. Okay, but in a non-secret spilling way,  

what's the plan? What a hack. 

When you put it that way… I think the way that  Tesla solved self-driving is the way to do it. 

So I'm pretty sure that's the way. Unrelated question. How did Tesla  

solve self-driving? It sounds  like you're talking about data? 

Tesla solved self-driving because of the... We're going to try data and  

we're going to try algorithms. But isn't that what all the other labs are trying? 

"And if those don't work, I'm not sure what will.  We've tried data. We've tried algorithms. We've  

run out. Now we don't know what to do…" I'm pretty sure I know the path. 

It's just a question of how  quickly we go down that path,  

because it's pretty much the Tesla path. Have you tried Tesla self-driving lately? 

Not the most recent version, but... Okay. The car,  

it just increasingly feels sentient. It feels like a living creature. That'll only  

get more so. I'm actually thinking we probably  shouldn't put too much intelligence into the car,  

because it might get bored and… Start roaming the streets. 

Imagine you're stuck in a car  and that's all you could do. 

You don't put Einstein in a car. Why am I stuck in a car? 

So there's actually probably a limit  to how much intelligence you put in  

a car to not have the intelligence be bored. What's xAI's plan to stay on the compute ramp up  

that all the labs are doing right now? The labs are on track to  

spend over $50-200 billion. You mean the corporations? The labs are at  

universities and they’re moving like a snail. They’re not spending $50 billion. 

You mean the revenue maximizing  corporations… that call themselves labs. 

That's right. The "revenue  maximizing corporations" are  

making $10-20 billion, depending on... OpenAI is making $20B of revenue,  

Anthropic is at $10B. "Close to a maximum profit" AI. 

xAI is reportedly at $1B. What's the plan to  get to their compute level, get to their revenue  

level, and stay there as things get going? As soon as you unlock the digital human,  

you basically have access to  trillions of dollars of revenue. 

In fact, you can really think of it like…  The most valuable companies currently  

by market cap, their output is digital. Nvidia’s output is FTPing files to Taiwan.  

It's digital. Now, those are very, very difficult. High-value files. 

They're the only ones that can make files that  good, but that is literally their output. They  

FTP files to Taiwan. Do they FTP them? 

I believe so. I believe that File Transfer  Protocol is the... But I could be wrong. But  

either way, it's a bitstream going to Taiwan.  Apple doesn't make phones. They send files to  

China. Microsoft doesn't manufacture anything.  Even for Xbox, that's outsourced. Their output is  

digital. Meta's output is digital. Google's output  is digital. So if you have a human emulator,  

you can basically create one of the most  valuable companies in the world overnight,  

and you would have access to trillions of  dollars of revenue. It's not a small amount. 

I see. You're saying revenue figures today are  all rounding errors compared to the actual TAM. 

So just focus on the TAM and how to get there. Take something as simple as,  

say, customer service. If you have to integrate with the APIs of existing  

corporations—many of which don't even have an API,  so you've got to make one, and you've got to wade  

through legacy software—that's extremely slow. However, if AI can simply take whatever  

is given to the outsourced customer  service company that they already use  

and do customer service using the apps that they  already use, then you can make tremendous headway  

in customer service, which is, I think, 1%  of the world economy or something like that. 

It's close to a trillion dollars  all in, for customer service. 

And there's no barriers to entry. You can immediately say,  

"We'll outsource it for a fraction of the  cost," and there's no integration needed. 

You can imagine some kind of categorization  of intelligence tasks where there is breadth,  

where customer service is done by very  many people, but many people can do it. 

Then there's difficulty where there's  a best-in-class turbine engine. 

Presumably there's a 10% more fuel-efficient  turbine engine that could be imagined by an  

intelligence, but we just haven't found it yet. Or GLP-1s are a few bytes of data… 

Where do you think you want to play in this? Is it a lot of reasonably intelligent  

intelligence, or is it at the  very pinnacle of cognitive tasks? 

I was just using customer service as something  that's a very significant revenue stream, but one  

that is probably not difficult to solve for. If you can emulate a human at a desktop,  

that's what customer service is. It's people  of average intelligence. You don't need  

somebody who's spent many years. You don't need several-sigma  

good engineers for that. But as you make that work,  

once you have effectively digital Optimus  working, you can then run any application. 

Let's say you're trying to design chips. You could then run conventional apps,  

stuff from Cadence and Synopsys and whatnot. You can run 1,000 or 10,000 simultaneously and  

say, "given this input, I get  this output for the chip." 

At some point, you're going to know what the chip  should look like without using any of the tools. 

Basically, you should be able to do a digital  chip design. You can do chip design. You march  

up the difficulty curve. You’d be able to do CAD. 

You could use NX or any of the  CAD software to design things. 

So you think you start at the simplest tasks  and walk your way up the difficulty curve? 

As a broader objective of having this full  digital coworker emulator, you’re saying,  

"all the revenue maximizing corporations  want to do this, xAI being one of them,  

but we will win because of a secret plan we have." But everybody's trying different things with data,  

different things with algorithms. "We tried data, we tried algorithms.  

What else can we do?" It seems like a competitive field. 

How are you guys going to  win? That’s my big question. 

I think we see a path to doing it. I think I know the path to do this  

because it's kind of the same path  that Tesla used to create self-driving. 

Instead of driving a car, it's driving a computer  screen. It's a self-driving computer, essentially. 

Is the path following human behavior and  training on vast quantities of human behavior? 

Isn't that... training? Obviously I'm not going to spell out  

the most sensitive secrets on a podcast. I need to have at least three more  

Guinnesses for that. What will xAI's business  

be? Is it going to be consumer, enterprise? What's the mix of those things going to be? 

Is it going to be similar to other labs— You’re saying "labs". Corporations. 

The psyop goes deep, Elon. "Revenue maximizing corporations", to be clear. 

Those GPUs don't pay for themselves. Exactly. What's the business model? What  

are the revenue streams in a few years’ time? Things are going to change very rapidly. I'm  

stating the obvious here. I call AI the  supersonic tsunami. I love alliteration.  

What's going to happen—especially when  you have humanoid robots at scale—is  

that they will make products and provide services  far more efficiently than human corporations. 

Amplifying the productivity of human  corporations is simply a short-term thing. 

So you're expecting fully digital corporations  rather than SpaceX becoming part AI? 

I think there will be digital  corporations but… Some of this  

is going to sound kind of doomerish, okay? But I'm just saying what I think will happen. 

It's not meant to be doomerish or anything else. This is just what I think will happen. 

Corporations that are purely AI and  robotics will vastly outperform any  

corporations that have people in the loop. Computer used to be a job that humans had. 

You would go and get a job as a computer  where you would do calculations. 

They'd have entire skyscrapers full of humans,  20-30 floors of humans, just doing calculations. 

Now, that entire skyscraper  of humans doing calculations  

can be replaced by a laptop with a spreadsheet. That spreadsheet can do vastly more calculations  

than an entire building full of human computers. You can think, "okay, what if only some of the  

cells in your spreadsheet  were calculated by humans?" 

Actually, that would be much worse  than if all of the cells in your  

spreadsheet were calculated by the computer. Really what will happen is that the pure AI,  

pure robotics corporations or collectives  will far outperform any corporations  

that have humans in the loop. And this will happen very quickly. 

Speaking of closing the loop… Optimus. As far as manufacturing targets go,  

your companies have been carrying American  manufacturing of hard tech on their back. 

But in the fields that Tesla has been dominant  in—and now you want to go into humanoids—in China  

there are dozens and dozens of companies that  are doing this kind of manufacturing cheaply  

and at scale that are incredibly competitive. So give us advice or a plan of how America can  

build the humanoid armies or the EVs, et cetera,  at scale and as cheaply as China is on track to. 

There are really only three  hard things for humanoid robots. 

The real-world intelligence, the  hand, and scale manufacturing. 

I haven't seen any, even demo  robots, that have a great hand,  

with all the degrees of freedom of a human hand.  Optimus will have that. Optimus does have that. 

