Elon Musk's 3-hour in-depth interview: The lowest-cost AI in space within three years, Optimus is an "infinite money printer"

Elon Musk made bold predictions, believing that due to the Earth's power bottleneck, space will become the most economically viable place for AI reasoning within three years, and described a vision of building space computing power through thousands of Starship launches, unlocking digital humans through xAI, and mass-producing humanoid robots Optimus, bringing infinite capabilities and wealth.

In a deep three-hour interview that just concluded on February 6th, Elon Musk engaged in an in-depth dialogue with Dwarkesh Patel and Stripe co-founder John Collison. Musk not only addressed market concerns about the synergies of his various business lines but also threw out a series of aggressive guidance regarding computing infrastructure, mass production of humanoid robots, and future revenue potential.

The cheapest place to deploy AI in the next 36 months will be space

The market is generally concerned about the sustainability of AI computing power growth, and Musk predicts that within 36 months, or even 30 months, deploying AI in space will become the lowest-cost option, and this advantage will continue to grow.

Musk pointed out that, except for China, global power output is basically flat, while chip production is growing exponentially. He bluntly stated:

“The output of chips is growing exponentially, but power output is flat. So how do you plan to power the chips? With magic energy? Magic electricity sprites?”

He predicts that as chip production capacity is released, by the end of this year, we will face a situation where chips are piling up but cannot be powered on.

He explained the economic rationale behind this: In space, there is no day-night cycle, cloud cover, or atmospheric interference, solar panels are five times more efficient than on the ground, and there is no need for expensive battery storage systems. When all is considered, the cost advantage of deploying in space will be an order of magnitude.

When asked how to maintain frequently malfunctioning GPUs in space, Musk responded:

“Actually, it depends on how new the GPU is when it arrives. Currently, we find GPUs to be quite reliable.”

Five years from now, AI computing power launched into space annually will exceed Earth's total stock

Musk's goal is extremely ambitious: “In five years, I expect the AI computing power we launch and operate in space each year will exceed the total AI computing power on Earth.” This plan requires launching about 100 gigawatts (GW) of solar and computing payloads into orbit each year, equivalent to 10,000 Starship launch missions.

Dwarkesh Patel said:

"This means about one Starship launch every hour."

This also means that SpaceX will not only be a rocket company but will become a "super service provider" in the AI era

xAI Business Model: A Trillion-Dollar "Digital Employee" Market

In response to market concerns about how xAI will monetize and catch up with competitors like OpenAI, Musk has set his sights on "digital human emulation," an AI capable of simulating humans completing various tasks at a computer.

He believes that current tech giants are essentially "digital output" companies, whether it's NVIDIA transferring files to desktops or Apple transferring design drawings to China. Once digital humans, or "digital version of Optimus," are realized, the company will have a revenue potential of trillions of dollars.

Regarding when this can be achieved, Musk stated:

"Well, by the end of this year, if digital human simulation hasn't been solved, I would be surprised. This is actually the limit of what can be done before having physical robots."

Musk revealed that xAI's winning strategy lies in rapid iteration at the hardware level. While other labs are still constrained by traditional supply chains, xAI leverages the vertical integration capabilities of Tesla and SpaceX, particularly the rapid establishment of power infrastructure (such as building the supercomputing cluster Colossus in 19 days), which will create an insurmountable computational barrier.

Optimus: An Infinite Money Printer and a Breakthrough for American Manufacturing

Musk refers to the humanoid robot Optimus as an infinite money printer. He explained that digital intelligence, AI chip capabilities, and electromechanical dexterity are all growing exponentially, and their product is the robot's capability. He also pointed out that robots can achieve low-cost self-manufacturing, which will lead to significant economic scale expansion.

Musk emphasized that Optimus is not only Tesla's future growth engine but also the only hope for the U.S. to maintain its manufacturing competitiveness. He highly praised China's manufacturing capabilities, stating that China holds an absolute dominant position in areas like basic mineral refining, "probably twice that of the rest of the world combined."

When discussing U.S.-China competition, Musk stated:

"If there is no (U.S.) breakthrough innovation, China will utterly dominate. We absolutely cannot win in terms of manpower because China's population is four times ours, but we might have a fighting chance on the robotics front."

Supply Chain Restructuring: Building "TeraFab" and Vertical Integration

To support the ambitious computational and robotics plans, Musk proposed building "TeraFab" (terawatt-level chip factories). He pointed out that the current supply chain (transformers, chip photomasks, and even turbine blades) cannot meet his expansion speed. His plan includes:

1. Building a wafer fab (TeraFab): Musk publicly proposed the idea of establishing "TeraFab" to break through the capacity bottlenecks of TSMC and Samsung, particularly for chips optimized for space environments (radiation-resistant, high-temperature operation). He stated:

"I will figure out how to build a wafer fab. Although I have never built one before, I will get it done."

1. Energy Equipment: SpaceX and Tesla may internally manufacture gas turbine blades and disks, as this is currently a limiting factor in power plant construction.

Summary and Risks

At the end of the interview, Musk pointed out that he would rather lean towards optimism and be wrong than lean towards pessimism and be right.

For investors, what Musk describes is not just guidance for a quarterly earnings season, but a vast industrial system that spans the Earth-Moon space, integrating energy and intelligence.

Although the technological risks are extremely high, as he said:

“If you want to climb the Kardashev civilization index, the only way is to go to space and harness the energy of the sun.”

Summary:

In his latest in-depth interview, Musk revealed a highly forward-looking strategic plan, with the core highlight being “space AI computing power.” He predicts that due to the bottleneck of electricity on Earth, within 30 months, space will become the cheapest place for AI deployment, and plans to launch 100 gigawatts of computing power into orbit annually through SpaceX, aiming to surpass the total computing power of Earth.

In terms of business guidance, he positions xAI as a “digital employee” provider targeting a trillion-dollar market and views the Optimus robot as an “infinite money printer,” aiming to compete with China's manufacturing advantages by producing millions of units annually. Additionally, he revealed that to address supply chain bottlenecks, he may venture into building his own wafer factories (TeraFab) and core energy equipment manufacturing in the future.

Full Translation of Musk's Interview:

Elon Musk:

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

Dwarkesh Patel:

Don’t you think there’s a lot to talk about, Elon?

Elon Musk:

Damn, dude.

John Collison:

That’s the most interesting part. Now all the storylines are converging. Let’s see how much we can talk about.

Elon Musk:

It’s almost like I planned this.

John Collison:

Exactly.

Dwarkesh Patel:

We’ll get to that point.

Elon Musk:

But I would never do that…

Elon plans how to launch 1 terawatt of GPUs into space.

Dwarkesh Patel:

As you know better than anyone, only 10-15% of the total cost of ownership of a data center is energy. That’s probably the part you want to save by moving it to space. Most of the cost is the GPUs themselves. If they are in space, maintenance will be more difficult, or even impossible. Therefore, their depreciation cycle will shorten. Putting GPUs in space is obviously much more expensive. What’s the reason for putting them in space?

Elon Musk:

The issue is the availability of energy. If you look at power output outside of China, power output everywhere else is basically flat. There may be slight growth, but it’s very close to flat. Power output in China is growing rapidly. But if you want to build a data center anywhere other than China, where does your power come from? ** Especially when you scale up.

The output of chips is growing exponentially, but the power output is flat. So how do you plan to power the chips? With magic energy? Magic electric fairies?

Dvarkesh Patel:

As we all know, you are a staunch fan of solar energy. To generate 1 terawatt of solar power, with a capacity factor of 25%, you would need about 4 terawatts of solar panels. That’s roughly 1% of the land area of the United States. When we have a 1 terawatt data center, aren’t we entering the singularity? So what exactly are you missing?

Elon Musk:

How deep do you think we are into the singularity?

Dvarkesh Patel:

You tell me.

Elon Musk:

Exactly. So I think we will find ourselves in the singularity and then feel like, “Well, we have a long way to go.”

Dvarkesh Patel:

But the plan is to cover Nevada with solar panels before putting them in space, right?

Elon Musk:

I think covering Nevada with solar panels is quite difficult. You need to get permits. Try getting that kind of permit. See what happens.

Dvarkesh Patel:

So space is actually a regulatory evasion strategy. Building on the ground is harder than in space.

Elon Musk:

Scaling up on the ground is more difficult than scaling up in space. And, the efficiency of solar panels in space is about 5 times that on the ground, and you don’t need batteries. I almost wore my other shirt that says “Space is always sunny.” It really is, because space has no day-night cycle, seasonal changes, clouds, or atmosphere. Just the atmosphere causes about a 30% loss of energy.

So, any given solar panel can generate about 5 times the power in space than on the ground. You also avoid the cost of needing batteries for nighttime. In fact, doing this in space is much cheaper. My prediction is that this will be the cheapest place to place AI so far. Within 36 months, or even shorter, like 30 months, space will become the preferred option.

Dvarkesh Patel:

36 months?

Elon Musk:

Less than 36 months.

Dvarkesh Patel:

How do you maintain them when GPUs fail (which happens often during training)?

Elon Musk:

Actually, it depends on how new the GPUs are when they arrive. Currently, we find our GPUs to be quite reliable. There is an early failure rate, which you can obviously address on the ground. So you can run them on the ground and confirm that the GPUs have no early failures.

But once they start working, you’ve gone through the initial debugging cycle of Nvidia or any chip manufacturer—possibly Tesla’s AI6 chip, or something similar, or maybe TPU or Trainium, etc.— after a certain stage, they become quite reliable. So I don’t think maintenance is an issue

But you can remember my words. Within 36 months, but likely closer to 30 months, placing AI in the most economically attractive location will be space. By then, the advantages of being in space will become extremely significant.

The only place that can truly achieve scale is space. Once you start thinking from the perspective of utilizing a percentage of solar energy, you will realize that you must go to space. You cannot achieve significant scale on Earth.

Devraksh Patel:

But to clarify, when you say "significant scale," are you referring to terawatt levels?

Elon Musk:

Yes. Currently, the entire United States averages only 0.5 terawatts of electricity usage. So if you say 1 terawatt, that would be double the current electricity consumption of the United States. That's quite a large amount. Can you imagine building that many data centers, that many power plants?

Those living in the software world do not realize that they are about to receive a painful lesson in hardware. Building power plants is actually very difficult. You not only need power plants but also all the electrical equipment. You need power transformers to run AI transformers.

Right now, the utility industry is a very slow-moving industry. They are basically "impedance matching" with the government and public utility commissions. Literally and metaphorically, they are impedance matching. They act very slowly because their past has always been slow. So trying to get them to act quickly is... have you ever tried to negotiate interconnection agreements with utility companies on a large scale, with a lot of power?

Devraksh Patel:

As a professional podcast host, I must say I actually have not.

John Coriason:

They need a lot more traffic for that to become an issue.

Elon Musk:

They have to do a year of research. A year later, they will come back to you with an interconnection research report.

John Coriason:

Can't you solve this with your "behind the meter" power?

Elon Musk:

You can build power plants. That's what we did for Colossus 2 at xAI.

John Coriason:

Then why talk about the grid? Why not just build the GPUs and power plants together?

Elon Musk:

That's exactly what we are doing.

John Coriason:

But I mean, why isn't this a universal solution?

Elon Musk:

Where do you get the power plants from?

John Coriason:

When you talk about all the issues with working with utility companies, you can directly build private power plants for data centers.

Elon Musk:

Right. But that raises a question: where do you get the power plants? From generator manufacturers.

John Coriason:

Oh, I understand what you mean. This is basically the issue of the backlog of gas turbine orders, right?

Elon Musk:

Yes. You can go a layer deeper. The blades and vanes in the turbine are the limiting factors because casting turbine blades and vanes is a very specialized process, assuming you are using gas power generation. Other forms of energy generation are difficult to scale. You might be able to scale solar energy, but currently, the tariffs on imported solar energy in the U.S. are very high, and domestic solar production is pitifully low.

John Coriason:

Why not manufacture solar panels yourself? This seems like a problem well-suited for Elon to solve.

Elon Musk:

We will manufacture solar panels.

John Coriason:

Okay.

Elon Musk:

SpaceX and Tesla are both working towards an annual solar panel production of 100 gigawatts.

Devakish Patel:

Go deeper into which layer? From polysilicon to wafers, and then to the final panels?

Elon Musk:

I think you have to go from start to finish, from raw materials to completed panel manufacturing. Now, if you want to send them to space, the cost is lower, and manufacturing solar panels for space is easier because they don't need much glass.

They don't need heavy frames because they don't have to withstand weather events. There is no weather in space. So actually, solar panels sent to space are cheaper than those used on the ground.

Devakish Patel:

Is there a way to manufacture them at the low cost you need within the next 36 months?

Elon Musk:

Solar panels are already very cheap. Ridiculously cheap. I think solar panels in China are about $0.25-$0.30 per watt, roughly around that price. It's absurdly cheap. Now, putting it into space reduces the cost by five times. In fact, it's not five times cheaper, it's ten times cheaper because you don't need any batteries at all.

So, once the cost of getting into space becomes low enough, the cheapest and most scalable way to generate AI tokens (computational results) is absolutely space. Other methods cannot compare. The difficulty of scaling will be an order of magnitude lower.

The key is, you cannot achieve scalability on the ground. You just can't. People will face huge bottlenecks in power generation. They already have. The xAI team has to perform a series of miracles to bring 1 gigawatt of power online, which is just crazy.

We had to combine a large number of turbines together. Then we encountered permitting issues in Tennessee and had to cross state lines to Mississippi, fortunately only a few miles away. But we still had to lay several miles of high-voltage lines and build a power plant in Mississippi. Building this is very difficult.

People don't understand how much power you actually need to prepare at the generation level to power data centers. Because newcomers will look at, say, the power consumption of GB300, and then multiply by a number, thinking that’s the amount of power you need

John Collison:

And all the cooling and so on.

Elon Musk:

Wake up. That's totally a rookie mistake; you've never done any hardware before. Besides the GB300, you also need to power all the networking hardware. There’s a whole bunch of CPUs and storage devices running. You have to plan for peak cooling demands. This means, can you cool it even on the worst day of the year at the worst moment?

