Musk has painted another astonishing growth blueprint for the AI industry.
On Friday local time, the 33rd Baron Investment Conference held a virtual fireside chat lasting approximately seventy minutes. The host was Ron Baron, founder of Baron Capital, with Elon Musk joining remotely.
During the interview, Musk delivered a series of blockbuster predictions: SpaceX is expected to surpass $100 billion in annualized revenue by December this year, with annualized recurring revenue potentially exceeding $500 billion by December next year; Terafab aims to produce one terawatt of chips annually, equivalent to 30 times next year's global AI chip output; Starlink computers co-designed with Nvidia are expected to begin reaching orbit in the third quarter of 2027; agentic internet traffic has already surpassed human traffic and could reach 1,000 times human traffic within five years.
SpaceX Revenue: From $100 Billion to $500 Billion in Just One Year
SpaceX was the most closely watched core subject of the interview. Musk explicitly provided two tiers of revenue guidance during the conversation:
"We are on track to achieve annualized revenue exceeding $100 billion in December. Obviously, December is very close, so I have good confidence in that. Unless some kind of natural disaster occurs, we will hit that target."
Going further, his expectations for next year were equally optimistic: "Then based on our plans for next year, annualized recurring revenue exceeding $500 billion in December next year seems quite likely. This is not a promise, it is just what I think will happen."
As for $1 trillion in annualized revenue, Musk said: "One trillion dollars is not impossible. It used to be impossible, but now it is no longer completely impossible. Maybe there is about a 2% probability. But what is interesting is that it is not 0%. That is the key. And that probability is increasing over time."
SpaceX generated approximately $15 billion to $16 billion in revenue last year. If the above trajectory materializes, it would mean the annualized revenue scale expanding by more than 30 times in less than two years.
Terafab: Producing 30 Times Global AI Chip Output Next Year
Musk elaborated during the interview on the Terafab (terawatt factory) plan being driven by xAI.
He explained the naming logic: "When I say terawatt factory, sometimes people think there are three possible meanings: one is terrifying 鈥?TERROR, then TERRA, which refers to Earth, but it is actually TERA, meaning terawatt. But one could say all three meanings apply."
In terms of scale, Terafab's goal is to produce chips corresponding to one terawatt per year. According to Musk's prediction, total global AI chip output next year will be approximately 35 to 45 gigawatts 鈥?meaning Terafab's single-factory capacity would be about 30 times the global annual output at that time.
Musk pointed out that there is enormous room for improvement in existing chip manufacturing efficiency, illustrating with one detail: "Right now, in a chip fab, a wafer spends more than half its time just on transport and waiting... Sometimes the time when atoms are actually usefully transformed is only about 20%."
His solution is to build a research fab in Austin, attempting to concentrate design, manufacturing, and testing under one roof. "I think we might be able to compress something that normally takes a year down to a month." The underlying logic: "For any given technology endeavor, the rate of progress depends on how many iterations you can do... As long as you can iterate an order of magnitude faster, roughly speaking, your technology progress will also be an order of magnitude faster."
Space Compute: First Batch of Starlink Computers Launching Next Year
Musk repeatedly positioned space compute as a strategic-level asset during the interview, with a clear logic chain.
In his view, ground-based compute expansion is constrained by electricity: U.S. average electricity consumption has just exceeded 500 gigawatts. "If you build 200 gigawatts of AI per year, that is equivalent to adding 40% to the entire U.S. power grid every year. That is simply not feasible."
The advantage of space lies in continuous sunlight and no atmospheric attenuation 鈥?solar energy utilization can be improved by an additional 30%, and virtually no large-scale energy storage is needed. "Once you have a very low cost per ton to orbit, the cost of intelligence in space will be much lower than on the ground."
Musk revealed that satellites already being launched are carrying Nvidia GPUs and Google TPUs for testing. "As far as we know, they are operating well in space." The first batch of Starlink computers co-designed with Nvidia is expected to begin launching in the third quarter of 2027, scaling to considerable volume in 2028 and reaching "very large scale" in 2029.
He further provided revenue calculations: "If we deploy 100 gigawatts of AI compute to space each year, and the value per gigawatt drops to even $10, that is $1 trillion a year. And it accumulates 鈥?every year you add another trillion dollars, because satellites do not evaporate." If the unit price is $50, that becomes $5 trillion per year.