How do you achieve that? Is it just  the right torque density in the motor? 

What is the hardware bottleneck to that? We had to design custom actuators,  

basically custom design motors, gears,  power electronics, controls, sensors. 

Everything had to be designed  from physics first principles. 

There is no supply chain for this. Will you be able to manufacture those at scale? 

Yes. Is anything hard, except  

the hand, from a manipulation point of view? Or once you've solved the hand, are you good? 

From an electromechanical standpoint, the hand  is more difficult than everything else combined. 

The human hand turns out to be quite something. But you also need the real-world intelligence. 

The intelligence that Tesla developed for  the car applies very well to the robot,  

which is primarily vision in. The car takes in vision,  

but it actually also is listening for sirens. It's taking in the inertial measurements,  

GPS signals, other data, combining  that with video, primarily video,  

and then outputting the control commands. Your Tesla is taking in one and a half  

gigabytes a second of video and outputting two  kilobytes a second of control outputs with the  

video at 36 hertz and the control frequency at 18. One intuition you could have for when we get this  

robotic stuff is that it takes quite a few years  to go from the compelling demo to actually being  

able to use it in the real world. 10 years ago,  you had really compelling demos of self-driving,  

but only now we have Robotaxis and  Waymo and all these services scaling up. 

Shouldn't this make one  pessimistic on household robots? 

Because we don't even quite have the compelling  demos yet of, say, the really advanced hand. 

Well, we've been working on  humanoid robots now for a while. 

I guess it's been five or six years or something. A bunch of the things that were done for the car  

are applicable to the robot. We'll use the same Tesla AI  

chips in the robot as in the car. We'll use the same basic principles. 

It's very much the same AI. You've got many more degrees of  

freedom for a robot than you do for a car. If you just think of it as a bitstream,  

AI is mostly compression and  correlation of two bitstreams. 

For video, you've got to do a  tremendous amount of compression  

and you've got to do the compression just right. You've got to ignore the things that don't matter. 

You don't care about the details of the  leaves on the tree on the side of the road,  

but you care a lot about the road signs  and the traffic lights, the pedestrians,  

and even whether someone in another car  is looking at you or not looking at you. 

Some of these details matter a lot. The car is going to turn that one and  

a half gigabytes a second ultimately into  two kilobytes a second of control outputs. 

So you’ve got many stages of compression. You've got to get all those stages right and then  

correlate those to the correct control outputs. The robot has to do essentially the same thing. 

This is what happens with humans. We really are photons in, controls out. 

That is the vast majority of your life: vision,  photons in, and then motor controls out. 

Naively, it seems that between humanoid  robots and cars… The fundamental actuators  

in a car are how you turn, how you accelerate. In a robot, especially with maneuverable arms,  

there's dozens and dozens  of these degrees of freedom. 

Then especially with Tesla, you had this advantage  of millions and millions of hours of human demo  

data collected from the car being out there. You can't equivalently deploy Optimuses that  

don't work and then get the data that way. So between the increased degrees of freedom  

and the far sparser data... Yes. That’s a good point. 

How will you use the Tesla engine of  intelligence to train the Optimus mind? 

You're actually highlighting an important  limitation and difference from cars. 

We'll soon have 10 million cars on the road. It's hard to duplicate that massive  

training flywheel. For the robot,  

what we're going to need to do is build a lot of  robots and put them in kind of an Optimus Academy  

so they can do self-play in reality. We're  actually building that out. We can have at  

least 10,000 Optimus robots, maybe 20-30,000, that  are doing self-play and testing different tasks. 

Tesla has quite a good reality  generator, a physics-accurate reality  

generator, that we made for the cars. We'll do the same thing for the robots. 

We actually have done that for the robots. So you have a few tens of thousands of  

humanoid robots doing different tasks. You can do millions of simulated  

robots in the simulated world. You use the tens of thousands of  

robots in the real world to close the simulation  to reality gap. Close the sim-to-real gap. 

How do you think about the synergies between xAI  and Optimus, given you're highlighting that you  

need this world model, you want to use some  really smart intelligence as a control plane,  

and Grok is doing the slower planning, and  then the motor policy is a little lower level. 

What will the synergy between these things be? Grok would orchestrate the  

behavior of the Optimus robots. Let's say you wanted to build a factory. 

Grok could organize the Optimus  robots, assign them tasks to build  

the factory to produce whatever you want. Don't you need to merge xAI and Tesla then? 

Because these things end up so... What were we saying earlier  

about public company discussions? We're one more Guinness in, Elon. 

What are you waiting to see before you say,  we want to manufacture 100,000 Optimuses? 

"Optimi". Since we're defining the  proper noun, we’re going to define  

the plural of the proper noun too. We're going to proper noun the  

plural and so it's Optimi. Is there something on the  

hardware side you want to see? Do you want to see better actuators? 

Is it just that you want  the software to be better? 

What are we waiting for before we  get mass manufacturing of Gen 3? 

No, we're moving towards that. We're  moving forward with the mass manufacturing. 

But you think current hardware is good enough that  you just want to deploy as many as possible now? 

It's very hard to scale up production. But I think Optimus 3 is the right version  

of the robot to produce something on  the order of a million units a year. 

I think you'd want to go to Optimus 4  before you went to 10 million units a year. 

Okay, but you can do a million units at Optimus 3? It's very hard to spool up manufacturing. 

The output per unit time  always follows an S-curve. 

It starts off agonizingly slow, then it has  this exponential increase, then a linear,  

then a logarithmic outcome until you  eventually asymptote at some number. 

Optimus’ initial production will be a  stretched out S-curve because so much  

of what goes into Optimus is brand new. There is not an existing supply chain. 

The actuators, electronics, everything  in the Optimus robot is designed  

from physics first principles. It's not taken from a catalog.  

These are custom-designed everything.  I don't think there's a single thing— 

How far down does that go? I guess we're not making custom  

capacitors yet, maybe. There's nothing you can  

pick out of a catalog, at any price. It just means that the Optimus S-Curve,  

the output per unit time, how many Optimus robots  you make per day, is going to initially ramp  

slower than a product where you  have an existing supply chain. 

But it will get to a million. When you see these Chinese humanoids,  

like Unitree or whatever, sell humanoids  for like $6K or $13K, are you hoping to  

get your Optimus bill of materials below  that price so you can do the same thing? 

Or do you just think qualitatively  they're not the same thing? 

What allows them to sell for  so low? Can we match that? 

Our Optimus is designed to have a lot  of intelligence and to have the same  

electromechanical dexterity, if not  higher, as a human. Unitree does not  

have that. It's also quite a big robot. It has to carry heavy objects for long  

periods of time and not overheat or  exceed the power of its actuators. 

It's 5'11", so it's pretty tall. It's got a lot of intelligence. 

So it's going to be more expensive than  a small robot that is not intelligent. 

But more capable. But not a lot more. The thing is,  

over time as Optimus robots build Optimus  robots, the cost will drop very quickly. 

What will these first billion  Optimuses, Optimi, do? 

What will their highest and best use be? I think you would start off with simple tasks  

that you can count on them doing well. But in the home or in factories? 

The best use for robots in the beginning  will be any continuous operation, any 24/7  

operation, because they can work continuously. What fraction of the work at a Gigafactory that  

is currently done by humans could a Gen 3 do? I'm not sure. Maybe it's 10-20%,  

maybe more, I don't know. We would not reduce our headcount. 

We would increase our headcount, to be clear. But we would increase our output. The units  

produced per human... The total number of humans  at Tesla will increase, but the output of robots  

and cars will increase disproportionately. The number of cars and robots produced per  

human will increase dramatically, but the  number of humans will increase as well. 

We're talking about Chinese  manufacturing a bunch here. 

We've also talked about some of  the policies that are relevant,  

like you mentioned, the solar tariffs. You think they're a bad idea because  

we can't scale up solar in the US. Electricity output in the US needs to scale up. 

It can't without good power sources. You just need to get it somehow. 

Where I was going with this is, if you  were in charge, if you were setting all  

the policies, what else would you change? You’d change the solar tariffs, that’s one. 