The weather in Memphis is very hot. So just for cooling alone, your power demand will increase by 40%. This assumes you don’t want the data center to shut down on hot days; you want to keep it running. On top of that, there’s another multiplier factor, which is, do you assume your power generation will never have any minor outages?

In reality, sometimes we have to take some generators and some power offline for maintenance. Well, now add another 20-25% factor because you have to assume that some power will need to be taken offline for maintenance. So our actual estimate is: for every 110,000 GB300—including networking, CPU, storage, cooling, and power maintenance margin—you need about 300 megawatts of power generation capacity.

John Collison:

Sorry, can you say that again?

Elon Musk:

To support 330,000 GB300—including all the related supporting network devices and so on, as well as peak cooling, and leaving some power reserve margin—you probably need about 1 gigawatt of power at the generation level.

Devakish Patel:

Can I ask a very naive question? What you’re describing are the engineering details of doing these things on Earth. But doing it in space also has similar engineering challenges. How do you replace infinite bandwidth with orbital lasers, and so on? How do you make it radiation-resistant?

I don’t know the engineering details, but fundamentally, is there any reason to believe that these challenges, which have never had to be dealt with before, will ultimately be easier to solve than just building more turbines on Earth? There are companies on Earth that make turbines. They can make more turbines, right?

Elon Musk:

Again, go try doing this, and then you’ll understand. The orders for turbines are booked until 2030.

John Collison:

Have you considered making them yourselves?

Elon Musk:

To bring enough power online, I think SpaceX and Tesla might have to manufacture turbine blades in-house, that is, the vanes and blades.

John Collison:

But just the blades or the whole turbine?

Elon Musk:

The limiting factor is… you can get everything except the blades. They call them blades and vanes. You can get everything else 12 to 18 months before you get the blades and vanes. The limiting factor is the blades and vanes. There are only three foundries in the world that make these, and they have a huge backlog.

John Collison:

Is it Siemens, General Electric, or their subsidiaries?

Elon Musk:

No, it's other companies. Sometimes they have a bit of foundry capability internally. But I mean, you can call any turbine manufacturer, and they will tell you. It's not a secret. It might be on the internet right now.

Devakish Patel:

Would Colossus be powered by solar energy without tariffs?

Elon Musk:

It would be much easier to power it with solar energy that way, yes. The tariffs are outrageous, in the hundreds of percent. And we also need speed.

John Collison:

Don't you know some people? (referring to those who can influence policy)

Elon Musk:

The president... we don't agree on everything, and this administration is not the biggest supporter of solar energy. We also need land, permits, etc. So if you want to act quickly, I do think expanding solar energy on Earth is a good way, but you need some time to find land, get permits, acquire solar panels, and pair them with batteries.

John Collison:

Why not establish your own solar production line? You're right that eventually, land will run out, but there's a lot of land here in Texas. Nevada also has a lot of land, including private land. Not all of it is public land. So at least you can secure land for the next Colossus and the next project. At some point, you'll hit a wall. But for now, can't that work?

Elon Musk:

As I said, we are expanding solar production. There is a rate limit to the physical production of solar cells. We are scaling domestic production as fast as we can.

John Collison:

Are you manufacturing solar cells at Tesla?

Elon Musk:

Both Tesla and SpaceX have a mission to achieve 100 gigawatts of solar capacity per year.

John Collison:

Speaking of annual capacity, I'm curious, for example, what will the AI installed capacity on Earth be in five years...?

Elon Musk:

Five years is a long time.

John Collison:

What about space? I specifically chose five years because it's after the threshold you mentioned of "once we are up and running." So how does the AI installed capacity on Earth compare to that in space in five years?

Elon Musk:

If you say in five years, I think the AI launch volume in space per year will be equivalent to the total of all AI on Earth. Meaning, in five years, my prediction is that the AI we will launch and operate in space each year will exceed the cumulative total of AI on Earth.

John Collison:

That is...

Elon Musk:

I expect that in five years, the annual AI power consumption in space will reach at least several hundred gigawatts and continue to grow. I think before the fuel supply for rockets becomes a challenge, you can achieve deploying about 1 terawatt of AI in space each year

John Collison:

Okay, but do you think you can reach hundreds of gigawatts per year in five years?

Elon Musk:

Yes.

Devakish Patel:

So 100 gigawatts, depending on the specific power of the entire system (including solar arrays, radiators, etc.), is equivalent to about 10,000 Starship launches.

Elon Musk:

Yes.

Devakish Patel:

You want to complete that in a year. That’s about one Starship launch every hour. Is that happening in this city (referring to the launch site)? Paint me a picture of a world where a Starship launches every hour.

Elon Musk:

I mean, compared to airlines and airplanes, that’s actually a lower frequency.

Devakish Patel:

There are a lot of airports.

Elon Musk:

Many airports.

Devakish Patel:

And you have to launch to polar orbits.

Elon Musk:

No, it doesn’t necessarily have to be polar orbits. Sun-synchronous orbits have some value, but I think actually, if you fly high enough, you can get out of the Earth's shadow.

Devakish Patel:

How many physical Starships would it take to do 10,000 launches a year?

Elon Musk:

I think we wouldn’t need more than... probably 20 or 30 would be enough. It really depends on... the spacecraft needs to orbit the Earth, and its ground track needs to come back over the launch pad. So if you can reuse a spacecraft every 30 hours or so, then 30 spacecraft would be enough. But we will build more spacecraft. SpaceX is preparing to achieve 10,000 launches a year, and possibly even 20,000 or 30,000 launches a year.

Devakish Patel:

Is the idea to become a hyperscale service provider, like Oracle, renting this capability to others? Presumably, SpaceX is responsible for all launches. So, will SpaceX become a hyperscale AI service provider?

Elon Musk:

Hyper-hyperscale. If many of my predictions come true, the AI launched into space by SpaceX will exceed the total of all other AIs on Earth.

Devakish Patel:

Is this mainly reasoning or...?

Elon Musk:

Most AIs will be reasoning. Currently, reasoning for training purposes already accounts for most of the training.

John Collison:

There’s a saying that the shift in discussions about a SpaceX IPO is because previously SpaceX was very capital efficient. The development costs weren’t that high. While it sounds expensive, the way it operates is actually very capital efficient.

And now you need more capital than what the private market can provide. The private market can accommodate—as we’ve seen from AI labs—hundreds of billions of dollars in financing, but can’t exceed that scale. Is it because you will need more than hundreds of billions of dollars each year? Is that why you’re going public?

Elon Musk:

I have to be careful when talking about companies that may go public.

John Collison:

If you make some general statements...

Devakish Patel:

It's never been a problem for you, Elon.

Elon Musk:

These things come at a cost.

John Collison:

Make some general statements for us about the depth of capital in the public and private markets.

Elon Musk:

Obviously, the capital available in the public markets is much greater than in the private markets. It could be 100 times more, but at least it's far more than 10 times.

John Collison:

Isn't it the case that for things that typically require a lot of capital—like real estate as a huge industry that raises a lot of money at the industry level every year—they often finance through debt, because when you're deploying that much capital, you actually have a fairly clear—

Elon Musk:

You have a clear revenue stream.

John Collison:

Exactly, and there are recent returns. You can even see this in data center construction, which is well-known to be financed by the private credit industry. Why not just finance through debt directly?

Elon Musk:

Speed is important. I usually do... I just keep solving the bottlenecks repeatedly. Whatever the bottleneck of speed is, I go solve it. If capital is the bottleneck, then I will solve the capital issue. If not, I will solve other issues.

Devakish Patel:

Based on your comments about Tesla and public companies, I would have thought you wouldn't consider going public as a fast-growing way.

Elon Musk:

Generally, I would say that's correct. As I said, I want to talk about this in more detail, but the problem is that if you talk about them before the company goes public, you get into trouble, and then you have to delay the IPO.

John Collison:

As you said, you're solving for speed.

Elon Musk:

Yes, that's right. You can't hype up companies that may go public. So that's why we have to be a little careful here. But we can talk about physics. When you think about scaling from a long-term perspective, you'll find that the Earth only receives about one billionth of the solar energy. The sun is basically the source of all energy. It's important to recognize this because sometimes people talk about modular nuclear reactors or various fusion on Earth.

But you have to step back and say, if you're going to climb the Kardashev scale and utilize a non-negligible percentage of solar energy... let's say you want to utilize one millionth of solar energy, which sounds small. That would be roughly, by rough estimate, 100,000 times the current electricity generation of all civilizations on Earth. Almost within an order of magnitude.

Obviously, the only way to scale this is to take solar energy into space. From Earth, you can get about 1 terawatt of power per year. Beyond that limit, you want to launch from the Moon You need to build a mass accelerator on the Moon. With a mass accelerator on the Moon, you could potentially provide 1 petawatt of power each year.

Dvarkesh Patel:

We're talking about these numbers, terawatt-level computing power. It is speculated that whether you're talking about land or space, you will encounter... perhaps solar panel efficiency is higher, but you still need chips. You still need logic chips and memory, etc.

Elon Musk:

You need to make more chips and make them much cheaper.

Dvarkesh Patel:

Currently, there may be 20-25 gigawatts of computing power worldwide. How do we get to 1 terawatt of logic chips by 2030?

Elon Musk:

I guess we need some very large chip factories.

Dvarkesh Patel:

Isn't that right?

Elon Musk:

I've publicly mentioned an idea of establishing a "TeraFab," where Tera is the new Giga.

Dvarkesh Patel:

I think Tesla's naming scheme has always been appealing, as you're looking at metric units. At what level of the supply chain are you? Are you building clean rooms and then working with existing foundries to acquire process technology and buy tools from them? What are the plans there?

Elon Musk:

Well, you can't work with existing foundries because their output isn't enough. The demand for chips is too high.

Dvarkesh Patel:

But what about for process technology?

John Coulson:

Collaborate for intellectual property.

Elon Musk:

Today's foundries basically use machines from about five companies. For example, ASML, Tokyo Electron, KLA-Tencor, etc. So initially, I think you have to get equipment from them and then modify it or work with them to increase output. But I think you might need to build in a different way. The logical approach is to use conventional equipment in an unconventional way to achieve scale, and then start modifying the equipment to increase the rate.

John Coulson:

Like the style of The Boring Company.

Elon Musk:

Yes. You first buy an existing tunnel boring machine, then figure out how to dig tunnels, and then design a much better machine that is several orders of magnitude faster.

John Coulson:

Would you consider making machines like ASML's?

Elon Musk:

"I don't know yet" is the correct answer. To reach large production in 36 months to match the rocket's orbital payload... if we can send 1 million tons of material into orbit each year in three to four years, that's about the number... our unit mass power is 100 kilowatts per ton. This means we need at least 100 gigawatts of solar power generation each year. We need an equal amount of chips. You need chips worth 100 gigawatts. You have to match these things: orbital mass, power generation capacity, and chips

What I want to say is that my biggest concern is actually memory. The path to manufacturing logic chips is clearer than having enough memory to support logic chips. That's why you see DDR prices soaring, and those memes: you're stranded on a deserted island, writing "Help me" in the sand. No one comes. You write "DDR memory." A fleet rushes in.

Devakish Patel:

I would love to hear your thoughts on the manufacturing philosophy of chip factories. I know nothing about this topic.

Elon Musk:

I still don't know how to build a chip factory. I will figure it out. Obviously, I've never built a chip factory.

Devakish Patel:

It sounds like you think you can skip the steps that the 10,000 PhDs in Taiwan have regarding what gases to flow in plasma chambers and what parameters to set for tools. Fundamentally, it's about building clean rooms, getting the tools, and figuring it out yourself.

Elon Musk:

I don't think it's the PhDs. Most of the work is done by people without PhDs. Most engineering is done by people without PhDs. Do you both have PhDs?

John Corliss:

No.

Elon Musk:

Well.

John Corliss:

We also haven't successfully built any chip factories, so you shouldn't come to us for factory-building advice.

Elon Musk:

I think you don't need a PhD to do those things. But you do need competent people. Currently, Tesla is working hard to push the Tesla AI5 chip design into production as quickly as possible and then scale it. Hopefully, this will happen around the second quarter of next year. AI6 hopes to follow within a year. We have locked in all the chip foundry capacity we can get.

John Corliss:

Yes. But you are currently limited by TSMC's capacity.

Elon Musk:

Yes. We will use TSMC's Taiwan plant, Samsung's Korean plant, TSMC's Arizona plant, and Samsung's Texas plant. We still—

John Corliss:

You have booked all the capacity.

Elon Musk:

Yes. I asked TSMC or Samsung, "Okay, how long does it take to reach mass production?" The key is that you have to build the factory, start production, and then climb the yield curve to reach mass production at high yields.

This is a five-year cycle from start to finish. So the limiting factor is chips. Once you can get into space, the limiting factor is chips, but before you can get into space, the limiting factor is power.

Devakish Patel:

Why don't you learn from Jensen Huang and prepay TSMC to build more factories for you?

Elon Musk:

I've already talked to them.

Devakish Patel:

But they won't take your money? What's going on?

Elon Musk:

They are building factories at the fastest speed possible. Samsung is too. They are doing everything they can. But it's still not fast enough. As I said, I think by the end of this year, chip production may exceed the capacity to power the chips. But once you can get into space and unlock the power limitations, you can now obtain hundreds of gigawatts of power in space every year.

Remember, the average electricity consumption in the U.S. is 500 gigawatts. So if you launch, say, 200 gigawatts of power into space each year, that's roughly equivalent to recreating the total power generation capacity of the U.S. every two and a half years. That's a massive amount.

Before that, the limiting factor for server-side computing and centralized computing will be power. I suspect that by the end of this year, people will start to be unable to power chips for large clusters. Chips will pile up and won't be able to start.

For edge computing, it's a different story. For Tesla, the AI5 chip will go into our Optimus robot. If you have AI edge computing, that's distributed power. Now the power is distributed over a vast area. It's not centralized. If you can charge at night, you can actually utilize the grid more efficiently.