Musk predicted that if one terawatt of compute is sent to space each year, it would be equivalent to sending digital superintelligence into space annually at twice the current average U.S. electricity consumption.
Starlink: 100-Fold Bandwidth Growth, Agent Traffic to Be 1,000 Times Human Traffic
Musk revealed that approximately 11,000 Starlink satellites are currently in orbit, primarily Version 2, with Version 3 satellites having just begun launching.
V3 has more than 10 times the capability of V2, and the number launched will also be 10 times greater. "Starlink's bandwidth in the next two to three years will be 100 times what it currently provides."
The key variable driving demand is AI agents.
Musk stated: "My understanding is that, in fact, agentic internet traffic has already surpassed human traffic, if I am reading Cloudflare's data correctly. And I believe Cloudflare's prediction that within five years, agentic AI traffic will be 1,000 times human traffic. And I think that even underestimates the scale of agentic AI traffic."
Data Centers: $35 Billion Investment, $50 Billion Return in One Year
Musk introduced the business logic of xAI's data centers: self-built speed is about 8 to 9 months, far faster than external contractors' two and a half years; investment is about $35 billion, below market levels.
"We will rent out the compute before our own Grok needs it, collecting about $50 billion a year. A $35 billion investment, getting back $50 billion a year."
He also revealed that xAI built an over-one-gigawatt internal power plant during the construction of the Memphis data center. "It took about one-fifth of the normal time to build such facilities 鈥?about a year to build, whereas it normally takes about five years." He also disclosed that he personally participated in cable laying, fiber optic wiring, and electrical installation.
Revenue growth from Q2 to Q3 this year exceeded 80%, with Q3 to Q4 expected to be approximately 70%.
Humanoid Robots and AI's Impact on the Economy
Musk estimated that digital superintelligence will add approximately 20% to 30% to the global economy, equivalent to about $25 trillion. "And this will happen very quickly," potentially doubling white-collar productivity.
On physical AI, he provided preliminary calculations for robots replacing labor: an efficient humanoid robot can approximately work 168 hours per week, with a minimum productivity expectation of about five times that of a human. "Rough estimate, about 20 million to 30 million robots could match U.S. goods and services output." In terms of business model, he leans toward leasing rather than selling.
Regarding OpenAI
Musk once again stated his position on OpenAI, speaking directly: "I did found OpenAI as an open-source nonprofit organization. Now it is a closed, maximum-profit entity, which is completely the opposite of what I originally founded."
He said he did not anticipate that OpenAI would one day convert to a for-profit company, and expressed confusion about the structure of holding commercial interests within a nonprofit organization. "I was confused at the time 鈥?how can you have a commercial interest in a nonprofit organization? That does not make sense to me."
Regarding Anthropic, he pointed out that it was founded by seven former OpenAI employees who left. "In effect, Anthropic is a spinoff of OpenAI."
Starship and Ultra-Rapid Reusability
Musk revealed that approximately one week after the interview, an attempt will be made to catch the returning Starship spacecraft using the tower's mechanical arm, known as "chopsticks," achieving full ultra-rapid reusability.
He explained the commercial logic of this system: "In principle, it could allow us to relaunch a rocket every hour 鈥?that giant booster, the Super Heavy booster." Starship is a fully reusable orbital-class launch system, fundamentally different from Falcon 9 鈥?Falcon 9's upper stage cannot be recovered. "It is equivalent to throwing away a medium-sized jet aircraft on every flight."
He predicted that in the extreme case, the cost per ton to orbit "could be lower than the shipping cost of air freight from California to Australia," a reduction of more than 100 times compared to the historical peak.
The following is the transcript of this interview (translated with AI assistance):
Ron Baron:
One image that came to mind when preparing for this conversation was E.T. E.T. crossing the moon, a boy riding a bicycle, carrying a communication device, wanting to phone home. This was obviously the precursor to Starlink. So I guess that did not work out.