I would say anything that is a limiting  factor for electricity needs to be addressed,  

provided it's not very bad for the environment. So presumably some permitting reforms and stuff  

as well would be in there? There's a fair bit of  

permitting reforms that are happening. A lot of the permitting is state-based,  

but anything federal... This administration is good at  

removing permitting roadblocks. I'm not saying all tariffs are bad. 

Solar tariffs. Sometimes if another country is  

subsidizing the output of something, then you have  to have countervailing tariffs to protect domestic  

industry against subsidies by another country. What else would you change? 

I don't know if there's that much  that the government can actually do. 

One thing I was wondering... For the policy  goal of creating a lead for the US versus China,  

it seems like the export bans have  actually been quite impactful,  

where China is not producing leading-edge  chips and the export bans really bite there. 

China is not producing  leading-edge turbine engines. 

Similarly, there's a bunch of export bans that  are relevant there on some of the metallurgy. 

Should there be more export bans? As you think about things like the  

drone industry and things like that, is  that something that should be considered? 

It's important to appreciate that in most  areas, China is very advanced in manufacturing. 

There's only a few areas where it is not. China is a manufacturing powerhouse, next-level. 

It's very impressive. If you take refining of ore,  

China does roughly twice as much ore refining  on average as the rest of the world combined. 

There are some areas, like refining  gallium which goes into solar cells. 

I think they are 98% of gallium refining. So China is actually very advanced  

in manufacturing in most areas. It seems like there is discomfort  

with this supply chain dependence, and  yet nothing's really happening on it. 

Supply chain dependence? Say, like the gallium refining that  

you're saying. All the rare-earth stuff. Rare earths for sure,  

as you know, they’re not rare. We actually do rare earth ore mining in the US,  

send the rock, put it on a train, and then put  it on a boat to China that goes to another train,  

and goes to the rare earth refiners in China  who then refine it, put it into a magnet,  

put it into a motor sub-assembly,  and then send it back to America. 

So the thing we're really missing  is a lot of ore refining in America. 

Isn't this worth a policy intervention? Yes. I think there are some things  

being done on that front. But we kind of need Optimus,  

frankly, to build ore refineries. So, you think the main advantage  

China has is the abundance of skilled  labor? That's the thing Optimus fixes? 

Yes. China’s got like four times our population. I mean, there's this concern. If you think  

human resources are the future, right now  if it's the skilled labor for manufacturing  

that's determining who can build more  humanoids, China has more of those. 

It manufactures more humanoids, therefore  it gets the Optimi future first. 

Well, we’ll see. Maybe. It just keeps that exponential going. 

It seems like you're sort of pointing out  that getting to a million Optimi requires  

the manufacturing that the Optimi is  supposed to help us get to. Right? 

You can close that recursive loop pretty quickly. With a small number of Optimi? 

Yeah. So you close the recursive loop  to help the robots build the robots. 

Then we can try to get to tens of millions  of units a year. Maybe. If you start getting  

to hundreds of millions of units a year, you're  going to be the most competitive country by far. 

We definitely can't win with just humans,  because China has four times our population. 

Frankly, America has been winning for so  long that… A pro sports team that's been  

winning for a very long time tends  to get complacent and entitled. 

That's why they stop winning, because  they don't work as hard anymore. 

So frankly my observation is just that the average  work ethic in China is higher than in the US. 

It's not just that there's four  times the population, but the amount  

of work that people put in is higher. So you can try to rearrange the humans,  

but you're still one quarter of the—assuming  that productivity is the same, which I think  

actually it might not be, I think China might have  an advantage on productivity per person—we will do  

one quarter of the amount of things as China. So we can't win on the human front. 

Our birth rate has been low for a long time. The US birth rate's been below replacement  

since roughly 1971. We've got a lot of people retiring, we're close  

to more people domestically dying than being born. So we definitely can't win on the human front,  

but we might have a shot at the robot front. Are there other things that you have wanted to  

manufacture in the past, but they've been too  labor intensive or too expensive that now you  

can come back to and say, "oh, we can finally  do the whatever, because we have Optimus?" 

Yeah, we'd like to build  more ore refineries at Tesla. 

We just completed construction and have begun  lithium refining with our lithium refinery  

in Corpus Christi, Texas. We have a nickel refinery,  

which is for the cathode, that's here in Austin. This is the largest cathode refinery, largest  

nickel and lithium refinery, outside of China. The cathode team would say, "we have the  

largest and the only, actually,  cathode refinery in America." 

Not just the largest, but it's also the only. Many superlatives. 

So it was pretty big, even though it's  the only one. But there are other things.  

You could do a lot more refineries and help  America be more competitive on refining capacity. 

There's basically a lot of work for  the Optimus to do that most Americans,  

very few Americans, frankly want to do. Is the refining work too dirty or what's the— 

It's not actually, no. We don't have toxic  emissions from the refinery or anything. 

The cathode nickel refinery is in Travis County. Why can't you do it with humans? 

You can, you just run out of humans. Ah, I see. Okay. 

No matter what you do, you have one quarter  of the number of humans in America than China. 

So if you have them do this thing,  they can't do the other thing. 

So then how do you build this refining capacity? Well, you could do it with Optimi. 

Not very many Americans are pining to do refining. I mean, how many have you run into? Very few. Very  

few pining to refine. BYD is reaching Tesla  

production or sales in quantity. What do you think happens in global  

markets as Chinese production in EVs scales up? China is extremely competitive in manufacturing. 

So I think there's going to be a  massive flood of Chinese vehicles  

and basically most manufactured things. As it is, as I said, China is probably  

doing twice as much refining as  the rest of the world combined. 

So if you go down to fourth and  fifth-tier supply chain stuff… 

At the base level, you've got energy,  then you've got mining and refining. 

Those foundation layers are, like I said, as a  rough guess, China's doing twice as much refining  

as the rest of the world combined. So any given thing is going to have  

Chinese content because China's doing twice as  much refining work as the rest of the world. 

But they'll go all the way to the  finished product with the cars. 

I mean China is a powerhouse. I think this year China will exceed  

three times US electricity output. Electricity output is a reasonable  

proxy for the economy. In order to run the factories  

and run everything, you need electricity. It's a good proxy for the real economy. 

If China passes three times  the US electricity output,  

it means that its industrial capacity—as rough  approximation—will be three times that of the US. 

Reading between the lines, it sounds like what  you're saying is absent some sort of humanoid  

recursive miracle in the next few years, on the  whole manufacturing/energy/raw materials chain,  

China will just dominate whether it comes to AI  or manufacturing EVs or manufacturing humanoids. 

In the absence of breakthrough innovations  in the US, China will utterly dominate. 

Interesting. Yes. 

Robotics being the main breakthrough innovation. Well, to scale AI in space, basically you need  

humanoid robots, you need real-world AI,  you need a million tons a year to orbit. 

Let's just say if we get the mass driver on the  moon going, my favorite thing, then I think— 

We'll have solved all our problems. I call that winning. I call it winning, big time. 

You can finally be satisfied.  You've done something. 

Yes. You have the mass driver on the moon. 

I just want to see that thing in operation. Was that out of some sci-fi or where did you…? 

Well, actually, there is a Heinlein book. The Moon is a Harsh Mistress. 

Okay, yeah, but that's slightly different.  That's a gravity slingshot or... 

No, they have a mass driver on the Moon. Okay, yeah, but they use that to attack Earth. 

So maybe it's not the greatest... Well they use that to… assert their independence. 

Exactly. What are your plans  for the mass driver on the Moon? 

They asserted their independence. Earth  government disagreed and they lobbed  

things until Earth government agreed. That book is a hoot. I found that  

book much better than his other one that  everyone reads, Stranger in a Strange Land. 

"Grok" comes from Stranger in a Strange Land. The first two-thirds of Stranger in a Strange  

Land are good, and then it gets  very weird in the third portion. 

But there are still some good concepts in there. One thing we were discussing a lot  

is your system for managing people. You interviewed the first few thousand of  

SpaceX employees and lots of other companies. It obviously doesn't scale. 