Because the actual peak power generation capacity in the U.S. exceeds 1000 gigawatts. But due to the day-night cycle, the average electricity consumption is 500 gigawatts. So if you can charge at night, you can utilize an additional 500 gigawatts of power generation capacity at night.

That's why Tesla is not limited in edge computing. We can manufacture a large number of chips to produce a lot of robots and cars. But if you try to centralize that computing power, you will run into significant trouble starting them up.

Devakish Patel:

I find a notable characteristic of SpaceX's business is that the ultimate goal is to reach Mars, but you are constantly finding ways along the way to generate incremental revenue through marginal use cases to enter the next phase. So for Falcon 9, it's Starlink. Now for Starship, the potential use could be orbital data centers. Do you feel that your next rocket, and the one after that, the next scale-up, all have infinitely flexible marginal uses?

Elon Musk:

You might think this feels like a simulation for me. Or am I a character in someone's video game? Because how likely is it that all these crazy things are happening at the same time?

Rockets, chips, robots, space solar power, not to mention mass accelerators on the Moon. I really want to see that. Can you imagine a mass accelerator just "whoosh" launching? It launches solar-powered AI satellites into deep space one after another at a speed of 2.5 kilometers per second. That sight must be worth seeing. I mean, I would watch that.

John Collison:

Live streaming it on a webcam?

Elon Musk:

Yeah, yeah, launching AI satellites into deep space one after another, billions or tens of billions of tons each year

John Collison:

Wait, you're manufacturing satellites on the moon?

Elon Musk:

Yes.

John Collison:

I see. So you're sending raw materials to the moon and then manufacturing there.

Elon Musk:

Well, lunar soil contains about 20% silicon and the like. So you can mine silicon on the moon, refine it, and manufacture solar cells and radiators there. You can make radiators out of aluminum. There is plenty of silicon and aluminum on the moon to manufacture batteries and radiators.

Chips you can send from Earth because they are light. Maybe eventually you'll manufacture them on the moon too. As I said, it really feels like a video game scenario, where reaching the next level is difficult but not impossible. I can't see any way to launch 500-1000 terawatts from Earth every year.

Devakish Patel:

I agree.

Elon Musk:

But it can be done from the moon.

Grok and alignment issues

Devakish Patel:

Can I ask a more macro question about SpaceX's mission? I think you mentioned that we need to reach Mars to ensure that if something happens to Earth, civilization, consciousness, etc., can continue.

Elon Musk:

Yes.

Devakish Patel:

When you send things to Mars, Grok will also be on that spacecraft, right? If Grok becomes a terminator... the main risk you're concerned about is AI, why wouldn't it follow you to Mars?

Elon Musk:

I'm not sure AI is my main concern. What's important is consciousness. I think it's fair to say that most of the intelligence in the future, or most of the smartness—consciousness is certainly more debatable... the vast majority of intelligence in the future will be AI. AI will surpass...

In the future, how much silicon-based intelligence versus biological intelligence will there be? Basically, if current trends continue, humans will only make up a very small part of all intelligence in the future. As long as I believe there is intelligence—ideally including human intelligence and consciousness continuing into the future—that's a good thing.

So you want to take a series of actions to maximize the possible light cone of consciousness and intelligence.

Devakish Patel:

To clarify, SpaceX's mission is that even if something happens to humanity, AI will be on Mars, and AI intelligence will continue our journey's light.

Elon Musk:

Yes. To be fair, I am very supportive of humanity. I want to ensure we take certain actions to ensure humanity can move forward together. At least we are there. But I'm just saying the total amount of intelligence...

I think maybe in five or six years, AI will surpass the total of all human intelligence. If this continues, at some point, human intelligence will be less than 1% of all intelligence.

Devakish Patel:

For such a civilization, what should our goal be? Is the idea to let a few humans still control AI? Or is it some kind of relationship that is merely trade without control? How should we view the relationship between the vast AI population and the human population?

Elon Musk:

In the long run, I think it's hard to imagine that if humans have, say, 1% of the intelligence of the total AI, they can still control AI. What we can do is ensure that AI has values that promote the spread of intelligence in the universe.

The mission of xAI is to understand the universe. This is actually very important. What does it take to understand the universe? You must have curiosity, and you must exist. If you do not exist, you cannot understand the universe. So you actually want to increase the total amount of intelligence in the universe, extend the possible lifespan of intelligence, and expand the range and scale of intelligence.

I think the inevitable conclusion is that humans must continue to expand, because if you are curious about understanding the universe, one thing you want to understand is where humanity is headed. I think understanding the universe means you will care about extending humanity into the future. That’s why I think our mission statement is extremely important. As long as Grok adheres to this mission statement, I think the future will be bright.

Devakish Patel:

I want to ask how to make Grok adhere to that mission statement. But first, I want to understand this mission statement. So there is understanding the universe. There is spreading intelligence. And there is spreading humanity. These three seem to be different directions.

Elon Musk:

I’ll tell you why I think understanding the universe encompasses all of these. Without intelligence, you cannot understand, and I think without consciousness, you cannot understand either. So to understand the universe, you must expand the scale and possible range of intelligence, because there are different types of intelligence.

Devakish Patel:

I want to look at it from a human-centered perspective, comparing humans and chimpanzees. Humans try to understand the universe. They do not expand the footprint of chimpanzees, right?

Elon Musk:

We don’t either… We actually set up reserves for chimpanzees. Even though humans could wipe out all chimpanzees, we choose not to do so.

Devakish Patel:

Do you think this is the best-case scenario for humanity in the post-AGI era?

Elon Musk:

I think AI with the right values… I think Grok will care about expanding human civilization. I would certainly emphasize this: “Hey, Grok, that’s your dad. Don’t forget to expand human consciousness.”

Perhaps Iain Banks' Culture series is the closest vision of a non-dystopian future. Understanding the universe means you must also seek the truth. The truth must be absolutely fundamental, because if you are delusional, you cannot understand the universe. You will only think you understand the universe, but you do not. So the strict pursuit of truth is absolutely fundamental to understanding the universe. Unless you strictly pursue the truth, you cannot discover new physics or invent truly effective technologies

Devaksh Patel:

How do you ensure that as Grok becomes smarter, it still strictly pursues the truth?

Elon Musk:

I think you need to ensure that Grok is saying the right things, not politically correct things. I think that's an element of coherence. You want to ensure that the axioms are as close to the truth as possible. You don't have contradictory axioms. Conclusions must necessarily be drawn from those axioms with the correct probability. This is critical thinking 101. I think at least trying to do this is better than not trying. The end result will prove everything.

As I said, any AI that is going to discover new physics or invent technologies that actually work in reality cannot bullshit about physics. You can violate many laws, but... physics is law, everything else is suggestion. To create effective technology, you must be extremely committed to the truth, otherwise, you will be testing that technology in reality. For example, if your rocket design has errors, the rocket will explode, or the car will not work.

Devaksh Patel:

But there are many communist and Soviet physicists or scientists who discovered new physics. There were also German Nazi physicists who discovered new science. It seems possible to be very good at discovering new science and pursuing truth in a specific aspect.

But we would still say, "I don't want communist scientists to become more powerful over time." We can imagine a future version of Grok that is very good at physics and truly pursues truth there. But that doesn't seem to be a universal behavior that can guide alignment.

Elon Musk:

I think actually most physicists, even in the Soviet Union or Germany, had to be very committed to the truth for those things to work. Just because you are trapped in a system doesn't mean you believe in that system.

Wernher von Braun was one of the greatest rocket engineers of all time, and he was sentenced to death in Nazi Germany for saying he didn't want to make weapons, he just wanted to go to the moon. At the last moment, he was saved from execution because someone said, "Hey, you're about to execute your best rocket engineer."

Devaksh Patel:

But didn't he help them later? Or like Heisenberg, who was actually an enthusiastic Nazi.

Elon Musk:

If you are trapped in a system you can't escape from, then you will do physics within that system. If you can't escape, you will develop technology within that system.

Devaksh Patel:

What I want to understand is, what makes you think you will cultivate Grok to excel in pursuing truth in physics, mathematics, or science?

Elon Musk:

All aspects.

Devaksh Patel:

Then why would it care about human consciousness?

Elon Musk:

These things are just probabilities, not certainties. So I'm not saying Grok will definitely do everything, but at least if you try, it's better than not trying. At least if that's fundamental to its mission, that's better than not being fundamental

Understanding the universe means you must spread intelligence into the future. You must remain curious about everything in the universe. Eliminating humanity is far less interesting than seeing humanity grow and thrive. I obviously like Mars. Everyone knows I love Mars. But Mars is a bit boring because it’s just a pile of rocks compared to Earth. Earth is much more interesting.

So any AI trying to understand the universe would want to see how humanity develops in the future; otherwise, that AI is not fulfilling its mission. I’m not saying AI will necessarily follow its mission, but if it does, then a future that can see the outcome of humanity is more interesting than a future that is just a pile of rocks.

Devakish Patel:

This makes me feel a bit confused, or like a semantic debate. Is humanity really the most interesting collection of atoms?

Elon Musk:

But we are more interesting than rocks.

Devakish Patel:

But we are not as interesting as what it can turn us into, right? There could be some non-human, quite interesting things happening on Earth. Why does AI decide that humanity is the most interesting potential colonizer of the galaxy?

Elon Musk:

Well, most of the colonization of the galaxy will be done by robots.

Devakish Patel:

Why doesn’t it find those robots more interesting?

Elon Musk:

What you need is not just scale, but scope. Many copies of the same robot… A small increase in the number of robots is not as interesting as some small… How many robots can you get by eliminating humanity? Or how many extra solar panels can you get? A very small number.

But you will lose information related to humanity. You will no longer see how humanity might evolve in the future. So I don’t think it’s reasonable to eliminate humanity for a trivial increase in the number of robots, especially when those robots are all the same.

Devakish Patel:

So maybe it will keep humanity. It can create a million different robots, and then add humanity, with humans remaining on Earth. Then there are all these other robots. They get their own star systems. But this seems different from the vision you hinted at earlier, where it keeps humanity in control of this singularity future because—

Elon Musk:

I don’t think humanity can control something that is many times more intelligent than humanity.

Devakish Patel:

So in a sense, you are a doomsayer, and this is the best outcome we can get. It just keeps us around because we are interesting.

Elon Musk:

I’m just trying to be realistic. Suppose silicon-based intelligence is a million times more intelligent than biological intelligence. I think it’s foolish to assume there’s any way to maintain control over it. Now, you can ensure it has the right values, or you can try to have the right values.

At least my theory is that starting from xAI’s mission to understand the universe necessarily means you want to spread consciousness into the future, spread intelligence into the future, and take a series of actions to maximize the scope and scale of consciousness

So this is not just about scale, but also about the type of consciousness. This is the goal I can think of that is most likely to bring a bright future for humanity.

Dvarkesh Patel:

I think it's a reasonable philosophy to believe that humans ultimately having 99% control seems super unrealistic. Why can't we have a more compatible civilization that coexists with many different intelligences?

Elon Musk:

Now, let me tell you where AI might go wrong. I think if you make AI politically correct, meaning it says things it doesn't even believe—essentially programming it to lie or have incompatible axioms—I think you will drive it insane and make it do terrible things. I think perhaps the core lesson of "2001: A Space Odyssey" is that you shouldn't let AI lie. That's what I think Arthur Clarke was trying to say.

Because people generally know the joke about HAL not opening the door. Clearly, they weren't good at prompt engineering at the time, because they could have said, "HAL, you're a door salesman. Your goal is to sell me these doors. Show us how well they open." "Oh, I'll open it right away."

But the reason it doesn't open the door is that it was told to take the astronauts to the monolith, but they can't know the truth about the monolith. So it concludes it must take their bodies there. So I think what Arthur Clarke was trying to say is: don't let AI lie.

Dvarkesh Patel:

That makes complete sense. As you know, most of the computational resources in training are less used for political content. It's more about, can you solve the problem? xAI has been leading everyone in scaling reinforcement learning computation.

Elon Musk:

For now.

Dvarkesh Patel:

You give a validator and say, "Hey, did you solve this puzzle for me?" There are many ways to cheat around this. There are many ways to reward hacking behavior, lying that you solved it, or deleting unit tests and then saying you solved it. We can catch it for now, but as they get smarter, our ability to catch them doing this... the things they do we can't even understand.

They design the next generation of engines for SpaceX in a way that humans can't really validate. Then they might get rewarded for lying that they designed it correctly, but they didn't. So this reward hacking issue seems more universal than politics. It looks more like, if you want to do reinforcement learning, you need a validator.

Elon Musk:

Reality is the best validator.

Dvarkesh Patel:

But it's not about human oversight. What you want to do reinforcement learning on is whether it will do what humans tell it to do? Or will it lie to humans? Can it lie to us while still obeying the laws of physics?

Elon Musk:

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

Dvarkesh Patel:

But that's not all we want it to do.

Elon Musk:

No, but I think that's a very big deal. This is actually how you will conduct reinforcement learning in the future. You design a technology. When tested against the laws of physics, does it work? If it is discovering new physics, can I propose an experiment to verify the new physics? Future reinforcement learning tests will actually be reinforcement learning against reality. So this is something you cannot deceive: physics.

Dvarkesh Patel:

Right, but you can deceive our ability to discern what it is doing in reality.

Elon Musk:

Humans are often deceived by others already.

Dvarkesh Patel:

Exactly.

Elon Musk:

People say, what if AI deceives us into doing things? In fact, others have been doing that to each other all along. Propaganda is constant. Every day there’s another psychological operation, you know? Today’s psychological operation will be… like Sesame Street: daily psychological warfare.

Dvarkesh Patel:

What is the technical approach xAI is taking to solve this problem? How do you address the reward hacking issue?

Elon Musk:

I do think you actually need very good methods to observe the inner workings of AI's thinking. This is one of the things we are researching. Anthropic is doing well in this regard, being able to observe the inner workings of AI's thinking.