Elon follows first principles: everything should be done the right way, not the fastest, quickest, or cheapest, but the best way, and everything should be re-examined. If someone invests on his behalf in a business, the judgment criterion is always: is this the best we can do? The fastest? The most effective? So I would say he is a builder. He is probably the only builder I know of in America who actually builds things, and not ordinary things 鈥?like 10 million square feet, 5 million, 20 million square feet buildings. Now he is considering or about to build a 100 million square foot building for Terap. No one has done this, no one has done anything like it.
The thing is, doing something like this yields economies of scale. You can procure and produce at much higher efficiency than smaller entities. So the question is: how do you do this? Why has no one tried to do something similar before? We are trying to bring silicon back to America. We need it, we need power, we need silicon.
No one other than you is doing this. Why?
Elon Musk:
That is actually not quite accurate, because TSMC and Samsung are both building fabs in the United States. TSMC is building a large number of fabs in Arizona, some of which I believe are already operational. Samsung has some operational fabs in Texas and has already built a fab for Tesla specifically for Tesla-designed AI5 and AI6 chips. But relative to Asia, U.S. chip manufacturing volume is very small. Even under the most optimistic assumptions, counting all announced fabs to be built in the United States, that is still only a small fraction 鈥?perhaps 10% of Asia, optimistically 10%. As most people know, Taiwan carries an extremely massive workload in chip production. TSMC is an amazing company, truly incredible, with the number, scale, and output of various chips it produces being unbelievable. Samsung is the same, and I think Micron does very well in memory. But the vast majority is manufactured in Asia.
My real concern is that there is always some kind of geopolitical risk in that region. I think that risk could materialize at some point. In that case, if we do not have all the pieces needed to manufacture chips at scale in the United States, we might have no chips at all, or very few 鈥?that 10% number. Looking further ahead, beyond any medium-term geopolitical bumps, even setting aside geopolitical crises, we are simply too slow at manufacturing chips.
If you want to reach, say, one terawatt of chip production per year, that is our goal, and that is why it is called Terap. Sometimes when people hear Terap, they think there are three possible meanings. One is terror, T-E-R-R-O-R, and people ask why you would want a Terap. There is also T-E-R-R-A, meaning Earth. But it is actually T-E-R-A, as in terawatt. One could say all three meanings apply. So it is terrifyingly large.
The rest of the world will manufacture about 20 gigawatts of AI chips this year, maybe 25 gigawatts. Next year, depending on whose forecast you believe, roughly 35 to 45 gigawatts. But if you want to do one terawatt, you need about 30 times next year's current global output.
Ron Baron:
Just for us to build.
Elon Musk:
Yes, we are building... we could also sell to others, but basically, what we are building in one factory is 30 times the current annual output. You would think this factory's output would be equivalent to world GDP, equivalent to the GDP of the 20th largest country. One factory.
Yes, it might even enter the top 10. Obviously, a plan like this would be crazy if it fails, and very bold if it succeeds. So it depends on the outcome. I think it will succeed at least to some degree.
Ron Baron:
Hopefully. This will cost a trillion dollars.
Elon Musk:
Over time, it could be more than that. But we will not need that money immediately. We will build in phases. We are studying ways to change the chip manufacturing process to improve chip production efficiency, just as we have done in many other areas. To give one example: right now, wafers in a chip fab spend more than half their time just on transport and waiting. That is, wafers spend more than half their time in the fab doing nothing 鈥?either moving or waiting for machines. I think there is great potential for improvement in chip manufacturing.
Ron Baron:
Why do these people not do it themselves? If their chips are just sitting idle half the time in their factories, what is stopping them from improving their own systems?
Elon Musk:
No. If you visit a fab, you will see wafers moving a great deal, moving rapidly everywhere, then waiting. But the time when atoms are being usefully transformed on the wafer is less than half. I would say sometimes it might be only 20% of the time that atoms on the wafer are being usefully changed. They do not want to try risky approaches in chip fabs because these facilities cost enormous capital. If there is a proven recipe, they simply do not want to risk something new. So the chip industry generally tends to repeat what was done last time, making only minor changes, not taking risks.