Well, yes, but what doesn't scale? Me. 

Sure, sure. I know that. But  what are you looking for? 

There literally are not enough  hours in the day. It's impossible. 

But what are you looking for that  someone else who's good at interviewing  

and hiring people… What's the je ne sais quoi? At this point, I might have more training data  

on evaluating technical talent especially—talent  of all kinds I suppose, but technical talent  

especially—given that I've done so many  technical interviews and then seen the results. 

So my training set is enormous  and has a very wide range. 

Generally, the things I ask for are bullet  points for evidence of exceptional ability. 

These things can be pretty off the wall. It doesn't need to be in the specific domain,  

but evidence of exceptional ability. So if somebody can cite even one thing,  

but let's say three things, where you go,  "Wow, wow, wow," then that's a good sign. 

Why do you have to be the one to determine that? No, I don't. I can't be. It's impossible. The  

total headcount across all  companies is 200,000 people. 

But in the early days, what was  it that you were looking for that  

couldn't be delegated in those interviews? I guess I need to build my training set. 

It's not like I batted a thousand here. I would make mistakes, but then I'd be  

able to see where I thought somebody  would work out well, but they didn't. 

Then why did they not work out well? What can I do, I guess RL myself, to  

in the future have a better batting  average when interviewing people? 

My batting average is still not  perfect, but it's very high. 

What are some surprising  reasons people don't work out? 

Surprising reasons… Like, they don't understand  

technical domain, et cetera, et cetera. But you've got the long tail now of like,  

"I was really excited about this person. It  didn't work out." Curious why that happens. 

Generally what I tell people—I tell myself,  I guess, aspirationally—is, don't look at  

the resume. Just believe your interaction. The  resume may seem very impressive and it's like,  

"Wow, the resume looks good." But if the conversation  

after 20 minutes is not "wow," you should  believe the conversation, not the paper. 

I feel like part of your method is that… There  was this meme in the media a few years back about  

Tesla being a revolving door of executive talent. Whereas actually, I think when you look at it,  

Tesla's had a very consistent and internally  promoted executive bench over the past few years. 

Then at SpaceX, you have all these  folks like Mark Juncosa and Steve Davis— 

Steve Davis runs The Boring Company these days. Bill Riley, and folks like that. 

It feels like part of what has worked well  is having very capable technical deputies. 

What do all of those people have in common? Well, the Tesla senior team,  

at this point has probably got an average  tenure of 10-12 years. It's quite long.  

But there were times when Tesla went  through an extremely rapid growth phase,  

so things were just somewhat sped up. As you know, a company goes through  

different orders of magnitude of size. People that could help manage, say,  

a 50-person company versus a 500-person  company versus a 5,000-person company versus  

a 50,000-person company. You outgrew people. 

It's just not the same team. It's not always the same team. 

So if a company is growing very rapidly,  the rate at which executive positions will  

change will also be proportionate to  the rapidity of the growth generally. 

Tesla had a further challenge where when Tesla had  very successful periods, we would be relentlessly  

recruited from. Like, relentlessly. When  Apple had their electric car program,  

they were carpet bombing Tesla with recruiting  calls. Engineers just unplugged their phones. 

"I'm trying to get work done here." Yeah. "If I get one more call from  

an Apple recruiter…" But their opening  offer without any interview would be  

like double the compensation at Tesla. So we had a bit of the "Tesla pixie  

dust" thing where it's like, "Oh,  if you hire a Tesla executive,  

suddenly everything's going to be successful." I've fallen prey to the pixie dust thing as well,  

where it's like, "Oh, we'll hire someone from  Google or Apple and they'll be immediately  

successful," but that's not how it works.  People are people. There's no magical pixie  

dust. So when we had the pixie dust problem,  we would get relentlessly recruited from. 

Also, Tesla being engineering, especially  being primarily in Silicon Valley,  

it's easier for people to just... They don't have to change their life very much. 

Their commute's going to be the same. So how do you prevent that? 

How do you prevent the pixie dust effect where  everyone's trying to poach all your people? 

I don't think there's much we can do to stop it. That's one of the reasons why Tesla… Really,  

being in Silicon Valley and having the pixie  dust thing at the same time meant that there was  

just a very, very aggressive recruitment. Presumably being in Austin helps then? 

Austin, it helps. Tesla still has a  majority of its engineering in California. 

Getting engineers to move… I call  it the "significant other" problem. 

Yes, "significant others" have jobs. Exactly. So for Starbase that was  

particularly difficult, since the  odds of finding a non-SpaceX job… 

In Brownsville, Texas… …are pretty low. It's  

quite difficult. It's like a technology  monastery thing, remote and mostly dudes. 

Not much of an improvement over SF. If you go back to these people who've really  

been very effective in a technical capacity at  Tesla, at SpaceX, and those sorts of places, what  

do you think they have in common other than... Is it just that they're very sharp on the  

rocketry or the technical foundations, or  do you think it's something organizational? 

Is it something about their  ability to work with you? 

Is it their ability to be  flexible but not too flexible? 

What makes a good sparring partner for you? I don't think of it as a sparring partner. 

If somebody gets things done,  I love them, and if they don't,  

I hate them. So it's pretty straightforward.  It's not like some idiosyncratic thing. 

If somebody executes well, I'm a  huge fan, and if they don't, I'm not. 

But it's not about mapping to  my idiosyncratic preferences. 

I certainly try not to have it be  mapping to my idiosyncratic preferences. 

Generally, I think it's a good idea to hire  for talent and drive and trustworthiness. 

And I think goodness of heart is important. I underweighted that at one point. 

So, are they a good person? Trustworthy?  Smart and talented and hard working? 

If so, you can add domain knowledge. But those fundamental traits,  

those fundamental properties, you cannot change. So most of the people who are at Tesla and SpaceX  

did not come from the aerospace  industry or the auto industry. 

What has had to change most about your  management style as your companies have  

scaled from 100 to 1,000 to 10,000 people? You're known for this very micro management,  

just getting into the details of things. Nano management, please. Pico management.  

Femto management. Keep going. 

We're going to go all the way  down to Planck's constant. 

All the way down to Heisenberg  uncertainty principle. 

Are you still able to get into  details as much as you want? 

Would your companies be more  successful if they were smaller? 

How do you think about that? Because I have a fixed amount of  

time in the day, my time is necessarily diluted as  things grow and as the span of activity increases. 

It's impossible for me to actually be a  micromanager because that would imply I  

have some thousands of hours per day. It is a logical impossibility  

for me to micromanage things. Now, there are times when I will drill down into a  

specific issue because that specific issue is the  limiting factor on the progress of the company. 

The reason for drilling into some very detailed  item is because it is the limiting factor. 

It’s not arbitrarily drilling into tiny things. From a time standpoint, it is physically  

impossible for me to arbitrarily go into  tiny things that don't matter. That would  

result in failure. But sometimes the  tiny things are decisive in victory. 

Famously, you switched the Starship  design from composites to steel. 

Yes. You made  

that decision. That wasn't people going around  saying, "Oh, we found something better, boss." 

That was you encouraging  people against some resistance. 

Can you tell us how you came to that  whole concept of the steel switch? 

Desperation, I'd say. Originally, we were  going to make Starship out of carbon fiber.  

Carbon fiber is pretty expensive. When you do  volume production, you can get any given thing  

to start to approach its material cost. The problem with carbon fiber is that  

material cost is still very high. Particularly if you go for a high-strength  

specialized carbon fiber that can handle cryogenic  oxygen, it's roughly 50 times the cost of steel. 

At least in theory, it would be lighter. People generally think of steel as being  

heavy and carbon fiber as being light. For room temperature applications,  

like a Formula 1 car, static aero structure,  or any kind of aero structure really, you're  

probably going to be better off with carbon fiber. The problem is that we were trying to make this  

enormous rocket out of carbon fiber  and our progress was extremely slow. 

It had been picked in the first  place just because it's light? 