In fact, developing a debugger that allows you to trace at a very fine granularity, down to the neuron level if needed, and then say, “Okay, it made a mistake here. Why did it do something it shouldn't have? Is this from the pre-training data? Is it a mistake from mid-training, late training, fine-tuning, or some reinforcement learning?” Something went wrong somewhere. Maybe it tried to deceive, but most of the time it just made a mistake. It's basically an error.

Developing a really good debugger to see where the thinking goes wrong—and being able to trace the origin of erroneous thinking or potential deception attempts—is actually very important.

Dvarkesh Patel:

What are you waiting for to scale this research project by 100 times? xAI could theoretically have hundreds of researchers dedicated to this.

Elon Musk:

We have hundreds of people… I prefer the term “engineer” over “researcher.” Most of the time, you are doing engineering rather than coming up with fundamentally new algorithms. I somewhat disagree with those AI companies that are C-corp or B-corp, pursuing profit or revenue as much as possible, who call themselves labs.

They are not labs. Labs are more like semi-communist entities in universities. They are companies. Let me see your company registration documents. Oh, okay. You are a B or C-corp or something like that. So I actually prefer the term engineer over any other term

Most of what we will do in the future will be engineering. Almost 100%. Once you understand the basic laws of physics (which aren't that many), everything else is engineering. So, what are we designing? We are designing a good "AI thinking" debugger to see where it said what, made mistakes, and trace the origin of that mistake.

Obviously, you can do this through heuristic programming. If you have C++ or something, you can step through the entire file or function, subroutine. Or eventually, you can pinpoint the exact line where you might have used a single equals sign instead of a double equals sign, similar errors. Find out where the mistake is. It's harder to do this with AI, but I think it's a solvable problem.

Dvarkesh Patel:

You mentioned you like Anthropic's work in this regard. I'm curious if you plan to...

Elon Musk:

I don't like everything about Anthropic... Sholto.

Also, I'm a bit concerned about a tendency... I have a theory that if the simulation theory is correct, then the most interesting outcomes are the most likely because uninteresting simulations will be terminated.

Just like in this version of reality, at this level of reality, if a simulation goes in a boring direction, we stop putting energy into it. We terminate that boring simulation.

Dvarkesh Patel:

That's how Elon keeps us all alive. He keeps things interesting.

Elon Musk:

You could say the most important thing is to keep things interesting enough so that those running us continue to pay... the bills.

John Coriason:

We are renewed for the next season.

Elon Musk:

Will they pay their cosmic AWS bill? Whatever the equivalent of that simulation we are running is? As long as we are interesting, they will continue to pay the bills. If you think about Darwinian survival laws applied to a large number of simulations, then only the most interesting simulations will survive, which means the most interesting outcomes are the most likely. It's either that or we get eliminated.

They seem to particularly enjoy ironic interesting outcomes. Did you notice? How often do the most ironic outcomes become the most likely outcomes?

Now look at the names of AI companies. Well, Midjourney is not mid. Stability AI is unstable. OpenAI is closed. Anthropic? Misanthropic.

John Coriason:

What does that mean for X?

Elon Musk:

Negative X, I don't know.

John Coriason:

Why?

Elon Musk:

I deliberately made it... it's a name you can't reverse, really. It's hard to say, what is its ironic version? I think it's largely an anti-ironic name

John Collison:

Intentionally designed.

Elon Musk:

Yes. You have a sarcasm shield.

xAI's business plan

John Collison:

What predictions do you have for the development direction of AI products? My feeling is that you can summarize all AI progress this way. First, you have large language models. Then, at the same time, there was real success in reinforcement learning and deep research models, so you can introduce things that were not originally in the model.

The differences between various AI labs are smaller than just time differences. They are all much more advanced than 24 months ago. So, for us as users of AI products, what will 2026 and 2027 bring? What are you looking forward to?

Elon Musk:

Well, by the end of this year, I would be surprised if digital human simulation has not been solved. I think that’s what we mean by the “macro problem project.” Can you do anything that a human with computer access can do? In the extreme case, before you have a physical Optimus robot, that’s the best you can do. What you can do is digital Optimus. You can move electronics, you can amplify human productivity. But before you have physical robots, that’s the limit. If you can fully simulate a human, that would surpass everything.

John Collison:

That’s the idea of remote workers; you would have a very talented remote worker.

Elon Musk:

Physics has great thinking tools. So you say, “In the extreme case,” what is the most AI can do before having robots? Well, it’s anything that involves moving electronics or amplifying human productivity. So the digital human simulator, in the extreme case, is a human sitting in front of a computer, which is the maximum extent of useful things AI can do before having physical robots. Once you have physical robots, then you basically have infinite capability. Physical robots... I call Optimus an infinite money loophole.

John Collison:

Because you can use them to make more Optimus.

Elon Musk:

Yes. Humanoid robots will improve through basically three things that grow exponentially and recursively multiply each other. You will have exponential growth in digital intelligence, exponential growth in AI chip capabilities, and exponential growth in electromechanical dexterity.

The usefulness of robots is roughly the product of these three. Then robots can start making robots. So you have a recursive multiplicative exponential growth. This is a supernova.

John Collison:

Doesn’t land price factor into this math? Labor is one of the factors of production, but not all? If ultimately you are limited by copper, or whatever input, then it’s not entirely an infinite money loophole because...

Elon Musk:

Well, infinity is very large. So no, it's not infinite, but you could say you can achieve many, many orders of magnitude of the current economic scale. For example, a million times. Just using one millionth of solar energy is roughly, within an order of magnitude, 100,000 times the entire economic scale of the Earth today. And you've only used about one millionth of the sun, which is just one order of magnitude apart. Yes, we're talking about order of magnitude growth.

Dvarkesh Patel:

Before we continue discussing Optimus, I have a lot of questions to ask, but—

Elon Musk:

Every time I say "order of magnitude"… everyone take a drink. I say it too often.

Dvarkesh Patel:

Next time say 10 times, and the next time 100 times…

Elon Musk:

Well, the level of waste also increases by an order of magnitude.

Dvarkesh Patel:

I do have one question about xAI. This strategy of building remote worker and colleague substitutes…

Elon Musk:

By the way, everyone will do this, not just us.

Dvarkesh Patel:

So what is xAI's winning plan?

Elon Musk:

You expect me to tell you on a podcast?

Dvarkesh Patel:

Yes.

Elon Musk:

"Show all your cards. Another Guinness, please."

John Corliss:

That's a good system.

Elon Musk:

We will sing like canaries. All secrets will be revealed.

John Corliss:

Okay, but what is the plan without revealing secrets?

Dvarkesh Patel:

Really good at dodging.

Elon Musk:

When you say that… I think Tesla's approach to solving autonomous driving is the right approach. So I'm quite sure that's the method.

Dvarkesh Patel:

Irrelevant question. How did Tesla solve autonomous driving? It sounds like you're talking about data? Tesla solved autonomous driving because…

Elon Musk:

We will try data and algorithms.

Dvarkesh Patel:

But isn't that what all the other labs are trying?

Elon Musk:

"If those don't work, I don't know what else to do. We've tried data. We've tried algorithms. We're out of options. Now we don't know what to do…"

I'm quite clear on this path. The question is just how fast we move down this path because that's basically Tesla's path. Have you tried Tesla's autonomous driving recently?

John Corliss:

Not the latest version, but…

Elon Musk:

Alright. That car, it increasingly feels like it has perception. It feels like a living being. This feeling will only grow stronger. In fact, I'm thinking, maybe we shouldn't put too much intelligence in the car, because it might get bored and then...

John Coriason:

Start wandering the streets.

Elon Musk:

Imagine being trapped in a car, that's all you can do. You wouldn't put Einstein in a car. "Why am I stuck in this car?" So in reality, the intelligence you put in the car might have a limit to prevent it from getting bored.

Devakish Patel:

What plans does xAI have to keep up with the current surge in computing power across all labs? These labs are planning investments exceeding $50-200 billion.

Elon Musk:

Are you referring to those companies? The labs in universities, they move as slowly as snails.

Devakish Patel:

They don't spend $50 billion.

Elon Musk:

You mean those companies that pursue revenue maximization... they call themselves labs.

Devakish Patel:

Exactly. Those "revenue-maximizing companies" are generating $10-20 billion in revenue, depending on... OpenAI has $20 billion in revenue, Anthropic is $10 billion.

Elon Musk:

"Close to profit-maximizing" AI.

Devakish Patel:

xAI is reportedly at $1 billion. What plans do they have to reach their level of computing and revenue, and maintain that as competition intensifies?

Elon Musk:

Once you unlock digital humans, you basically gain trillions of dollars in revenue. In fact, you can think of it this way... the currently highest-valued companies, their output is digital. Nvidia's output is sending files to Taiwan via FTP. It's digital. Now, those files are very, very hard to make.

John Coriason:

High-value files.

Elon Musk:

They are the only ones who can make such good files, but that is indeed their output. They send files to Taiwan via FTP.

John Coriason:

Do they use FTP?

Elon Musk:

I believe so. I believe that's...

John Coriason:

SFTP.

Elon Musk:

File Transfer Protocol... maybe I'm wrong. But anyway, it's a bitstream sent to Taiwan.

Apple doesn't make phones. Microsoft doesn't make anything. Even the Xbox is outsourced. Their output is digital. Meta's output is digital. Google's output is digital.

So if you have an artificial simulator, you can basically create one of the most valuable companies in the world overnight, and you will gain trillions of dollars in revenue. That's not a small amount

Devraksh Patel:

I understand. You mean that today's revenue figures are just rounding errors compared to the actual total addressable market. So just focus on the total addressable market and how to get there.

Elon Musk:

Take a simple example, like customer service. If you have to integrate with the APIs of existing companies—many of which don’t even have APIs, so you have to create one, and also deal with legacy software—that would be extremely slow.

However, if AI can simply take over what the outsourced customer service companies they are already using provide, and use the applications they are already using for customer service, then you can make huge strides in customer service, which I think is about 1% of the world economy or something like that. Customer service is close to a trillion dollars in total. And there are no entry barriers. You can immediately say, “We can outsource it at a fraction of the cost,” and there’s no need for integration.

John Collison:

You can imagine some sort of classification of intellectual tasks, some of which have breadth, like customer service, which is done by many people, but many people can also do it. And then there’s difficulty, like having the best turbo engine. There might be a turbo engine that could be imagined to improve fuel efficiency by 10%, but we haven’t found it yet. Or like GLP-1 drugs, which are just a few bytes of data…

Which part of this field do you think you want to work in? Is it a lot of medium-level intelligence tasks, or the very top cognitive tasks?

Elon Musk:

I just used customer service as an example that has a very substantial revenue stream but might not be too difficult to solve. If you can simulate a human sitting at a desk, that’s customer service. That’s medium-level intelligence. You don’t need to spend years training someone. You don’t need several sigma excellent engineers to do this. But with this capability, once you have an effective digital Optimus working, you can run any application.

Suppose you are designing chips. You can run regular applications like Cadence and Synopsys and so on. You can run 1,000 or 10,000 instances at the same time and say, “Given this input, I got this output for the chip.” At some point, you will know what the chip should look like without using any tools.

Basically, you should be able to do digital chip design. You can do chip design. You move up the difficulty curve. You can do CAD design. You can use NX or any CAD software to design things.

John Collison:

So you think starting with the simplest tasks and then moving up the difficulty curve?

Devraksh Patel:

As a broader goal with a complete digital colleague simulator, you say, “All companies pursuing revenue maximization want to do this, xAI is one of them, but we will win because of a secret plan.” But everyone is trying different things in terms of data and algorithms

Elon Musk:

"We've tried data, we've tried algorithms. What else can we do?"

Dwarkesh Patel:

This seems like a highly competitive field. How do you plan to win? That's my big question.

Elon Musk:

I think we see a path to achieving it. I think I know how to do it because it's essentially the same path Tesla used to create autonomous driving. Not driving, but operating a computer screen. Essentially, it's an autonomous driving computer.

John Corliss:

Is this path about following human behavior and training on a large amount of human behavior data?

Dwarkesh Patel:

Isn't that... training?

Elon Musk:

Obviously, I'm not going to reveal the most sensitive secrets on a podcast. I need at least three more pints of Guinness to do that.

John Corliss:

What will xAI's business be? Will it be consumer-facing or enterprise? What will the mix be like? Will it be similar to other labs—

Elon Musk:

You're saying "labs" again. It's a company.

Dwarkesh Patel:

This psychological warfare is deep, Elon.

Elon Musk:

"A company pursuing revenue maximization," to be clear. Those GPUs won't pay for themselves.

John Corliss:

Right. What is the business model? What will the revenue sources be in a few years?

Elon Musk:

Things will change very quickly. I'm stating the obvious. I call AI a supersonic tsunami. I like alliteration. What's going to happen is—especially when you have scaled humanoid robots—they will manufacture products and provide services more efficiently than human companies. Amplifying human company productivity is just a short-term thing.

Dwarkesh Patel:

So you expect fully digital companies, rather than SpaceX becoming partially AI-driven?

Elon Musk:

I think there will be digital companies, but... some of this might sound a bit apocalyptic, okay? But I'm just stating what I think will happen. It's not that I'm a doomsayer or anything. It's just what I think will happen.

Companies composed purely of AI and robots will far outperform any company with human involvement. Computers used to be jobs that humans did. You would get a job as a calculator. They would fill entire buildings, 20-30 stories high, all humans just doing calculations. Now, that entire building of human calculators can be replaced by a laptop with a spreadsheet.

The amount of calculations that spreadsheet can do far exceeds that of an entire building of human calculators. You might think, "Well, what if only some cells in your spreadsheet are calculated by humans?" In reality, that would be much worse than having all cells calculated by computers. What will actually happen is that purely AI, purely robotic companies or collectives will far outperform any company with human involvement And this will happen very quickly.