We are building a research laboratory at Gigafactory Texas in Austin, Texas. This lab will try many different chip manufacturing methods. Many will not succeed, but I am confident some will. We will then transfer these improvements to Terap itself, the production Terap, located in Grimes County, Texas, near Texas A&M. There we will scale up AI chip production. I think one interesting thing about the Austin research lab is that I want to see if we can do logic, memory, packaging, and lithography under one roof, so that we can design new chips, manufacture the entire chip under the same roof, test it, and advance with very fast iteration cycles. Something that normally takes a year, I think could be shortened to a month. For any technology endeavor, the rate of progress depends on how many iterations you can do and how much progress you can make between each iteration. If you can increase iteration speed, increase the number of iterations per unit time, and increase the progress between each iteration, you will accelerate progress far faster than anyone else. If you can iterate an order of magnitude faster, as a first approximation, your technology advancement will also be an order of magnitude faster.
Ron Baron:
Most enterprises have enormous inertia. Because if someone spends money and fails, that is a stain on them; if they succeed, and they want to compete for the top job, it just means they have to work harder. So success means working harder, and failure means it is over, losing the chance to compete for the top job. Elon has built a culture in his companies: always come to me with ideas, and if you think it will succeed, and I believe it truly will, we abandon what we did before and start over. Always do the best thing. One thing I find interesting about your culture is that you introduced me to Key Luck, who is the head of Intel and will now be interested in Terap. I spent some time with him, and he came to visit me. I said, what is it like working with Elon? He said, other people come to visit my facility in Oregon and leave after an hour and a half.
Elon came, sat there for four and a half hours, interviewing engineers one after another, questioning everything related to speed. Later I went to Texas to visit the car factory they were building at the time, and they introduced me to the facility head, the manager, a giant of a man. Of course, everyone looks like a giant to me, but he was truly very tall. He had previously been Ford's head of production. I said, what is the difference between Ford and Tesla? He said, at Ford, if I want to make a change, it has to go through six levels of approval.
Here, when I need a change, I just walk upstairs and Elon is there. So when the factory was being built, Elon was not some distant person. He was there, even had an Airstream, living in the Airstream. Do you know any other executive of a major company, possibly the largest company in the world, who would live in an Airstream to ensure that all problems during factory construction can be answered quickly? This is an amazing culture, amazing people, and the amazing people he recruits. The talent working there is brilliant. Getting a job there is harder than getting into Harvard. Of course, I do not know why people still want to go to Harvard now.
Elon Musk:
By the way, the Airstream is actually quite nice. Not that bad. Honestly, I kind of like my little Airstream. But it is parked right next to the power plant we built. This power plant generates over one gigawatt, possibly reaching one and a half to two gigawatts of internal generation in about one-fifth the time normally required to build such facilities. About a year to build, where it normally takes about five years. And it is not that someone has to come upstairs to find me. There is no upstairs. We do not have an upper floor at Colossus in Memphis. I sometimes walk the entire site twice a day, sometimes three times, sometimes four or five times. The entire site, the power plant and all the buildings.
I also insist on doing some work myself. I have pulled power cables as thick as a python, very heavy and long. I have done fiber optic wiring, done electrical switching, participated in power plant management software. I try to do a bit of everything so that I do not become numb to the suffering of the people who actually do this work, and so that I have some understanding of the difficulties and how to improve. So it is not that I sit in an office waiting for people to come to me. I walk the site and proactively ask them questions.
Ron Baron:
That is remarkable. Let us talk about data centers, which are in the news now. There is considerable resistance to data centers. But if you want AI, you must have data centers. Let us first talk about data centers on this planet. Elon believes the obstacle to data center construction will be whether there are enough chips and enough compute. So what he did is, many people... we have Grok, he is at the forefront with OpenAI, Anthropic, those people were his people, of course he no longer likes OpenAI, but...
Elon Musk:
I did indeed found OpenAI as an open-source nonprofit organization. Now it has become a closed, maximum-profit entity. It is actually the complete opposite of what I founded.
Ron Baron:
Right.
Elon Musk:
Anthropic was founded by seven people who left OpenAI because they did not trust Sam Altman, with good reason. They have said this publicly; it is not a secret. So in effect, Anthropic is a spinoff of OpenAI. Yes, I think I should have founded OpenAI as a regular company rather than a nonprofit. I thought it would be some kind of charity for the world. I did not expect it would one day become a closed-source for-profit entity.
Ron Baron:
When you left because you were disappointed with its direction, they gave you an ownership stake, and you said no, I do not want it, and then you left.