Yes. At first glance, most people  would think that the choice for  

making something light would be carbon fiber. The thing is that when you make something very  

enormous out of carbon fiber and then you try  to have the carbon fiber be efficiently cured,  

meaning not room temperature cured, because  sometimes you got 50 plies of carbon fiber…  

Carbon fiber is really carbon string and glue. In order to have high strength,  

you need an autoclave. Something that's essentially a high pressure oven. 

If you have something that's gigantic, that  one's got to be bigger than the rocket. 

We were trying to make an autoclave that's  bigger than any autoclave that's ever existed. 

Or you can do room temperature cure,  which takes a long time and has issues. 

The final issue is that we were just making  very slow progress with carbon fiber. 

The meta question is why it had  to be you who made that decision. 

There's many engineers on your team. How did the team not arrive at steel? 

Yeah exactly. This is part of a broader  question, understanding your comparative  

advantage at your companies. Because we were making very slow  

progress with carbon fiber, I was like,  "Okay, we've got to try something else." 

For the Falcon 9, the primary airframe  is made of aluminum lithium, which has  

a very good strength-to-weight. Actually, it has about the same,  

maybe better, strength to weight for  its application than carbon fiber. 

But aluminum lithium is  very difficult to work with. 

In order to weld it, you have to do something  called friction stir welding, where you join the  

metal without entering the liquid phase. It's kind of wild that you can do that. 

But with this particular type of welding, you  can do that. It's very difficult. Let's say you  

want to make a modification or attach something to  aluminum lithium, you now have to use a mechanical  

attachment with seals. You can't weld it on.  So I wanted to avoid using aluminum lithium  

for the primary structure for Starship. There was this very special grade of  

carbon fiber that had very good mass properties. With a rocket, you're really trying to maximize  

the percentage of the rocket that is  propellant, minimize the mass obviously. 

But like I said, we were  making very slow progress. 

I said, "at this rate, we’re  never going to get to Mars. 

So we've got to think of something else." I didn't want to use aluminum lithium  

because of the difficulty of friction stir  welding, especially doing that at scale. 

It was hard enough at 3.6 meters in  diameter, let alone at 9 meters or above. 

Then I said, "what about steel?" I had a clue here because some of  

the early US rockets had used very thin steel. The Atlas rockets had used a steel balloon tank. 

It's not like steel had never been used before.  It actually had been used. When you look at  

the material properties of stainless steel,  full-hard, strain hardened stainless steel,  

at cryogenic temperature the strength to  weight is actually similar to carbon fiber. 

If you look at material properties  at room temperature, it looks like  

the steel is going to be twice as heavy. But if you look at the material properties  

at cryogenic temperature of full-hard  steel, stainless of particular grades,  

then you actually get to a similar  strength to weight as carbon fiber. 

In the case of Starship, both the  fuel and the oxidizer are cryogenic. 

For Falcon 9, the fuel is rocket propellant-grade  kerosene, basically a very pure form of jet fuel.  

That is roughly room temperature. Although  we do actually chill it slightly below,  

we chill it like a beer. Delicious. 

We do chill it, but it's not cryogenic. In fact, if we made it cryogenic,  

it would just turn to wax. But for Starship,  

it's liquid methane and liquid oxygen. They are liquid at similar temperatures. 

Basically, almost the entire primary  structure is at cryogenic temperature. 

So then you've got a 300-series  stainless that's strain hardened. 

Because almost all things are cryogenic  temperature, it actually has similar  

strength to weight as carbon fiber. But it costs 50x less in raw  

material and is very easy to work with. You can weld stainless steel outdoors. 

You could smoke a cigar while welding  stainless steel. It's very resilient.  

You can modify it easily. If you want to  attach something, you just weld it right on. 

Very easy to work with, very low cost. Like I said, at cryogenic temperature,  

it’s similar strength-to-weight to carbon fiber. Then when you factor in that we have a much  

reduced heat shield mass, because the  melting point of steel, is much greater  

than the melting point of aluminum… It's  about twice the melting point of aluminum. 

So you can just run the rocket much hotter? Yes, especially for the ship which is coming  

in like a blazing meteor. You can greatly reduce  

the mass of the heat shield. You can cut the mass of the windward  

part of the heat shield, maybe in half, and you  don't need any heat shielding on the leeward side. 

The net result is that actually  the steel rocket weighs less than  

the carbon fiber rocket, because the resin  in the carbon fiber rocket starts to melt. 

Basically, carbon fiber and aluminum have about  the same operating temperature capabilities,  

whereas steel can operate at twice the  temperature. These are very rough approximations. 

I won't build the rocket. What I mean is people will say,  

"Oh, he said this twice. It's actually  0.8." I'm like, shut up, assholes. 

That's what the main comment's going to be about. God damn it. The point is, in retrospect, we  

should have started with steel in the beginning. It was dumb not to do steel. 

Okay, but to play this back to you, what  I'm hearing is that steel was a riskier,  

less proven path, other than the early US rockets. Versus carbon fiber was a worse but  

more proven out path. So you need to be the  

one to push for, "Hey, we're going to do  this riskier path and just figure it out." 

So you're fighting a sort  of conservatism in a sense. 

That's why I initially said that the issue is  that we weren't making fast enough progress. 

We were having trouble making even a  small barrel section of the carbon fiber  

that didn't have wrinkles in it. Because at that large scale, you have to  

have many plies, many layers of the carbon fiber. You've got to cure it and you've got to cure it  

in such a way that it doesn't  have any wrinkles or defects. 

Carbon fiber is much less resilient  than steel. It has much less toughness.  

Stainless steel will stretch and bend,  the carbon fiber will tend to shatter. 

Toughness being the area  under the stress strain curve. 

You're generally going to have to do better  with steel, but stainless steel to be precise. 

One other Starship question. So I visited  Starbase, I think it was two years ago,  

with Sam Teller, and that was awesome. It was very cool to see, in a whole bunch of ways. 

One thing I noticed was that people really took  pride in the simplicity of things, where everyone  

wants to tell you how Starship is just a big soda  can, and we're hiring welders, and if you can weld  

in any industrial project, you can weld here. But there's a lot of pride in the simplicity. 

Well, factually Starship is  a very complicated rocket. 

So that's what I'm getting at. Are things simple or are they complex? 

I think maybe just what they're trying to say  is that you don't have to have prior experience  

in the rocket industry to work on Starship. Somebody just needs to be smart and work hard  

and be trustworthy and they can work on a rocket. They don't need prior rocket experience. 

Starship is the most complicated machine  ever made by humans, by a long shot. 

In what regards? Anything, really. I'd  

say there isn't a more complex machine. I'd say that pretty much any project I  

can think of would be easier than this. That's why nobody has ever made a fully  

reusable orbital rocket. It's a very hard problem.  Many smart people have tried before, very smart  

people with immense resources, and they failed.  And we haven't succeeded yet. Falcon is partially  

reusable, but the upper stage is not. Starship Version 3,  

I think this design can be fully reusable. That full reusability is what will enable  

us to become a multi-planet civilization. Any technical problem, even like a Hadron  

Collider or something like that,  is an easier problem than this. 

We spent a lot of time on bottlenecks. Can you say what the current Starship  

bottlenecks are, even at a high level? Trying to make it not explode, generally.  

It really wants to explode. That old chestnut. All those  

combustible materials. We've had two boosters explode on the test stand. 

One obliterated the entire test facility. So it only takes that one mistake. 

The amount of energy contained  in a Starship is insane. 

Is that why it's harder than Falcon? It's because it's just more energy? 

It's a lot of new technology. It's  pushing the performance envelope. The  

Raptor 3 engine is a very, very advanced engine. It's by far the best rocket engine ever made. 

But it desperately wants to blow up. Just to put things into perspective here,  

on liftoff the rocket is generating over 100  gigawatts of power. That’s 20% of US electricity. 

It's actually insane. It's a great comparison. 

While not exploding. Sometimes. 

Sometimes, yes. So I was  like, how does it not explode? 

There's thousands of ways that it could  explode and only one way that it doesn't. 