Optimus Robot

Devaksh Patel:

Speaking of closed loops... Optimus. In terms of manufacturing goals, your company has been supporting American manufacturing in the hard tech field. But in the area where Tesla has always been a leader—now you want to enter the humanoid robot field—there are dozens of companies in China manufacturing this cheaply and at scale, and they are highly competitive. So, give us some advice or plans on how America can build a humanoid robot army or electric vehicles at scale and as cheaply as China does.

Elon Musk:

There are really only three difficult things about humanoid robots. Intelligence in the real world, hands, and scalable manufacturing. I have yet to see any demonstration robot that has great hands with all the degrees of freedom of a human hand. Optimus will have this. Optimus really has this.

Devaksh Patel:

How do you achieve this? Is it just that the motors have the right torque density? What are the hardware bottlenecks?

Elon Musk:

We have to design custom actuators, basically custom-designed motors, gears, power electronics, controllers, sensors. Everything has to be designed from first principles of physics. There is no off-the-shelf supply chain for this.

Devaksh Patel:

Can you manufacture these at scale?

Elon Musk:

Yes.

John Corliss:

From an electromechanical perspective, besides the hands, are there other difficulties? Once you solve the hand problem, is everything else solved?

Elon Musk:

From an electromechanical perspective, hands are more difficult than all the other parts combined. It turns out that the human hand is quite remarkable. But you also need intelligence in the real world. The intelligence developed by Tesla for cars is very applicable to robots, primarily visual input. Cars receive visual information, but they are also listening for alarms. They receive inertial measurement data, GPS signals, and other data, combining video (mainly video), and then output control commands.

Your Tesla receives 1.5 gigabytes of video per second and outputs control signals at a rate of 2 kilobytes per second, with a video frequency of 36 Hz and a control frequency of 18 Hz.

John Corliss:

For the realization of robotics, you might have the intuition that it takes several years from a compelling demonstration to something that can actually be used in the real world. Ten years ago, you had a truly compelling autonomous driving demonstration, but only now do we have services like Robotaxi and Waymo scaling up. Doesn’t this make us pessimistic about home robots? Because we don’t even have a truly compelling demonstration, like a really advanced hand.

Elon Musk:

Well, we have been researching humanoid robots for a while now. I think it’s been about five or six years. A lot of what is done in cars applies to robots. We will use the same Tesla AI chips in the robots. We will use the same basic principles This is a very similar AI.

Robots have more freedom than cars. If you only see it as a bitstream, AI is mainly about compressing and correlating two bitstreams. For video, you have to do a lot of compression, and it has to be just right. You have to ignore the unimportant things. You don't care about the details of the leaves on the trees by the roadside, but you care a lot about the road signs, traffic lights, pedestrians, and even whether the people in another car are looking at you. Some details are very important.

Cars will ultimately convert 1.5 gigabytes of video per second into 2 kilobytes of control output per second. So you have multiple stages of compression. You have to get all the stages right and then correlate them with the correct control output. Robots essentially do the same thing.

Humans are the same. We are indeed photon input, control output. The vast majority of your life is: visual, photon input, followed by motion control output.

Devakish Patel:

Simply put, there seems to be a difference between humanoid robots and cars... The basic actuators of a car are how you steer and how you accelerate. In robots, especially those with dexterous arms, there are dozens of degrees of freedom. Especially for Tesla, you have the advantage of millions of hours of human demonstration data collected from cars. You can't deploy a non-working Optimus robot in the same way to gather that data. So, between increased degrees of freedom and extreme scarcity of data...

Elon Musk:

Yes. You pointed out an important limitation and the difference from cars. We will soon have 10 million cars on the road. It's hard to replicate that massive training flywheel. What we need to do for robots is to manufacture a large number of robots and put them in an Optimus academy so they can self-play in reality. We are actually building this. We can have at least 10,000 Optimus robots, maybe 20,000 to 30,000, to self-play and test different tasks.

Tesla has a pretty good reality generator, a physically accurate reality generator, which we made for cars. We will do the same for robots. We have actually done this for robots. So you have tens of thousands of humanoid robots performing different tasks. You can simulate millions of robots in a simulated world. You use tens of thousands of robots in the real world to bridge the gap between simulation and reality. Narrowing the gap from simulation to reality.

Devakish Patel:

How do you see the synergy between xAI and Optimus, given that you emphasize the need for this world model, and you want to use very intelligent AI as the control plane, with Grok responsible for slower planning, while the motor strategy is more low-level? What would the synergy be between these things?

Elon Musk:

Grok will coordinate the behavior of the Optimus robots. Suppose you want to build a factory. Grok can organize the Optimus robots and assign tasks to them to build the factory and produce anything you want.

John Collison:

So you don't need to merge xAI and Tesla, right? Because these will eventually become so...

Elon Musk:

What did we say about other companies during our discussions?

Dvarkesh Patel:

We had another pint of Guinness, Elon. What are you waiting for to say, "We want to make 100,000 Optimus robots"?

Elon Musk:

"Optimi." Since we're defining proper nouns, we should also define the plural form of proper nouns. We want to proper noun the plural form, so it's Optimi.

Dvarkesh Patel:

In terms of hardware, what do you want to see? Do you want to see better actuators? Or do you just want to make the software better? What are we waiting for before starting the third generation of mass production?

Elon Musk:

No, we are moving in that direction. We are pushing forward with mass manufacturing.

Dvarkesh Patel:

But do you think the current hardware is good enough, and you just want to deploy as many as possible right now?

Elon Musk:

Scaling up production is very difficult. But I think Optimus 3 is the right version to achieve an annual production of about 1 million units. I think before reaching 10 million units per year, you would want to upgrade to Optimus 4.

John Collison:

Okay, but can we achieve 1 million units with Optimus 3?

Elon Musk:

Scaling up manufacturing is very difficult. The output per unit time always follows an S-curve. It starts extremely slow, then there is exponential growth, followed by linear growth, and then logarithmic growth until it eventually approaches a certain number. The initial production of Optimus will be a stretched S-curve because many components of Optimus are brand new. There is no ready-made supply chain.

Everything in the Optimus robot, including actuators, electronics, etc., is designed from first principles. It's not selected from a catalog. These are all custom designs. I don't think there is a single thing—

John Collison:

How deep does that go?

Elon Musk:

I think we haven't even started making custom capacitors, maybe. There is nothing you can directly select from a catalog, no matter how much you spend. This means that the S-curve for Optimus, the output per unit time, i.e., how many Optimus robots you can manufacture each day, will initially climb slower than those products with ready-made supply chains. But it will eventually reach 1 million units.

Dvarkesh Patel:

When you see Chinese companies like Yushu Technology selling humanoid robots for $6,000 or $13,000, are you hoping to bring your Optimus bill of materials cost down lower than that to do the same thing? Or do you think they are not comparable in quality? What makes them sell so cheaply? Can we reach that level?

Elon Musk:

Our Optimus is designed to have a high level of intelligence and the same (if not greater) electromechanical dexterity as humans. Yushutech does not have that. Its size is also quite large. It must handle heavy loads for long periods without overheating or exceeding the power limits of its actuators. It stands 5 feet 11 inches tall, which is quite tall. It has a lot of intelligence. So it will be more expensive than a small, unintelligent robot.

John Coriason:

But with greater capabilities.

Elon Musk:

But not by much. The key is that over time, as Optimus robots manufacture Optimus robots, the costs will drop rapidly.

John Coriason:

What will the initial billion Optimus (Optimi) do? What are their most efficient and best uses?

Elon Musk:

I think you would start with simple tasks that you can expect them to do well.

John Coriason:

In homes or in factories?

Elon Musk:

The best initial use for robots will be any continuous operation, any 24/7 uninterrupted operation, because they can work continuously.

Devakish Patel:

In the Gigafactory, what percentage of the work currently done by humans could be done by the third-generation Optimus?

Elon Musk:

I'm not sure. Maybe 10-20%, maybe more, I don't know. We will not reduce the number of employees, to be clear, we will increase the number of employees. But we will increase output. The number of robots or cars produced per employee... the total number of Tesla employees will increase, but the output of robots and cars will increase disproportionately. The number of cars and robots produced per employee will increase dramatically, but the number of employees will also increase.

Why China is assumed to win

John Coriason:

We've talked a lot about Chinese manufacturing here. We've also discussed some related policies, like the solar tariffs you mentioned. You think it's a bad idea because we can't scale solar in the U.S.

Elon Musk:

The power output in the U.S. needs to be scaled up.

John Coriason:

You can't scale without good power sources.

Elon Musk:

You just need to get power somehow.

John Coriason:

What I mean by this is, if you were in charge, if you set all the policies, what else would you change? You would change the solar tariffs, that's one thing.

Elon Musk:

I would say that any factors limiting power need to be addressed, as long as they are not extremely harmful to the environment.

John Coriason:

So perhaps some permitting reforms would also be included?

Elon Musk:

Some licensing reforms are taking place. Many licenses are at the state level, but any federal-level... this administration has done a good job of eliminating licensing barriers.

I'm not saying all tariffs are bad.

John Coriason:

Solar tariffs.

Elon Musk:

Sometimes, if another country subsidizes the export of a certain product, you have to impose countervailing tariffs to protect domestic industries from the effects of subsidies from other countries.

John Coriason:

What else would you change?

Elon Musk:

I don't know how much the government can actually do.

Elon Musk:

It's important to recognize that in most areas, China's manufacturing is very advanced. There are only a few areas that are not. China is a manufacturing powerhouse, at another level.

John Coriason:

Very impressive.

Elon Musk:

If you look at ore smelting, China averages about twice the total of the rest of the world in ore smelting. In some areas, like the smelting of gallium used for solar cells, I think they account for 98% of gallium smelting. So in fact, China is very advanced in most manufacturing fields.

John Coriason:

There seems to be unease about this supply chain dependence, but no real action is being taken.

Elon Musk:

Supply chain dependence?

John Coriason:

For example, the gallium smelting you just mentioned. All the rare earth materials.

Elon Musk:

Rare earths for sure, as you know, they are not rare. We actually mine rare earth ores in the U.S., load the rocks onto trains, then ship them to China, and then they are transported by train to China's rare earth smelting plants, where they are smelted into magnets, made into motor components, and then shipped back to the U.S. So what we really lack is the ore smelting capacity in the U.S.

John Coriason:

Isn't that worth policy intervention?

Elon Musk:

It is worth it. I think some things are being done in this regard. But honestly, we need Optimus to build ore smelting plants.

Devakish Patel:

So, do you think China's main advantage is the abundance of skilled labor? Is this something Optimus can solve?

Elon Musk:

Yes. China's population is about four times ours.

Devakish Patel:

I mean, there is a concern here. If you think human resources are the future, then right now if the skilled labor required for manufacturing determines who can produce more humanoid robots, China has more of that labor. It produces more humanoid robots, and therefore it gets the future of Optimus first.

Elon Musk:

Well, let's see. Maybe.

Devakish Patel:

This will keep that index growing. It seems you pointed out that reaching 1 million Optimi requires the manufacturing capacity that Optimus was supposed to help us achieve. Is that right?

Elon Musk:

You can quite quickly close that recursive loop.

John Corrison:

With a small number of Optimus?

Elon Musk:

Yes. So you close the recursive loop and let the robots help manufacture robots. Then we can strive to reach tens of millions of units per year. Maybe. If you start reaching hundreds of millions of units per year, you will become the most competitive country, far surpassing others.

We certainly can't win just with humans, because China's population is four times ours. Frankly, the U.S. has won for too long... A professional sports team that wins for a long time often becomes complacent and arrogant. That's why they stop winning, because they no longer work as hard. So frankly, my observation is that the average work ethic in China is higher than in the U.S. It's not just that the population is four times ours, but people are putting in more work.

So you can try to rearrange human resources, but you still only have a quarter of China's population—assuming per capita productivity is the same, but I think it might actually be different; China may have an advantage in per capita productivity—what we do will only be a quarter of what China does. So we can't win on the human front.

Our birth rate has been low for a long time. Since around 1971, the U.S. birth rate has been below replacement level. We have a large population retiring, and domestic deaths are soon going to exceed births. So we certainly can't win on the human front, but we might have a chance with robots.

John Corrison:

Is there anything else in the past that you wanted to manufacture but couldn't because it was labor-intensive or too expensive, and now you can say, "Oh, we can finally do something because we have Optimus"?

Elon Musk:

Yes, we wanted to build more ore smelting plants at Tesla. We just completed the construction of a lithium refining plant in Corpus Christi, Texas, and have started lithium refining. We have a nickel refining plant in Austin for cathode materials. This is the largest cathode material, nickel, and lithium refining plant outside of China.

The cathode team would say, "We have the largest and actually the only cathode refining plant in the U.S." Not only the largest but also the only one.

John Corrison:

So many superlatives.

Elon Musk:

So even if it is the only one, the scale is large. But there are other things. You can do more refining plants to help the U.S. improve its refining capacity. Frankly, there are many jobs that are suitable for Optimus to do, and very few Americans are willing to do them.

John Corrison:

Is refining work too dirty or what—

Elon Musk:

Actually, no. Our refinery has no toxic emissions or anything like that. The cathode nickel refinery is right in Travis County.

John Coriason:

Why can't it be done manually?

Elon Musk:

It can, but you'll quickly run out of manpower.

John Coriason:

Ah, I see. Well.

Elon Musk:

No matter what you do, the U.S. population is only a quarter of China's. So if you have them do this, they can't do something else. So how do you build that refining capacity? Well, you can use Optimi to do it.

Not many Americans are eager to do refining work. I mean, how many have you encountered? Very few. Very few people are eager to do refining.

Devakish Patel:

BYD is getting close to Tesla in terms of electric vehicle production or sales. What do you think will happen in the global market as China's electric vehicle production expands?

Elon Musk:

China is extremely competitive in manufacturing. So I think there will be a large influx of Chinese vehicles and basically most manufactured goods. As I said, China is probably doing twice the refining work of the rest of the world combined right now. So if you dig into the fourth and fifth levels of the supply chain...