Elon Musk:
Yes, I was confused about how you can have a commercial interest in a nonprofit organization. That does not make sense to me. At the beginning I provided almost all of the company's capital. It was my idea, I named it, I recruited the key people, and I taught them everything I knew, and beyond that, nothing more.
Ron Baron:
Back to data centers. You think putting all AI into operation requires data centers, people want to build them, and normally building one takes about two and a half years, costing about $60 billion 鈥?$50 or $60 billion to build one gigawatt. Those who build for hyperscalers and lease out need to borrow money or obtain it somehow to get capital, must have long-term leases, at least five years. So what people pay in a five-year lease is about $12 billion a year on a $60 billion investment. Instead, we build them ourselves. We build because we have uses, because we are vertically integrated 鈥?for cars, robots, rockets, everything. Because we need them, we can take on the construction risk without needing to know there is someone else's contract. But as a result, we build them, and they are not... so we need them, sometimes four months, now it takes eight or nine months to build. They need two and a half years; we need eight or nine months. It costs them $60 billion; it costs us $35 billion. The demand for compute is so great that people who built data centers have them sitting idle because there is not enough power to turn them on. We will rent out compute until we need it for Grok ourselves, leasing to them at about $50 billion a year. So we get $50 billion in annual revenue from a $35 billion investment. So SpaceX's revenue is about to explode.
Elon Musk:
Yes, they are exploding. I think our Q2 to Q3 revenue growth this year exceeded 80%. I am not sure of the exact figure for Q3 to Q4, but probably about 70%. Q4 to Q1 next year might be similar. Bringing compute online at scale is extremely difficult.
I think there are many speculators out there who believe you just order some GPUs, hand a purchase order to some company to build a data center, and then sit back and collect cash. The reality is that it is extremely difficult. There is a meme I find quite amusing, called the Jose meme, you may have seen it: a guy digging a hole in the road, eight people watching him dig. Now in that meme, one person is actually working, and eight people are watching that person work. Unless you increase the number of Joses in this country, the amount of work completed will not increase. Adding more people who just send purchase orders to Jose will not cause more things to happen; only more Joses means more work. So I aspire to be Jose, and SpaceX wants to be Jose. We do the actual work. This is a big deal.
As the United States massively expands and builds, frankly, it is not a large influx of people doing actual physical work, but a large increase in people sending purchase orders. This means that the people doing actual work are extremely stretched. So there are a large number of partially completed data center projects across the country because the peanut butter is spread too thin to even taste. The difference is that at SpaceX, we are accustomed to doing extremely difficult physical work, like building giant rockets and massive launch sites. Tesla is also very good at large-scale construction. Probably the two best companies on Earth at physical hardware, besides China, are SpaceX and Tesla. I do not even know who the third would be.
I think even Tesla might be the best in the world, even including China, which is very, very formidable. Notably, the Tesla Model Y is the best-selling electric vehicle in China. Let me tell you, Chinese car companies are super impressive.
Ron Baron:
The guy from Ford who went there. He said he was shocked by the quality of Chinese cars.
Elon Musk:
Yes, partly... China is great. By the way, I love China. Going there is very inspiring.
Ron Baron:
We were talking about SpaceX's massive revenue growth. I remember it was $15 or $16 billion last year, and by the end of this year it will reach about $100 billion in annualized revenue. Then next year we think it will be between $500 billion and $1 trillion. So we go from $15 billion to about $100 billion annualized, to $500 billion to $1 trillion. And this is really just the beginning. The really cool thing about selling our data centers to competitors is that they are basically paying us $50 billion for something that cost us $35 billion, and we will have a steady stream of cash flow from renting out data centers we still own. So we got them for free, right?
Elon Musk:
Well, sort of. I would like to comment on a few of those points. We are heading toward over $100 billion in annualized revenue by December. Obviously December is very close, so I have good insight. Unless some kind of natural disaster occurs, we will hit that target. Based on our plans for next year, annualized recurring revenue exceeding $500 billion in December next year seems quite likely.
This is not a promise, just what I think will happen, my best guess. I just want to distinguish between promising to do something and what I think will happen. What I think will happen is $500 billion in ARR in December next year. I am just saying that $1 trillion is not impossible, meaning it used to be impossible, and now it is no longer impossible. Maybe about a 2% probability. But what is interesting is that it is not 0%. That is the interesting part. And the probability is increasing over time.