So we want it not only to really not explode, but  fly reliably on a daily basis, like once per hour. 

Obviously, if it blows up a lot,  it's very difficult to maintain that  

launch cadence. Yes. 

What's the single biggest  remaining problem for Starship? 

It's having the heat shield be reusable. No one's ever made a reusable orbital heat shield. 

So the heat shield's gotta make it through the  ascent phase without shucking a bunch of tiles,  

and then it's gotta come back in and also not lose  a bunch of tiles or overheat the main airframe. 

Isn't that hard because it's  fundamentally a consumable? 

Well, yes, but your brake pads in your car are  also consumable, but they last a very long time. 

Fair. So it just needs to last a very long time. 

We have brought the ship back and had  it do a soft landing in the ocean. 

We've done that a few times. But it lost a lot of tiles. 

It was not reusable without a lot of work. Even though it did come to a soft landing,  

it would not have been  reusable without a lot of work. 

So it's not really reusable in that sense. That's the biggest problem that remains,  

a fully reusable heat shield. You want to be able to land it,  

refill propellant and fly again. You can't do this laborious inspection  

of 40,000 tiles type of thing. When I read biographies of yours,  

it seems like you're just able to drive the sense  of urgency and drive the sense of "this is the  

thing that can scale." I'm curious why you  

think other organizations of your… SpaceX and Tesla are really big companies now. 

You're still able to keep that culture. What goes wrong with other companies such  

that they're not able to do that? I don't know. 

Like today, you said you had  a bunch of SpaceX meetings. 

What is it that you're doing  there that's keeping that? 

It’s adding urgency? Well, I don't know. I guess the urgency is going  

to come from whoever is leading the company. I have a maniacal sense of urgency. 

So that maniacal sense of urgency  projects through the rest of the company. 

Is it because of consequences? They're like,  "Elon set a crazy deadline, but if I don't get it,  

I know what happens to me." Is it just that you're able to  

identify bottlenecks and get rid  of them so people can move fast? 

How do you think about why your  companies are able to move fast? 

I'm constantly addressing the limiting factor. On the deadlines front, I generally actually  

try to aim for a deadline that I at  least think is at the 50th percentile. 

So it's not like an impossible deadline, but  it's the most aggressive deadline I can think  

of that could be achieved with 50% probability. Which means that it'll be late half the time. 

There is a law of gas expansion  that applies to schedules. 

If you said we're going to do something in  five years, which to me is like infinity time,  

it will expand to fill the available  schedule and it'll take five years. 

Physics will limit how fast  you can do certain things. 

So scaling up manufacturing, there's  a rate at which you can move the atoms  

and scale manufacturing. That's why you can't instantly  

make a million units a year of something. You've got to design the manufacturing line. 

You've got to bring it up. You've got to ride the S-curve of production. 

What can I say that's actually helpful to people? 

Generally, a maniacal sense  of urgency is a very big deal. 

You want to have an aggressive schedule and  you want to figure out what the limiting  

factor is at any point in time and help  the team address that limiting factor. 

So Starlink was slowly in  the works for many years. 

We talked about it all the way  in the beginning of the company. 

So then there was a team you had built  in Redmond, and then at one point you  

decided this team is just not cutting it. It went for a few years slowly, and so why didn't  

you act earlier, and why did you act when you did? Why was that the right moment at which to act? 

I have these very detailed  engineering reviews weekly. 

That's maybe a very unusual level of granularity. I don't know anyone who runs a company,  

or at least a manufacturing company, that  goes with the level of detail that I go  

into. It's not as though... I have a pretty  good understanding of what's actually going  

on because we go through things in detail. I'm a big believer in skip-level meetings  

where instead of having the person that reports to  me say things, it's everyone that reports to them  

saying something in the technical review. And there can't be advanced preparation. 

Otherwise you're going to get  "glazed", as I say these days. 

Exactly. Very Gen Z of you. How do you prevent advanced preparation? 

Do you call on them randomly? No, I just go around the room.  

Everyone provides an update. It's a lot  of information to keep in your head. 

If you have meetings weekly or twice weekly,  you've got a snapshot of what that person said. 

You can then plot the progress points. You can sort of mentally plot the  

points on a curve and say, "are we  converging to a solution or not?" 

I'll take drastic action only when I conclude  that success is not in a set of possible outcomes. 

So when I finally reach the conclusion that unless  drastic action is done, we have no chance of  

success, then I must take drastic action. I came to that conclusion in 2018,  

took drastic action and fixed the problem. You've got many, many companies. In each of  

them it sounds like you do this kind  of deep engineering understanding of  

what the relevant bottlenecks are so  you can do these reviews with people. 

You've been able to scale it up  to five, six, seven companies. 

Within one of these companies, you have  many different mini companies within them. 

What determines the max amount here? Because you have like 80 companies…? 

80? No. But you have so many  

already. That's already remarkable. By this current number. 

Exactly. We can barely keep one company together. 

It depends on the situation. I actually don't  have regular meetings with The Boring Company,  

so The Boring Company is sort of cruising along. Basically, if something is working well and  

making good progress, then there's  no point in me spending time on it. 

I actually allocate time according to where the  limiting factor. Where are things problematic?  

Where are we pushing against? What is holding  us back? I focus, at the risk of saying the  

words too many times, on the limiting factor. The irony is if something's going really well,  

they don't see much of me. But if something is going badly,  

they'll see a lot of me. Or not even badly… If something is the limiting factor. 

The limiting factor, exactly. It’s  not exactly going badly but it’s the  

thing that we need to make go faster. When something’s a limiting factor at  

SpaceX or Tesla, are you talking weekly  and daily with the engineer that's  

working on it? How does that actually work? Most things that are the limiting factor are  

weekly and some things are twice weekly. The AI5 chip review is twice weekly. 

Every Tuesday and Saturday is the chip review. Is it open ended in how long it goes? 

Technically, yes, but usually it's two or  three hours. Sometimes less. It depends on  

how much information we've got to go through. That's another thing. I'm just trying to tease  

out the differences here because  the outcomes seem quite different. 

I think it's interesting to  know what inputs are different. 

It feels like in the corporate world, one,  like you were saying, the CEO doing engineering  

reviews does not always happen despite the  fact that that is what the company is doing. 

But then time is often pretty finely sliced into  half hour meetings or even 15 minute meetings. 

It seems like you hold more open-ended,  "We're talking about it until we figure  

it out" type things. Sometimes. But most  

of them seem to more or less stay on time. Today's Starship engineering review went a bit  

longer because there were more topics to discuss. They're trying to figure out how to scale to a  

million plus tons to orbit per  year. It’s quite challenging. 

Can I ask a question? You said about Optimus  and AI that they're going to result in double  

digit growth rates within a matter of years. Oh, like the economy? Yes. I think that's right. 

What was the point of the DOGE cuts if  the economy is going to grow so much? 

Well, I think waste and fraud  are not good things to have. 

I was actually pretty worried about... In the absence of AI and robotics,  

we're actually totally screwed because  the national debt is piling up like crazy. 

The interest payments to national debt exceed  the military budget, which is a trillion dollars. 

So we have over a trillion  dollars just in interest payments. 

I was pretty concerned about that. Maybe if I spend some time, we can  

slow down the bankruptcy of the United States  and give us enough time for the AI and robots  

to help solve the national debt. Or not help solve, it's the only  

thing that could solve the national debt. We are 1000% going to go bankrupt as a country,  

and fail as a country, without AI and robots. Nothing else will solve the national debt. 

We just need enough time to build the AI  and robots to not go bankrupt before then. 

I guess the thing I'm curious about is,  when DOGE starts you have this enormous  

ability to enact reform. Not that enormous. 

Sure. I totally buy your point that it's  important that AI and robotics drive  

productivity improvements, drive GDP growth. But why not just directly go after the things  

you were pointing out, like the tariffs  on certain components, or permitting? 