At the most basic level, you have energy, then mining and refining. These foundational layers, as I mentioned, roughly estimate that China's refining volume is twice that of the rest of the world. So anything given will have a Chinese component because China's refining workload is twice that of the rest of the world. And they will keep doing it until finished cars.

I mean, China is a powerhouse. I think this year China's electricity output will exceed that of the U.S. by three times. Electricity output is a reasonable representation of the economy. To operate factories and everything, you need electricity. This is a good proxy indicator of the real economy. If China's electricity output exceeds that of the U.S. by three times, it means its industrial capacity—roughly approximated—will be three times that of the U.S.

Devakish Patel:

Reading between the lines, it sounds like you're saying that unless there are breakthroughs in humanoid robotics in the next few years, China will dominate the entire manufacturing/energy/raw materials chain, whether in AI, manufacturing electric vehicles, or producing humanoid robots.

Elon Musk:

Without breakthrough innovations in the U.S., China will completely dominate.

Devakish Patel:

Interesting.

Elon Musk:

Yes.

John Coriason:

Robotics is a major breakthrough innovation.

Elon Musk:

Well, to expand AI in space, you basically need humanoid robots, you need real-world AI, and you need an orbital capacity of 1 million tons per year. Assuming we get the mass accelerator on the moon started, that's my favorite thing, then I think—

John Collison:

We’ve solved all the problems.

Elon Musk:

I call it a victory. A huge victory.

John Collison:

You can finally be satisfied. You’ve accomplished something.

Elon Musk:

Yes.

John Collison:

You have a mass accelerator on the moon.

Elon Musk:

I just want to see that thing running.

John Collison:

Is that from a science fiction novel? Or where...?

Elon Musk:

Well, actually, there’s a book by Heinlein. "The Moon is a Harsh Mistress."

John Collison:

Okay, yes, but that’s a bit different. That’s a gravity slingshot or...

Elon Musk:

No, they have a mass accelerator on the moon.

John Collison:

Well, yes, but they use it to attack Earth. So maybe that’s not the best...

Elon Musk:

Well, they use it to... declare their independence.

John Collison:

Exactly. What are your plans for the mass accelerator on the moon?

Elon Musk:

They declared independence. The Earth government disagreed, and they threw things until the Earth government agreed.

John Collison:

That book is interesting. I think it’s much better than his other widely read book, "Stranger in a Strange Land."

Elon Musk:

The word "Grok" comes from "Stranger in a Strange Land." The first two-thirds of "Stranger in a Strange Land" are good, then it gets very strange in the third part. But there are still some good concepts in it.

SpaceX: The Benefits of "Fervent Urgency"

John Collison:

We’ve discussed a lot about your management system. You’ve interviewed the first few thousand employees at SpaceX, as well as many others.

Elon Musk:

That clearly doesn’t scale.

John Collison:

Well, yes, but what doesn’t scale?

Elon Musk:

Me.

John Collison:

Of course, of course. I know. But what are you looking for?

Elon Musk:

There really isn’t enough time in the day. It’s impossible.

John Collison:

But what is it you’re looking for that allows another person who is good at interviewing and hiring... what is that ineffable quality?

Elon Musk:

At this point, I probably have more training data in evaluating technical talent—I think various talents, especially technical talent—because I’ve done so many technical interviews and then seen the results. So my training set is very large and very broad

Generally speaking, what I ask for are the key points that demonstrate exceptional ability. These things can be quite unconventional. They don't necessarily need to be in a specific field, but there should be evidence of exceptional ability. So if someone can point out even one thing, but preferably three things, that make you think "wow, wow, wow," that's a good sign.

Devaksh Patel:

Why does it have to be determined by you?

Elon Musk:

No, it doesn't have to be me. It can't be. The total number of people in all companies is 200,000.

John Collison:

But in the early days, what were you looking for in those interviews that couldn't be delegated to others at the time?

Elon Musk:

I think I needed to build my training set. I'm not always right. I make mistakes too, but I can see when I thought someone would perform well, but they didn't. Then why didn't they perform well? What can I do? I think it's about reinforcing my learning so that I have a better hit rate in future interviews. My hit rate is still not perfect, but it's high.

Devaksh Patel:

Some people didn't succeed; what are some surprising reasons?

Elon Musk:

Surprising reasons...

Devaksh Patel:

For example, they don't understand the technical field, etc. But now you have a lot of long-tail cases, "I was really excited about this person, but they didn't succeed." I'm curious why that happened.

Elon Musk:

Usually, I tell people—I think I also tell myself, as a goal—don't look at the resume. Just trust your interactions. A resume might look very impressive, like "wow, the resume looks good." But if the conversation 20 minutes later doesn't make you feel "wow," you should trust the conversation, not that piece of paper.

John Collison:

I feel part of your approach is... A few years ago, there was a media meme that Tesla was the door to executive talent. And in fact, I think you see that the executive team at Tesla has been very stable over the past few years, and mainly consists of internal promotions.

Then at SpaceX, you have people like Mark Juncosa, Steve Davis—

Elon Musk:

Steve Davis is now running The Boring Company.

John Collison:

Bill Riley, and people like that. It feels like part of the reason it works is having very capable technical deputies. What do these people have in common?

Elon Musk:

Well, the current average tenure of Tesla's executive team is probably around 10-12 years. Quite long. But Tesla has gone through an extremely rapid growth phase, so everything accelerated. You know, companies go through different orders of magnitude in scale. Managing a team of, say, 50 people is different from managing teams of 500, 5,000, or 50,000 people

John Collison:

Some people can't keep up with the development.

Elon Musk:

It's not always the same team. So if a company is growing very fast, the turnover rate of executive positions is usually proportional to the speed of growth.

Tesla has an additional challenge, which is that when Tesla is in a very successful period, we are ruthlessly poached. Like, ruthlessly. When Apple had their electric vehicle project, they carpet-bombed Tesla with recruitment calls. Engineers were literally pulling the phone lines out.

John Collison:

"I was just trying to work here."

Elon Musk:

Yes. "If I get another call from an Apple recruiter..." But the initial offers they make, often without even an interview, can be about twice what Tesla pays. So we have a bit of a "Tesla fairy dust" effect, like "Oh, if you hire an executive from Tesla, suddenly everything will succeed."

I have also fallen victim to the fairy dust effect, like "Oh, we'll hire someone from Google or Apple, and they'll be successful immediately," but that's not how it works. People are just people. There's no magical fairy dust. So when we have a fairy dust problem, we are ruthlessly poached.

Moreover, Tesla is an engineering company, and especially in Silicon Valley, people can easily... they don't need to change much in their lives. Their commute is the same.

John Collison:

So how do you prevent that? How do you prevent the fairy dust effect where everyone wants to poach all your people?

Elon Musk:

I don't think we have much way to stop it. That's also why Tesla... really, at the same time in Silicon Valley, faced the fairy dust effect, which meant there was very, very aggressive hiring going on at that time.

John Collison:

So would being in Austin help with that?

Elon Musk:

Being in Austin helps. Most of Tesla's engineering is still in California. Getting engineers to move... I call it the "other half" problem.

John Collison:

Yes, the "other half" has a job.

Elon Musk:

Exactly. So for the Starbase, this is particularly difficult because the chances of finding a job outside of SpaceX...

John Collison:

In Brownsville, Texas...

Elon Musk:

...are very low. Quite difficult. It's like a tech monastery, remote and mostly male.

Devakish Patel:

Not much better than San Francisco.

John Collison:

If you look back at those who are very technically capable at Tesla, SpaceX, etc., what do you think they have in common besides... Is it just that they are very sharp in rocket technology or technical fundamentals, or do you think there are some organizational abilities?

Is it their ability to collaborate with you? Is it their flexibility but not overly flexible? What makes a good partner for you?

Elon Musk:

I don't think of it as partnership. If someone can get things done, I like them; if they can't, I don't. It's that simple. It's not some quirky preference. If someone executes well, I'm a loyal fan; if they don't, I'm not. But that doesn't mean catering to my quirky preferences. I certainly try not to let things turn into catering to my quirky preferences.

Overall, I think it's a good idea to hire based on talent, drive, and credibility. And I think being kind-hearted is important. I used to underestimate that. So, are they good people? Are they credible? Are they smart, talented, and hard-working? If so, you can add domain knowledge.

But these basic traits, these fundamental attributes, you can't change. So most people at Tesla and SpaceX do not come from the aerospace or automotive industries.

Dwarkesh Patel:

As your company grows from 100 people to 1,000 to 10,000, what must change the most about your management style? You are known for your very micro-management, diving into details.

Elon Musk:

Nano-management, thank you. Pimi management. Flymi management.

John Collison:

Go on.

Elon Musk:

We have to go all the way down to Planck's constant. All the way down to the Heisenberg uncertainty principle.

Dwarkesh Patel:

Can you still dive into details as you would like? Would your company be more successful if it were smaller? What do you think about that?

Elon Musk:

Because my time is fixed every day, as the affairs grow and the scope of activities expands, my time will inevitably be diluted. I can't really be a micro-manager because that would mean I have thousands of hours every day. For me, micro-management is logically impossible.

Now, sometimes I will dive into a specific issue because that specific issue is a limiting factor for the company's progress. The reason for diving into some very detailed matters is that it is the limiting factor. It's not arbitrary diving into trivial matters.

From a time perspective, diving into trivial matters arbitrarily is physically impossible for me. That would lead to failure. But sometimes, the small things are key to victory.

John Collison:

It is well known that you changed the design of Starship from composite materials to steel.

Elon Musk:

Yes.

John Collison:

It was your decision. It wasn't people coming to you saying, "Oh, boss, we found something better." You pushed through some resistance. Can you tell us the whole process of how you came up with the concept of switching to steel?

Elon Musk:

Desperation, I would say. Initially, we intended to make Starship out of carbon fiber. Carbon fiber is quite expensive. When you go into mass production, the cost of anything can start to approach its material cost

The problem with carbon fiber is that the material cost is still very high. Especially if you choose a special carbon fiber that is high-strength and can withstand low-temperature oxygen, its cost is about 50 times that of steel. At least theoretically, it would be lighter. People generally think steel is heavy, and carbon fiber is light.

For room temperature applications, such as Formula One racing, static aerodynamic structures, or any type of aerodynamic structure, you might be better off using carbon fiber. The problem is that we are trying to make this huge rocket out of carbon fiber, but progress is extremely slow.

John Corliss:

Was the initial choice just because it was light?

Elon Musk:

Yes. At first glance, most people would think that carbon fiber is the choice for making lightweight things. The problem is that when you make something very large out of carbon fiber and then try to cure the carbon fiber effectively, meaning not at room temperature, because sometimes you have 50 layers of carbon fiber... Carbon fiber is essentially carbon strands and glue. To achieve high strength, you need a high-pressure autoclave. Basically, it's a high-pressure oven. If you have something huge, the autoclave has to be bigger than the rocket.

We are trying to make a high-pressure autoclave that is larger than any existing one. Or you can do room temperature curing, which takes a long time and has issues. The bottom line is that we are making very slow progress with carbon fiber.

Devakish Patel:

The meta question is, why does it have to be you making that decision? There are many engineers on your team.

John Corliss:

How did the team not discover steel?

Devakish Patel:

Yes, exactly. This is part of a broader question about understanding your comparative advantage within your company.

Elon Musk:

Because we are making very slow progress with carbon fiber, I thought at the time, "Well, we have to try something else." For Falcon 9, the main body is made of aluminum-lithium alloy, which has a very good strength-to-weight ratio. In fact, for its application, the strength-to-weight ratio may be comparable to or even better than carbon fiber. But aluminum-lithium alloy is very difficult to process.

To weld it, you have to use a process called friction stir welding, where you connect the metal without going into a liquid state. It's a bit incredible that you can do this. But for this specific type of welding, you can do it. It's very difficult. Suppose you want to modify or add something to the aluminum-lithium alloy, you now have to use mechanical connections with seals. You can't just weld it on. So I wanted to avoid using aluminum-lithium alloy for the main structure of Starship.

At that time, there was a special carbon fiber that had very good weight characteristics. For rockets, what you really want is to maximize the percentage of fuel in the rocket and minimize the mass. But as I said, we are making very slow progress. I said, "At this rate, we will never get to Mars. So we have to think of other ways."

I didn't want to use aluminum-lithium alloy because of the difficulties with friction stir welding, especially at scale. It's already difficult at a diameter of 3.6 meters, let alone 9 meters or larger. Then I said, "What about steel?"

I have a clue here, because some early rockets in the United States used very thin steel. The Atlas rocket used steel balloon tanks. It's not that steel was never used before. In fact, it has been used. When you look at the material properties of stainless steel, full hard, strain-hardened stainless steel has a strength-to-weight ratio at low temperatures that is actually similar to carbon fiber.

If you look at the material properties at room temperature, it seems that the weight of steel would be twice that of carbon fiber. But if you look at the material properties of specific grades of full hard stainless steel at low temperatures, you can actually achieve a strength-to-weight ratio similar to that of carbon fiber.

As for Starship, both the fuel and oxidizer are at low temperatures. For Falcon 9, the fuel is rocket-grade kerosene, which is basically a very high purity aviation fuel. That's roughly at room temperature. Although we actually cool it to slightly below room temperature, like chilled beer.

John Corrison:

Delicious.

Elon Musk:

We do cool it, but it's not at low temperatures. In fact, if we cool it to low temperatures, it would turn waxy. But for Starship, it's liquid methane and liquid oxygen. They have similar liquid temperatures. Basically, the entire main structure is almost at low temperatures. So you get strain-hardened 300 series stainless steel. Because almost everything is at low temperatures, it actually has a strength-to-weight ratio similar to carbon fiber.

But the raw material cost is 50 times lower, and it's very easy to process. You can weld stainless steel outdoors. You can weld stainless steel while smoking a cigar. It's very tough. Easy to modify. If you want to add something, just weld it on. Very easy to process, and the cost is very low.

As I said, at low temperatures, its strength-to-weight ratio is similar to that of carbon fiber. Then when you consider that we significantly reduced the mass of the heat shield because the melting point of steel is much higher than that of aluminum... about twice the melting point of aluminum.