As my biological neural network's prediction of digital superintelligence's impact on the global economy, it is about 20% to 30% growth, and that does not even factor in robotics. I think about 20% to 30% growth, roughly speaking, is about $25 trillion in global economic addition. And this will happen very quickly.
Ron Baron:
That is a $130 trillion global economy. You think digital AI will add 25%, 30% to it?
Elon Musk:
20% to 30%. Yes, that is my guess.
Ron Baron:
What is your guess for how much it will add to the economy?
Elon Musk:
I think probably 20% to 30%. I think it could be higher. But it could double white-collar productivity.
Ron Baron:
That is just digital, not physical. Digital AI.
Elon Musk:
Physical is what we will dominate.
Ron Baron:
I think... yes.
Elon Musk:
The way to think about physical AI, that is robotics, is: any economy is the number of people in it multiplied by average productivity per person. If you have an extremely efficient humanoid robot that can work 24/7, 168 hours a week, no vacation, no bathroom breaks, no need to eat, and operates faster than a human, you can expect average weekly productivity of roughly five times a human, roughly speaking, because a person might work 40 hours a week, needs rest, and does not move as fast as a robot. So I say the minimum productivity expectation for a humanoid robot is five times a human. Then, for example, take the U.S. population 鈥?how many people are in school or retired, that is about half the population, about half are in school or retired; among working-age people, not everyone works or is in high-productivity employment. What remains, maybe about, I do not know, 120 million people in the U.S. have net positive goods and services output. Then you say, how many robots are needed to match that productivity? I think you would need about 20 to 30 million robots to match U.S. goods and services output.
Ron Baron:
And we will build those. Building a robot is cheaper than building a car.
Elon Musk:
Yes. Initially about the same, but over time much cheaper.
Ron Baron:
So robot cost $30,000, heading toward $20,000.
Elon Musk:
Yes.
Ron Baron:
Will we lease them or sell them? Do we already have ideas?
Elon Musk:
I think we will probably lease them.
Ron Baron:
Back to data centers. If people are really unhappy about data centers in their communities, since these are so profitable for us, why do we not just say, okay, your community's utility bills will remain unchanged, we will subsidize them, we will use that subsidy, and we will also give you $50 million for schools and hospitals. Do we have such a program, going to communities and saying, yes, we will help you in your community, give you more jobs? We really are a job-creating company. No one creates jobs like we do. Is there a way? I do not think we have done it, I do not think our industry has done a good job explaining that this is beneficial rather than harmful to people.
Elon Musk:
Actually, in the Memphis area locally, I think we are quite popular. Strangely, the popularity is there; the unpopularity is in places that do not have data centers, where people think data centers will have negative effects. But in Memphis, Memphis is booming because of SpaceX. Restaurants are full, hotels are full. We sometimes pay people twice what they earned before. Why do people not write about this? Why are there no stories about this?
Ron Baron:
Why is this not...
Elon Musk:
There are some stories... we also helped local schools, and the tax revenue the county and city receive is enormous. So I think this popularity thing may be a temporary concern, because in reality data centers bring enormous prosperity. If you talk to people in town, they all feel it, because their businesses are booming.
Ron Baron:
Then let us talk about data centers in space, the obvious solution, so that people do not have to worry about them in their communities if they do not want to. For us, putting data centers in space is cheaper. Each time we launch a new generation, they are more powerful than before. Is it not... everyone says, that is 10 years away. And Elon says, no, we are starting. Google has already tested with a satellite we operated for them. It has been working, so we will begin deploying them in '27 or '28. How soon will there be an impact?