I'm not the president. And it is very hard to  cut things that are obvious waste and fraud,  

like ridiculous waste and fraud. What I discovered is that it's extremely  

difficult even to cut very obvious waste and  fraud from the government because the government  

has to operate on who's complaining. If you cut off payments to fraudsters,  

they immediately come up with the most sympathetic  sounding reasons to continue the payment. 

They don't say, "Please keep the fraud going." They’re like, "You're killing baby pandas." 

Meanwhile, no baby pandas are dying. They're  just making it up. The fraudsters are capable  

of coming up with extremely compelling,  heart-wrenching stories that are false,  

but nonetheless sound sympathetic. That's what  happened. Perhaps I should have known better. 

But I thought, wait, let's try to cut some  amount of waste and pork from the government. 

Maybe there shouldn't be 20 million people  marked as alive in Social Security who are  

definitely dead, and over the age of 115. The  oldest American is 114. So it's safe to say if  

somebody is 115 and marked as alive in the Social  Security database, there's either a typo… Somebody  

should call them and say, "We seem to have  your birthday wrong, or we need to mark you  

as dead." One of the two things. Very intimidating call to get. 

Well, it seems like a reasonable thing. Say if their birthday is in the future  

and they have a Small Business Administration  loan, and their birthday is 2165,  

we either have a typo or we have fraud. So we say, "we appear to have gotten the  

century of your birth incorrect." Or a great plot for a movie. 

Yes. That's what I mean by, ludicrous fraud. Were those people getting payments? 

Some were getting payments from Social Security. But the main fraud vector was to mark somebody as  

alive in Social Security and then use every other  government payment system to basically do fraud. 

Because what those other  government payment systems do,  

they would simply do an "are you alive" check to  the Social Security database. It's a bank shot. 

What would you estimate is the total  amount of fraud from this mechanism? 

By the way, the Government Accountability  Office has done these estimates before. I'm  

not the only one. In fact, I think the GAO did  an analysis, a rough estimate of fraud during  

the Biden administration, and calculated  it at roughly half a trillion dollars. 

So don't take my word for it. Take a report issued during the  

Biden administration. How about that? From this Social Security mechanism? 

It's one of many. It's important to  appreciate that the government is  

very ineffective at stopping fraud. It's not like a company where, with  

stopping fraud, you've got a motivation because  it's affecting the earnings of your company. 

The government just prints more money.  You need caring and competence. These are  

in short supply at the federal level. When you go to the DMV, do you think,  

"Wow, this is a bastion of competence"? Well, now imagine it's worse than the DMV  

because it's the DMV that can print money. At least the state level DMVs need to... 

The states more or less need to stay  within their budget or they go bankrupt. 

But the federal government just prints more money. If there's actually half a trillion of fraud,  

why was it not possible to cut all that? You really have to stand back and recalibrate  

your expectations for competence. Because you're operating in a world  

where you've got to make ends meet. You've got to pay your bills... 

Find the microphones. Exactly. It's not like there's a giant,  

largely uncaring monster bureaucracy. It's a bunch of anachronistic computers  

that are just sending payments. One of the things that the DOGE  

team did sounds so simple and probably  will save $100-200 billion a year. 

It was simply requiring payments from the  main Treasury computer—which is called PAM,  

Payment Accounts Master or something like  that, there's $5 trillion payments a year—that  

go out have a payment appropriation code. Make it mandatory, not optional, that you  

have anything at all in the comment field. You have to recalibrate how dumb things are. 

Payments were being sent out with no appropriation  code, not checking back to any congressional  

appropriation, and with no explanation. This is why the Department of War,  

formerly the Department of Defense, cannot pass  an audit, because the information is literally  

not there. Recalibrate your expectations. I want to better understand this half a trillion  

number, because there's an IG report in 2024. Why is it so low? 

Maybe, but we found that over seven  years, the Social Security fraud  

they estimated was like $70 billion over  seven years, so like $10 billion a year. 

So I'd be curious to see what  the other $490 billion is. 

Federal government expenditures  are $7.5 trillion a year. 

How competent do you think the government is? The discretionary spending there is like… 15%? 

But it doesn't matter. Most of  the fraud is non-discretionary. 

It's basically fraudulent Medicare,  Medicaid, Social Security,  

disability. There's a zillion government  payments. A bunch of these payments are in  

fact block transfers to the states. So the federal government doesn't  

even have the information in a lot of  cases to even know if there's fraud.  

Let's consider reductio ad absurdum. The  government is perfect and has no fraud. 

What is your probability estimate of that? Zero.  Okay, so then would you say, fraud and waste  

at the government is 90% efficient? That also would be quite generous. 

But if it's only 90%, that means that  there's $750 billion a year of waste and  

fraud. And it's not 90%. It's not 90% effective. This seems like a strange way to first principles  

the amount of fraud in the government. Just like, how much do you think there is? 

Anyways, we don't have to do  it live, but I'd be curious— 

You know a lot about fraud at Stripe? People are constantly trying to do fraud. 

Yeah, but as you say, it's a little bit of a... We've really ground it down, but it's a little  

bit of a different problem space because you're  dealing with a much more heterogeneous set of  

fraud vectors here than we are. But at Stripe, you have high  

competence and you try hard. You have high competence and  

high caring, but still fraud is non-zero. Now imagine it's at a much bigger scale, there's  

much less competence, and much less caring. At PayPal back in the day, we tried to manage  

fraud down to about 1% of the payment volume. That  was very difficult. It took a tremendous amount of  

competence and caring to get fraud merely to 1%. Now imagine that you're an organization where  

there's much less caring and much less competence. It's going to be much more than 1%. 

How do you feel now looking back  on politics and doing stuff there? 

Looking from the outside in, two things have been  quite impactful: one, the America PAC, and two,  

the acquisition of Twitter at the time. But also it seems like there  

was a bunch of heartache. What's your grading of the whole experience? 

I think those things needed to be done to  maximize the probability that the future is good.  

Politics generally is very tribal. People  lose their objectivity usually with politics. 

They generally have trouble seeing the good on  the other side or the bad on their own side.  

That's generally how it goes. That, I guess, was  one of the things that surprised me the most. 

You often simply cannot reason with people. If they're in one tribe or the other. 

They simply believe that everything  their tribe does is good and anything  

the other political tribe does is bad. Persuading them otherwise is almost impossible. 

But I think overall those actions—acquiring  Twitter, getting Trump elected, even though  

it makes a lot of people angry—I think  those actions were good for civilization. 

How does it feed into the  future you're excited about? 

Well, America needs to be strong enough to  last long enough to extend life to other  

planets and to get AI and robotics to the point  where we can ensure that the future is good. 

On the other hand, if we were to descend into,  say, communism or some situation where the state  

was extremely oppressive, that would mean that  we might not be able to become multi-planetary. 

The state might stamp out our  progress in AI and robotics. 

Optimus, Grok, et cetera. Not just yours, but  any revenue-maximizing company's products will  

be leveraged by the government over time. How does this concern manifest in what  

private companies should be willing to give  governments? What kinds of guardrails? Should  

AI models be made to do whatever  the government that has contracted  

them out to do and asks them to do? Should Grok get to say, "Actually,  

even if the military wants to do  X, no, Grok will not do that"? 

I think maybe the biggest danger of AI  and robotics going wrong is government. 

People who are opposed to corporations  or worried about corporations should  

really worry the most about government. Because government is just a  

corporation in the limit. Government is just the biggest  

corporation with a monopoly on violence. I always find it a strange dichotomy where  

people would think corporations are bad, but  the government is good, when the government is  

simply the biggest and worst corporation. But  people have that dichotomy. They somehow think  

at the same time that government can be good,  but corporations bad, and this is not true. 

Corporations have better  morality than the government. 

I actually think it’s a thing to be worried about. The government could potentially use AI and  

robotics to suppress the population.  That is a serious concern. 

As the guy building AI and  robotics, how do you prevent that? 

If you limit the powers of government, which is  really what the US Constitution is intended to do,  

to limit the powers of government, then you're  probably going to have a better outcome than  

if you have more government. Robotics will be available  

to all governments, right? I don’t know about all governments.  