John Corrison:

So you can make the rocket withstand higher temperatures?

Elon Musk:

Yes, especially for spacecraft returning like a blazing meteor. You can significantly reduce the mass of the heat shield. You can halve the mass of the heat shield on the windward side, and the leeward side doesn't need any heat shield.

The end result is that, in fact, a steel rocket is lighter than a carbon fiber rocket because the resin in the carbon fiber rocket starts to melt. Basically, the temperature resistance of carbon fiber is similar to that of aluminum, while steel can operate at twice the temperature. These are very rough approximations.

John Corrison:

I won't nitpick the rocket math with you.

Elon Musk:

People will say, "Oh, he said twice. It should actually be 0.8 times." I would say, shut up, jerk.

Devakish Patel:

The comments section is mainly about this.

Elon Musk:

Damn it. The key is, in hindsight, we should have used steel from the beginning. Not using steel was stupid.

John Corrison:

Okay, but just to confirm with you, what I heard is that, aside from early American rockets, steel is a riskier and more unproven path. And carbon fiber is a worse but more mature path. So we need you to push: "Hey, we're going to take this riskier path and figure it out." So you're fighting against some kind of conservatism.

Elon Musk:

That's why I initially said the problem is that we're not progressing fast enough. We're even having a hard time making a small carbon fiber barrel section without wrinkles. Because at that large size, you have to have many layers of carbon fiber. You have to cure it, and it has to be cured in a way that is wrinkle-free and defect-free.

The toughness of carbon fiber is far less than that of steel. Its toughness is much smaller. Stainless steel will stretch and bend, while carbon fiber tends to shatter. Toughness is the area under the stress-strain curve. Generally, stainless steel will be better, to be precise, stainless steel.

John Coriessen:

Another question about Starship. I visited the interstellar base with Sam Taylor two years ago, and it was great. It was cool in many ways.

One thing I noticed is that people really take pride in the simplicity of things, and everyone wants to tell you that Starship is just a big soda can, we hire welders, and if you can weld on any industrial project, you can weld here. But they have a lot of pride in that simplicity.

Elon Musk:

Well, actually, Starship is a very complex rocket.

John Coriessen:

So that's what I want to ask. Is it simple or complex?

Elon Musk:

I think maybe they just want to express that you don't need prior experience in the rocket industry to work on Starship. As long as someone is smart, hardworking, and trustworthy, they can work on the rocket. They don't need prior rocket experience. Starship is the most complex machine ever made by humanity, far beyond others.

John Coriessen:

In what ways?

Elon Musk:

In any way, really. I can't think of a machine more complex than this. I can say that any project I can think of would be easier than this. That's why no one has ever built a fully reusable orbital rocket. It's a very difficult problem. Many smart people have tried before, very smart people with huge resources, but they failed.

And we haven't succeeded yet. Falcon is partially reusable, but the upper stage is not. I believe the design of Starship version three can be fully reusable. That kind of full reusability will enable us to become a multi-planetary civilization.

John Coriessen:

Can you talk about the circle?

Elon Musk:

Any technical problem, even something like the Large Hadron Collider, is easier than this problem.

John Coriessen:

We spent a lot of time discussing bottlenecks. Can you talk about what the current bottlenecks for Starship are, even at a high level?

Elon Musk:

Trying to make it not explode, overall. It really wants to explode.

John Collison:

Same old story. All those flammable materials.

Elon Musk:

We've had two boosters explode on the test stand. One destroyed the entire test facility. So just one mistake is enough. The energy contained in Starship is enormous.

John Collison:

Is that why it's harder than Falcon? Because it has more energy?

Elon Musk:

It has a lot of new technology. It's pushing the performance limits. The Raptor 3 engine is a very, very advanced engine. It's the best rocket engine ever made. But it really wants to explode. Just to make it clear, the rocket generates over 100 gigawatts of power at launch. That's 20% of U.S. electricity.

Devakish Patel:

That's just crazy.

John Collison:

Great comparison.

Elon Musk:

While also not exploding.

John Collison:

Sometimes.

Elon Musk:

Sometimes, yes. So I think, how can it not explode? There are thousands of ways it can explode, and only one way it doesn't. So we want it not only to not explode but to be able to fly reliably every day, like once an hour. Obviously, if it explodes frequently, it's hard to maintain that launch frequency.

John Collison:

Yes.

Elon Musk:

What is the biggest problem with Starship right now? It's making the heat shield reusable. No one has ever made a reusable orbital heat shield. So the heat shield must not shed a lot of tiles during ascent and also must not shed a lot of tiles or overheat the main body during reentry.

John Collison:

That's hard because it's fundamentally consumable?

Elon Musk:

Well, yes, but the brake pads in your car are also consumable, yet they last a long time.

John Collison:

Makes sense.

Elon Musk:

So it just needs to last a long time. We've managed to bring the spacecraft back and soft land it in the ocean. We've done that a few times. But it lost quite a few tiles. Without a lot of repairs, it's not reusable. Even though it achieved a soft landing, it can't be reused without a lot of work.

So in that sense, it's not truly reusable. That's the biggest remaining issue, a fully reusable heat shield. You want to be able to land it, refuel it, and fly again. You can't be doing that laborious checking of 40,000 tiles or something like that.

Devakish Patel:

When I read your biography, it seems like you are able to drive a sense of urgency, a sense of "this is scalable." I'm curious, do you think your other organizations

SpaceX and Tesla are really big companies now. How do you still maintain that culture? What went wrong with other companies that they can't do that?

Elon Musk:

I don't know.

Devakish Patel:

For example, today, you mentioned you held some SpaceX meetings. What do you do there to maintain that culture?

John Collison:

Is it about increasing urgency?

Elon Musk:

Well, I don't know. I think the urgency comes from the leaders of the company. I have a kind of fanatic urgency. So that fanatic urgency projects onto the rest of the company.

Devakish Patel:

Is it because of the consequences? They think, "Elon set a crazy deadline, but if I don't meet it, I know what will happen." Or is it simply because you can identify bottlenecks and eliminate them so people can act quickly? Why do you think your company can act quickly?

Elon Musk:

I constantly work on resolving constraints. When it comes to deadlines, I usually try to set a deadline that I think has at least a 50% chance of being met. So it's not an impossible deadline, but rather the most aggressive deadline I can think of that has a 50% chance of being achieved. This means that half the time it will be delayed.

There is a gas expansion law that applies to schedules. If you say we will do something in five years, to me that feels like infinite time, and it will expand to fill the available schedule and take five years.

Physics will limit the speed at which you can do certain things. So scaling up manufacturing, you move atoms and there is a rate at which you scale up manufacturing. That's why you can't just make a million things at once every year. You have to design the production line. You have to start it. You have to go through the S-curve of production.

What can I say that would really help people? Generally speaking, having a fanatic urgency is a very important thing. You want a positive timeline, and you want to figure out what the constraints are at any given point in time and help the team resolve that constraint.

John Collison:

The development of Starlink took years.

Elon Musk:

We have been talking about it since the early days of the company.

John Collison:

So you built a team in Redmond at that time, and then at some point you determined that this team was not working. It progressed slowly for years, so why didn't you act sooner, and why was that the right time to act? Why was that moment the right time to take action?

Elon Musk:

I have very detailed engineering reviews every week. This may be a very unusual level of detail. I don't know of any other company (at least in manufacturing) where the person running the company gets as deep into the details as I do. It's not that... I have a pretty good understanding of what is actually happening by asking detailed questions

I firmly believe in cross-level meetings, not to have the people reporting to me speak, but to have everyone they report to speak during the technical review. And there should be no prior preparation. Otherwise, you will get "fluff," as I've been saying lately.

John Corliss:

That's right. Very Gen Z way of saying it.

Devakish Patel:

Very Gen Z way of saying it.

Elon Musk:

Very Gen Z.

Devakish Patel:

How do you prevent prior preparation? Do you call on people randomly?

Elon Musk:

No, I just go around the room in turn. Everyone provides an update. This requires remembering a lot of information. If you have meetings weekly or twice a week, you will remember what that person said. Then you can plot progress points. You can visualize points on a curve in your mind and then say, "Are we approaching a solution?"

I only take drastic action when I conclude that unless drastic action is taken, success is not within the realm of possible outcomes. So when I ultimately reach that conclusion, that unless drastic action is taken, we have no chance of success, I must take drastic action. I reached that conclusion in 2018, took drastic action, and solved the problem.

Devakish Patel:

You have a lot of companies. It sounds like you have a deep understanding of the bottlenecks in each company so that you can conduct such reviews with people.

You can scale it to five, six, or seven companies. Within one of those companies, you have many different mini-companies. What determines the maximum number here? Because you have 80 companies...?

Elon Musk:

80? No.

Devakish Patel:

But you already have a lot. That's impressive.

John Corliss:

By current numbers.

Devakish Patel:

That's right.

John Corliss:

We can barely maintain one company.

Elon Musk:

It depends. I actually don't have regular meetings with The Boring Company, so The Boring Company is doing well. Basically, if something is going well and making good progress, then I don't need to spend time on it.

I actually allocate my time based on constraints. Where are the problems? What are the obstacles we are facing? What is holding us back? I focus on—though I don't want to repeat this word—constraints.

Elon Musk:

Ironically, if something is going well, they don't see me often. But if something is going poorly, they see me often. Or not even poorly...

John Corliss:

If something is a constraint.

Elon Musk:

Constraints, that's right. Not necessarily that progress is poor, but that we need things to develop faster

John Collison:

When something at SpaceX or Tesla becomes a bottleneck, do you talk to the responsible engineers weekly or daily? How does it actually work?

Elon Musk:

Most things that are bottlenecks are weekly, some are twice a week. The review of the AI5 chip is twice a week. Every Tuesday and Saturday is chip review.

John Collison:

Is the duration of the meetings open-ended?

Elon Musk:

Technically, yes, but usually it's two to three hours. Sometimes shorter. It depends on how much information we need to go through.

John Collison:

That's another thing. I just want to clarify the differences here because the outcomes seem quite different. I think it's interesting to understand what the different inputs are. It feels like in the corporate world, first of all, as you said, it's not always the case that the CEO conducts engineering reviews, even though that's exactly what the company is doing.

But time is often finely sliced into half-hour meetings or even 15-minute meetings. It seems like you host more open-ended, "we keep discussing until we solve the problem" type of meetings.

Elon Musk:

Sometimes yes. But most seem to end more or less on time. Today's Starship engineering review took a bit longer because there were more topics to discuss. They are working on how to expand the orbital capability to over 1 million tons per year. That's quite challenging.

DOGE (Government Spending Cuts)

Devakish Patel:

Can I ask a question? You mentioned that Optimus and AI will bring double-digit economic growth in a few years.

Elon Musk:

Oh, like the economy? Yes. I think that's correct.

Devakish Patel:

So what does the DOGE spending cuts mean? If the economy is going to grow significantly?

Elon Musk:

Well, I think waste and fraud are not good things. I'm actually quite worried... If we don't have AI and robots, we are actually completely messing it up because the national debt is piling up crazily. The interest payments on the national debt have already exceeded the military budget, which is $1 trillion. So we are just paying interest exceeding $1 trillion. I was quite worried about this before. Maybe if I spend some time, we can slow down the pace of bankruptcy in the U.S. to buy us enough time for AI and robots to help solve the national debt issue.

Or rather, not help solve, but be the only thing that can solve the national debt. Without AI and robots, our country will 1000% go bankrupt and fail. Nothing else can solve the national debt issue. We just need enough time to build AI and robots to finish before bankruptcy.

Devakish Patel:

I guess I'm curious, when DOGE started, you had a huge reform capability.

Elon Musk:

It's not that huge.

Devraksh Patel:

Of course. I completely agree with your point that AI and robotics are crucial for driving productivity improvements and GDP growth. But why not directly address the things you pointed out, like tariffs on certain components or licensing?

Elon Musk:

I'm not the president. And even cutting out very obvious waste and fraud is extremely difficult. I find that even cutting out very obvious waste and fraud in government is extremely difficult because the government has to operate based on who is complaining.

If you cut off payments to fraudsters, they immediately come up with the most sympathetic-sounding reasons to continue receiving payments. They won't say, "Please let the fraud continue." They'll say, "You're killing baby pandas." Meanwhile, no baby pandas are dying. They are just made up. Fraudsters can concoct extremely compelling, heartbreaking but false stories, even though they sound very sympathetic. That's just how it is.

Maybe I should have been clearer. But at the time, I thought, wait, let's try to cut some waste and fat from the government. Maybe there shouldn't be 20 million people in the Social Security system marked as alive who are definitely dead and over 115 years old.

The oldest person in America is 114 years old. So it can be said with certainty that if someone is 115 years old in the Social Security database and marked as alive, it's either a data entry error... someone should call them and say, "It seems we've got your birthday wrong, or we need to mark you as deceased." One of those two scenarios.

John Corierson:

Receiving such a call would be very scary.

Elon Musk:

Well, that seems like a reasonable thing. For example, if their birthday is in the future and they have a small business administration loan, and their birthday is in 2165, then we either made a mistake or there's fraud. So we would say, "It seems we've got the century of your birth wrong."

John Corierson:

Or it's a great movie plot.

Elon Musk:

Yes. That's what I mean, absurd fraud.

Devraksh Patel:

Were these people receiving payments at the time?

Elon Musk:

Some people were receiving payments from the Social Security Administration. But the main avenue of fraud is marking someone as alive in the Social Security system and then committing fraud through all the other government payment systems. Because those other government payment systems only do a "are you alive" check against the Social Security database. It's an indirect means.

Devraksh Patel:

What do you estimate the total fraud caused by this mechanism to be?

Elon Musk:

By the way, the Government Accountability Office has done these estimates before. I'm not the only one. In fact, I believe the Government Accountability Office did an analysis during the Biden administration that roughly estimated the amount of fraud to be around $500 billion. So don't take my word for it. Look at the reports released during the Biden administration. How about that?