Elon Musk:
Yes. People have already tested NVIDIA GPUs. There is a satellite with an Nvidia GPU up there that we launched; and a satellite with a Google TPU we recently launched. These are operating, and as far as I know, they are operating well in space. We expect to begin launching the first batch of StarMind computers, which we co-designed with Nvidia, possibly 鈥?I am just guessing 鈥?possibly in the third quarter of next year. We will start slowly, then possibly scale to significant production in '28, and then reach very, very large production in '29. To reach a scale of one terawatt per year, or even what I said, maybe 200 gigawatts per year, you really need to go to space, because there is not enough power on the ground to realistically reach those levels. As a reference, U.S. average electricity consumption is slightly above 500 gigawatts. U.S. peak power capacity is about 1.2 terawatts, but average electricity consumption is slightly above 500 gigawatts. So you can see, if you do, say, 200 gigawatts of AI per year, that is like adding 40% to the entire U.S. power grid every year, which is simply not feasible. When you put solar panels in space, because space is basically always sunny, solar panels can always face the sun directly, not at an angle to the sun, and there is no atmospheric attenuation of solar energy. So you get 30% more solar energy in space. Always sunny, no seasons. You can use a very small battery pack because you do not need to survive the night. The cost of intelligence in space is significantly lower than on the ground, once you have low cost per ton to orbit. Starship, because it will be the first fully reusable rocket, will reduce the cost of reaching space by 100 times or more. I think actually in the extreme case, probably within 2 to 3 years, the cost per ton to orbit will be lower than the cost of air freight from California to Australia.
Ron Baron:
NASA said the cost to orbit in 2010 was $18,500 per kilogram. That was because the rocket was only used once. Later we figured out Falcon 9, which could be reused, I think it was $2,500, then down to $1,500...
Elon Musk:
Now it will go down to $100 or $15 or $50. I think we will reach below $100 per kilogram to orbit.
It is important to understand how profound full reusability is. With Falcon 9, we have about 80% reusability. The booster comes back and lands. You may have seen videos of Falcon 9 booster landings, and then the fairing or nose cone containing the satellite 鈥?those fairing halves land in the ocean by parachute, and we re-fly those, but we do not re-fly the upper stage. We cannot recover Falcon 9's upper stage, and the upper stage costs about as much as a medium-sized jet. This means the minimum cost of each Falcon 9 flight is one medium-sized jet. That is still far better than anything else. But one could say, if every time you fly somewhere you have to throw away a medium-sized jet, there would not be much air travel. You would say, I think I will just drive. Now with Starship, the entire system is designed for reusability. Not just reusable, but extremely rapid reusability.
The reason for catching the booster and ship with the tower arm, which we call chopsticks. A bit like Mr. Miyagi and the Karate Kid, catching a fly, but just scaled up. The reason is that the rocket takes off from the launch pad. The rocket is placed on the launch pad by the tower arm, and then when it returns from space it is caught by the same tower arm and placed back on the launch pad. So you do not just have full reusability, you have extremely rapid reusability. You really cannot get faster reusability than landing at the takeoff site. This in principle will allow us to re-fly a rocket every hour, the giant booster, the Super Heavy booster.
Ron Baron:
10,000 launches per year.
Elon Musk:
Yes, eventually possibly more, more than one launch per hour. When I was young, people would line up to watch every rocket launch, which were very rare. And we could do one per hour, and it would not even make the newspapers anymore.
Ron Baron:
That would be quite... you really want to watch this one. This one is cool.
Elon Musk:
It is the largest flying object ever, and the highest thrust object of any kind ever manufactured by humans. This thing weighs over 5,000 tons at takeoff, going straight up, fast. Then that giant thing comes back, comes back from space, the booster is coming down about 5 minutes after takeoff, anyway. Hopefully in a good way, but either way it is coming down. So excitement is guaranteed. Success is not always guaranteed, but excitement certainly is.
Ron Baron:
When you watch it come back and land at the takeoff site, it is very cool. Chopsticks.
Elon Musk:
It looks crazy. When I first proposed this to the team, they looked at me like I was crazy, but I said, look, I think this can be done. There is nothing in physics preventing it from happening. The only thing I can think of that would be slightly better than catching with the tower arm is landing directly on the launch pad. But if something goes wrong, you would damage the launch pad, which would require a lot of work to repair. So catching with the arm is slightly better, because if something goes wrong, it does not necessarily destroy the launch pad. But that is just a few minutes of difference between catching with the arm and landing directly on the takeoff pad. Now what is much harder than the booster is catching the ship. The ship still has to do many things that have never been done before. No one has made a truly reusable orbital heat shield. The space shuttle's heat shield needed 9 months of repair between flights. Even though the space shuttle orbiter partially came back, they threw away the main fuel tank on every flight, which was actually the primary body structure, the primary structure that everything was attached to, but they would recover the orbiter, and the heat shield needed 9 months of refurbishment before it could fly again. So no one has made a fully reusable orbital heat shield. This is an extremely difficult thing.