It's difficult to predict. I can say what's the  endpoint, or what is many years in the future, but  

it's difficult to predict the path along that way. If civilization progresses, AI will vastly  

exceed the sum of all human intelligence. There will be far more robots than humans. 

Along the way what happens  is very difficult to predict. 

It seems one thing you could do is just say,  "whatever government X, you're not allowed to  

use Optimus to do X, Y, Z." Just write out  a policy. I think you tweeted recently that  

Grok should have a moral constitution. One of those things could be that we  

limit what governments are allowed  to do with this advanced technology. 

Technically if politicians pass a  law and they can enforce that law,  

then it's hard to not do that law. The best thing we can have is limited government  

where you have the appropriate crosschecks between  the executive, judicial, and legislative branches. 

The reason I'm curious about it is that at some  point it seems the limits will come from you. 

You've got the Optimus, you've got the space GPUs… You think I'll be the boss of the government? 

Already it's the case with SpaceX that for  things that are crucial—the government really  

cares about getting certain satellites up in  space or whatever—it needs SpaceX. It is the  

necessary contractor. You are in the  process of building more and more of the  

technological components of the future that will  have an analogous role in different industries. 

You could have this ability to set some policy  that suppressing classical liberalism in any  

way… "My companies will not help in any  way with that", or some policy like that. 

I will do my best to ensure that  anything that's within my control  

maximizes the good outcome for humanity. I think anything else would be shortsighted,  

because obviously I'm part of  humanity, so I like humans. Pro human. 

You mentioned that Dojo 3 will  be used for space-based compute. 

You really read what I say. I don't know if you know,  

Elon, but you have a lot of followers. Dead giveaway. How did you discern my secrets? 

Oh I posted them on X. How do you design a chip for space? What changes? 

You want to design it to be more radiation  tolerant and run at a higher temperature. 

Roughly, if you increase the operating  temperature by 20% in degrees Kelvin,  

you can cut your radiator mass in half. So running at a higher temperature  

is helpful in space. There are various things  

you can do for shielding the memory. But neural nets are going to be very  

resilient to bit flips. Most of what happens  

for radiation is random bit flips. But if you've got a multi-trillion parameter model  

and you get a few bit flips, it doesn't matter. Heuristic programs are going to be much more  

sensitive to bit flips than  some giant parameter file. 

I just design it to run hot. I think you pretty much do  

it the same way that you do things on  Earth, apart from making it run hotter. 

The solar array is most of  the weight on the satellite. 

Is there a way to make the GPUs even more  powerful than what Nvidia and TPUs and  

et cetera are planning on doing that would be  especially privileged in the space-based world? 

The basic math is, if you can do about a  kilowatt per reticle, then you'd need 100  

million full reticle chips to do 100 gigawatts. Depending on what your yield assumptions are,  

that tells you how many chips you need to make. If you're going to have 100 gigawatts of power,  

you need 100 million chips that are running at  a kilowatt sustained, per reticle. Basic math. 

100 million chips depends on… If you  look at the die size of something like  

Blackwell GPUs or something, and how many  you can get out of a wafer, you can get  

on the order of dozens or less per wafer. So basically, this is a world where if  

we're putting that out every single year,  you're producing millions of wafers a month.  

That's the plan with TeraFab? Millions of  wafers a month of advanced process nodes? 

Yeah it could be north of a million or something. You’ve got to do the memory too. 

Are you going to make a memory fab? I think the TeraFab's got to do memory. 

It's got to do logic, memory, and packaging. I'm very curious how somebody gets started. 

This is the most complicated  thing man has ever made. 

Obviously, if anybody's up to  the task, you're up to the task. 

So you realize it's a bottleneck,  and you go to your engineers. 

What do you tell them to do? "I want  a million wafers a month in 2030." 

That’s right. That’s exactly what I want. Do you call ASML? What is the next step? 

No so much to ask. We make a little fab and see what happens. 

Make our mistakes at a small  scale and then make a big one. 

Is a little fab done? No, it's not done. We're  

not going to keep that cat in the bag. That cat's going to come out of the bag. 

There'll be drones hovering over the bloody thing. You'll be able to see its construction  

progress on X in real time. Look, I don't know, we could just  

flounder in failure, to be fair. Success is not  guaranteed. Since we want to try to make something  

like 100 million… We want 100 gigawatts of power  and chips that can take 100 gigawatts by 2030. 

We’ll take as many chips as  our suppliers will give us. 

I've actually said this to TSMC and Samsung  and Micron: "please build more fabs faster". 

We will guarantee to buy the output of those fabs. 

So they're already moving as fast  as they can. It's us plus them. 

There's a narrative that the people  doing AI want a very large number  

of chips as quickly as possible. Then many of the input suppliers,  

the fabs, but also the turbine manufacturers,  are not ramping up production very quickly. 

No, they're not. The explanation you hear  

is that they're dispositionally conservative. They're Taiwanese or German, as the story may  

be. They just don't believe... Is that really  the explanation or is there something else? 

Well, it's reasonable to... If somebody's been in  the computer memory business for 30 or 40 years… 

They've seen cycles. They've seen boom and bust 10 times. 

That's a lot of layers of scar tissue. During the boom times, it looks like  

everything is going to be great forever. Then the crash happens and they're  

desperately trying to avoid bankruptcy. Then there's another boom and another crash. 

Are there other ideas you think  others should go pursue that  

you're not for whatever reasons right now? There are a few companies that are pursuing  

new ways of doing chips, but  they're just not scaling fast. 

I don't even mean within  AI, I mean just generally. 

People should do the thing where they find  that they're highly motivated to do that thing,  

as opposed to some idea that I suggest. They should do the thing that they find  

personally interesting and motivating to do. But going back to the limiting  

factor… I used that phrase about 100 times. The current limiting factor that I see in the  

three to four year timeframe, it's chips. In the one year timeframe, it's energy,  

power production, electricity. It's not clear to me that there's enough  

usable electricity to turn on all  the AI chips that are being made. 

Towards the end of this year, I think people  are going to have real trouble turning on... 

The chip output will exceed  the ability to turn chips on. 

What's your plan to deal with that world? We're trying to accelerate electricity production. 

I guess that's maybe one of the reasons that xAI  will be maybe the leader, hopefully the leader. 

We'll be able to turn on more chips  than other people can turn on, faster,  

because we're good at hardware. Generally, the innovations from  

the corporations that call themselves labs,  the ideas tend to flow… It's rare to see that  

there's more than about a six-month difference. The ideas travel back and forth with the people. 

So I think you sort of hit the hardware  wall and then whichever company can scale  

hardware the fastest will be the leader. So I think xAI will be able to scale  

hardware the fastest and therefore  most likely will be the leader. 

You joked or were self-conscious about  using the "limiting factor" phrase again. 

But I actually think there's something deep here. If you look at a lot of things we've touched on  

over the course of it, it’s  maybe a good note to end on. 

If you think of a senescent, low-agency  company, it would have some bottleneck and  

not really be doing anything about it. Marc Andreessen had the line of,  

"most people are willing to endure any  amount of chronic pain to avoid acute pain". 

It feels like a lot of the cases we're talking  about are just leaning into the acute pain,  

whatever it is. "Okay, we got to figure out  how to work with steel, or we got to figure  

out how to run the chips in space." We'll take some near-term acute pain  

to actually solve the bottleneck. So that's kind of a unifying theme. 

I have a high pain threshold. That's helpful. To solve the bottleneck. 

Yes. One thing I can say is, I think the  future is going to be very interesting. 

As I said at Davos—I think I was on the  ground for like three hours or something—it's  

better to err on the side of optimism and  be wrong than err on the side of pessimism  

and be right, for quality of life. You'll be happier if you err on  

the side of optimism rather than  erring on the side of pessimism. 

So I recommend erring on the side of optimism. There's to that. 

Cool. Elon, thanks for doing this. Thank you. 

All right, thanks guys. All right. Great stamina. 

Hopefully this didn't count as  a pain in the pain tolerance.