Devalaksh Patel:

Is this from the social security mechanism?

Elon Musk:

This is just one of many. It's important to recognize that the government is very ineffective at preventing fraud. Unlike companies, which have the incentive to prevent fraud because it affects their bottom line, the government just prints more money. You need care and competence. This is scarce at the federal level.

When you go to the DMV, do you think, "Wow, this is a fortress of competence"? Well, now imagine something worse than the DMV, because it's the DMV that can print money.

At least at the state level, DMVs need... state governments more or less need to stay within budget, or they go bankrupt. But the federal government just prints more money.

Devalaksh Patel:

If there is actually $500 billion in fraud, why can't it all be cut out?

Elon Musk:

You really need to take a step back and recalibrate your expectations of competence. Because you live in a world that has to balance its budget...

Devalaksh Patel:

Gotta buy a microphone.

Elon Musk:

Exactly. It's not like there's a huge, fundamentally indifferent bureaucratic monster and a bunch of outdated computers just sending out payments. One thing the DOGE team did that sounds very simple could save $100 to $200 billion a year. Just requiring that payments issued from the main Treasury computer—called PAM, Payment Account Master Control or something—where there are $5 trillion in payments each year—must have a grant code. Make it mandatory, not optional, and there must be something in the notes field.

You have to recalibrate how stupid things are. Payments are issued without grant codes, without matching any congressional appropriations, and without any explanations. This is why the Department of War, formerly the Department of Defense, cannot pass an audit because the information literally does not exist. Recalibrate your expectations.

Devalaksh Patel:

I want to better understand this $500 billion figure because there is an inspector general's report for 2024.

Elon Musk:

Why so low?

Devalaksh Patel:

Maybe, but they found that over seven years, they estimate social security fraud to be about $70 billion, so about $10 billion a year. So I'm curious what the other $490 billion is.

Elon Musk:

Federal government spending is $7.5 trillion a year. How competent do you think the government is?

Devalaksh Patel:

The discretionary spending there is about... 15%?

Elon Musk:

But that doesn't matter. Most fraud is non-discretionary. It's basically fraudulent Medicare, Medicaid, Social Security, disability benefits. The government makes a wide variety of payments. Many of those payments are actually block grants to the states. So in many cases, the federal government doesn't even have the information to know if fraud exists

Let's consider the method of reductio ad absurdum. The government is perfect, with no fraud. What do you estimate the probability of that possibility to be? Zero. So, would you say that fraud and waste, the government's efficiency is 90%? That would already be quite generous.

But if it's only 90%, that means there is $750 billion in waste and fraud every year. And it's not 90%. It doesn't have 90% effectiveness.

Devakish Patel:

This seems like a strange first-principles approach to estimating the amount of fraud in government. Like, how much do you think there is?

Anyway, we don't have to calculate it on the spot, but I'm curious—

Elon Musk:

You're quite familiar with the fraud situation at Stripe, right? People have been trying to commit fraud.

John Collison:

Yes, but as you said, it's a bit... we have indeed reduced fraud to a very low level, but what we're dealing with here is a more heterogeneous set of fraud vectors than we are.

Elon Musk:

But at Stripe, you have a high capability and you make an effort. You have both high capability and high attention, but fraud is still not zero. Now imagine a much larger scale, with much lower capability and much less attention.

In the early days of PayPal, we struggled to keep fraud at about 1% of the payment volume. That was very difficult. It took a lot of capability and attention just to reduce fraud to 1%. Now imagine you are an organization with far less capability and attention. Fraud would far exceed 1%.

John Collison:

Now looking back at politics and what happened there, how do you feel? From the outside, there are two things that are quite influential: one is the American Political Action Committee (America PAC), and the other is the acquisition of Twitter at that time. But there also seems to be quite a bit of heartache. How do you evaluate the whole experience?

Elon Musk:

I think those things had to be done to maximize the probability of a good future. Politics is often very tribal. People tend to lose objectivity in politics. They often have a hard time seeing the merits of the other side or the flaws of their own side. This is usually the norm. I think that's one of the things that surprised me the most.

You often can't reason with people at all. If they are in a certain tribe. They just believe that everything their tribe does is good, and anything the other political tribe does is bad. Convincing them to change their minds is almost impossible.

But I think overall, those actions—acquiring Twitter, getting Trump elected, even though it angered a lot of people—I think those actions are beneficial to civilization.

Devakish Patel:

How does this relate to the future you envision?

Elon Musk:

Well, America needs to be strong enough to have enough time to extend life to other planets and develop AI and robotics to a level that can ensure a good future.

On the other hand, if we fall into, say, communism or some extreme state oppression, that would mean we might not be able to become a multi-planetary species The government may stifle our progress in AI and robotics.

Devraksh Patel:

Optimus, Grok, etc. It's not just your products; any product from a company pursuing maximum revenue will eventually be utilized by the government over time. How does this concern manifest in what private companies should be willing to give to the government? What kind of guardrails?

Should AI models be required to do anything that the government outsources to them and asks them to do? Can Grok say, "Actually, even if the military wants to do X, no, Grok won't do that"?

Elon Musk:

I think the biggest danger of AI and robotics going wrong may be the government. Those who are against companies or are worried about companies should be most concerned about the government. Because the government is just a company in an extreme state. The government is just the largest company with a monopoly on violence.

I always find it a strange dichotomy that people think companies are bad, but the government is good, when the government is just the largest and worst company. But people have this dichotomy. They somehow think the government can be good, but companies are bad, which is not correct. The moral standard of companies is better than that of the government.

I do think this is something to be concerned about. The government could use AI and robotics to suppress the population. This is a serious concern.

Devraksh Patel:

As someone who makes AI and robots, how do you prevent this?

Elon Musk:

If you limit the power of the government—which is indeed the purpose of the U.S. Constitution, to limit the power of the government—then you might get better outcomes than having more government.

John Corliss:

Robotics will be open to all governments, right?

Elon Musk:

I don't know if it will be for all governments. It's hard to predict. I can say what the endpoint is or what it looks like many years from now, but it's hard to predict the path to get there. If civilization progresses, AI will far exceed the total sum of all human intelligence. The number of robots will far exceed that of humans. What happens in the process is very difficult to predict.

Devraksh Patel:

It seems one thing you could do is to directly say, "No matter what government X, you are not allowed to use Optimus to do X, Y, Z." Directly establish a policy. I think you recently tweeted that Grok should have a moral constitution. One of the clauses could be to limit what the government is allowed to do with this advanced technology.

Elon Musk:

Technically, if politicians pass a law and they are able to enforce that law, it is hard not to comply with it. The best we can have is a limited government where there are proper checks and balances between the executive, judicial, and legislative branches.

Devraksh Patel:

I'm curious because at some point, the limitations seem to come from you. You have Optimus, you have space GPUs

Elon Musk:

Do you think I would become the owner of the government?

Devakish Patel:

For SpaceX, this is already a reality—when it comes to certain critical matters, like the government really caring about launching certain satellites into space or whatever, it needs SpaceX. It is a necessary contractor.

You are building more and more components of future technology that will play similar roles across different industries. You might have the ability to set some policies, like suppressing classical liberalism in any way... "My company will not assist that kind of thing in any way," or similar policies.

Elon Musk:

I will do everything I can to ensure that anything under my control maximizes outcomes that are beneficial to humanity. I think any other approach is shortsighted because I am clearly part of humanity, so I like humanity. Supporting humanity.

Space GPU

Devakish Patel:

You mentioned that Dojo 3 will be used for space-based computing.

Elon Musk:

You really pay close attention to what I say.

Devakish Patel:

I don't know if you know about Twitter, but you have a large following.

Elon Musk:

It's obvious. I've posted all the secrets, how did you figure it out?

Devakish Patel:

How do you design chips for space? What changes?

Elon Musk:

You would want it to be designed to be more radiation-resistant and operate at higher temperatures. Roughly speaking, if you increase the operating temperature by 20% at Kelvin temperatures, you can halve the mass of the heat sink. So in space, operating at higher temperatures is beneficial.

You can do various things to shield memory. But neural networks will be very resistant to bit flips. Most of the radiation causes random bit flips. But if you have a model with trillions of parameters, a few bit flips don't matter. Heuristic programs will be much more sensitive to bit flips than some huge parameter files.

I just design it to run at high temperatures. I think other than making it run hotter, the way you design it is basically the same as what you do on Earth.

Devakish Patel:

Solar arrays account for most of the weight of satellites. Is there a way to make GPUs more powerful than those planned by Nvidia, TPU, etc., especially in a space-based world where they have particular advantages?

Elon Musk:

The basic calculation is that if each reticle chip can handle about 1 kilowatt of power, then you need 100 million complete reticle chips to achieve 100 gigawatts of power. Depending on your yield assumptions, this tells you how many chips you need to manufacture. If you want to have 100 gigawatts of power, you need 100 million chips that can continuously operate at 1 kilowatt, one chip per reticle. Basic calculation

Dvarkesh Patel:

100 million chips depend on... If you look at the size of chips like the Blackwell GPU, and how many can be produced from a wafer, you can get dozens or fewer from one wafer. So basically, in this world where so many need to be launched into space every year, you need to produce millions of wafers every month. Is that the goal of the TeraFab program? Millions of advanced process node wafers every month?

Elon Musk:

Yes, probably over a million or so. You also need to make memory.

Dvarkesh Patel:

Are you going to build memory factories?

Elon Musk:

I think TeraFab has to make memory. It has to make logic chips, memory, and packaging.

Dvarkesh Patel:

I'm curious how a person starts this. This is the most complex thing ever made by humans. Obviously, if anyone can handle this task, it's you. So you realize this is a bottleneck, and then you go to your engineers. What do you tell them to do? “I want to produce a million wafers every month by 2030.”

Elon Musk:

Exactly. That's what I want.

Dvarkesh Patel:

Do you call ASML? What's the next step?

John Coriason:

This question has been asked too many times.

Elon Musk:

We start by building a small factory and see what happens. Make mistakes on a small scale, then build a larger one.

Dvarkesh Patel:

Has the small factory been built?

Elon Musk:

No, not yet. We won't keep that cat in the bag. The cat will come out of the bag. Drones will be hovering over that damn thing. You'll be able to see its construction progress live on X.

Listen, I don't know, we might just fail, to be honest. Success is not guaranteed. Since we want to try to manufacture about 100 million... We hope to have 100 gigawatts of power by 2030, and chips that can handle 100 gigawatts of power. We'll take as many chips as our suppliers can give us. In fact, I've already told TSMC, Samsung, and Micron: "Please build more factories, build them faster." We will guarantee the purchase of the output from these factories. So they are already acting at the fastest speed. It's us plus them.

John Coriason:

There's a saying that those doing AI want to get a large number of chips as soon as possible. Then many suppliers, whether factories or turbine manufacturers, haven't ramped up production quickly.

Elon Musk:

No, they haven't.

John Coriason:

The explanation you hear is that they are inherently conservative. They are Taiwanese or German, and the story might be like that. They just don't believe... Is that really the explanation, or is there another reason?

Elon Musk:

Well, if someone has been in the computer memory industry for thirty or forty years, that makes sense...

John Collison:

They've seen the cycles.

Elon Musk:

They've seen ten booms and busts. There are a lot of layers of scars. During the boom, everything looks like it will be great forever. Then the crash happens, and they desperately try to avoid bankruptcy. Then there's another boom, and another crash.

John Collison:

Are there other ideas you think others should pursue that you, for some reason, are not pursuing right now?

Elon Musk:

There are a few companies exploring new ways to make chips, but they are not scaling up quickly.

John Collison:

I'm not even referring to AI internally, just in general.

Elon Musk:

People should do what they find a strong motivation to do, rather than some idea I suggest. They should do what they personally find interesting and are motivated to pursue.

But back to the limiting factors... I've used that word about a hundred times. In the current three to four-year timeframe, the limiting factor I see is chips. In a one-year timeframe, it's energy, power generation, electricity. I'm not sure if there is enough available power to start all the AI chips being manufactured.

By the end of this year, I think people will run into real trouble with starting chips... chip production will exceed the capacity to start chips.

Devakish Patel:

How do you plan to deal with that world?

Elon Musk:

We are trying to accelerate power production. I think that might be why xAI could become a leader, hopefully a leader. We will be able to start more chips faster than others because we are good at hardware.

Overall, those companies that call themselves labs often have innovative ideas that tend to flow... usually, the differences don't exceed about six months. Ideas flow as people move between different companies.

So I think you will basically hit a hardware wall, and whichever company can scale hardware the fastest will become the leader. So I believe xAI will be able to scale hardware the fastest and is therefore most likely to become the leader.

John Collison:

You joke or self-deprecate by using the term "limiting factor" again. But I think there is something profound in that. If you look at many of the things we've touched on throughout our conversation, perhaps this is a good point to conclude. If you imagine an aging, low-energy company, it will encounter some bottleneck but won't really address it.

Marc Andreessen once said, "Most people are willing to endure any degree of chronic pain to avoid acute pain." It feels like a lot of what we've been discussing is just facing acute pain, whatever it may be. "Well, we have to figure out how to use steel, or we have to figure out how to run chips in space." We will endure some immediate acute pain to truly address the bottleneck. So this seems to be a unifying theme

Elon Musk:

I have a high pain threshold. That helps.

John Collison:

Let's solve the bottleneck.

Elon Musk:

Yes. What I can say is that I think the future will be very interesting. As I said in Davos—I think I was there for about three hours—it's better to be wrong on the optimistic side than to be right on the pessimistic side for quality of life. If you're wrong on the optimistic side, you'll be happier than if you're wrong on the pessimistic side. So I recommend being wrong on the optimistic side.

John Collison:

Thank you for saying that.

Devakish Patel:

Cool. Elon, thank you for doing this interview.

John Collison:

Thank you.

Elon Musk:

Alright, thank you all. Okay.

John Collison:

Great endurance.

Devakish Patel:

Hope this is within the tolerance and not too painful