This is something we have been working on for a decade. It is too hard. I think it is one of the most difficult technical problems humans have attempted to solve. But we have an extremely talented team that has worked incredibly hard for a long time, and we have been able to iterate. Back to iteration. We have been able to do many iterations, whereas the space shuttle actually could not iterate. They always had people on the space shuttle, so they could not take any risks, because someone could die. With Starship, we take many risks to explore the physics needed to achieve full reusability of the heat shield and the entire vehicle. We have actually achieved precise soft landings of the ship at sea multiple times, and if there were a tower there, the ship would be caught. So this is perhaps an interesting point: if the ship thinks it is landing with a simulated tower, and if there were a tower there, it would be caught. Now we want to be very sure that when we bring it back to the actual tower, everything will work properly, so that we do not smash the tower or shower debris on people's heads, which would certainly affect our popularity. So we want to be highly confident in the ship's reliability. I think we have finally reached the point where we think it is worth trying with Ship 42, hopefully that is prophetic. 42 is a special number, and we can try to catch with the tower. That is what we will do in about a week.
Ron Baron:
We are all cheering for it. Starship is the big unlock here. By the way, I was looking at the young man in charge of Starship operations, and he was speaking in front of an audience wearing sport jackets, suits, and white shirts, and he came up wearing engineer boots, blue jeans, and a plaid shirt, unshaven, with long hair. He started speaking, saying F this, F that, F this. I looked at him and said, oh my God, that was me at 40. Really good person, rides motorcycles. So this is the big unlock. Everything Elon does may not make sense to you when you first hear it, and then you say, oh wow, I get it, he may have been thinking about that all along.
When you think about reusable rockets, the cost of reaching space was so high that the only ones who could go to space, at $18,000, were governments protecting their countries, or maybe a television company spending $2 billion on a satellite so it could broadcast to one place and no one else. Bring the cost down so low, and suddenly you make Starlink possible. Starlink now has 10,000 satellites in orbit, heading toward 20,000, which is what I used to think; now he says no, it will not be 20,000, it will be 100,000. Why 100,000? Because in the past we thought the revenue he would generate from Starlink would come from consumers, governments, and enterprises. And now he says agentic will be several times what we previously thought, and every time you...
Elon Musk:
Oh yes. My understanding is that, in fact, agentic internet traffic has already surpassed human traffic, if I am reading Cloudflare's data correctly. I believe Cloudflare predicts that within 5 years, traffic will be a thousand times human traffic. If anything, I think that underestimates the amount of agentic AI traffic.
Ron Baron:
A thousand times in how many years? Five.
Elon Musk:
Yes.
Ron Baron:
We have always thought Starlink was a business that could potentially reach $1 trillion in revenue within 10 years, $700 or $800 billion in EBITDA, worth $14 or $15 trillion. The entire SpaceX is now worth less than $2 trillion. So that is an opportunity. But now I think we may actually be underestimating how much SpaceX and Starlink could be worth. Because you are adding so many satellites, and now we have space centers, data centers.
Elon Musk:
Starlink revenue is growing very rapidly by any normal business standard. You are roughly doubling revenue annually at a large scale. I remember that used to be a big deal. Not long ago, doubling revenue every year with high gross margins and positive cash flow would normally be an impressive thing. Although I have to say, with AI, it really surpasses that. With AI, you have the potential for order-of-magnitude growth per year. I mean, if you think about the things I just said, because I said a bunch of things that sound extremely crazy, but I think we will get there. Sometimes I am a bit optimistic on timelines, but I always deliver in the end.
And I am not always late. I want to say this because the media always reports when I am late, but they never report when I am early. Actually I am often early. But the media does not write positive articles; they just ignore when I am early and only print when I am late. So it looks like I am always late, but actually that is not the case. My predictions, I aim for my predictions to be at the 50th percentile, meaning they are basically right about half the time in terms of timeline, and things roughly work out that way. We now have 11,000 satellites up there, primarily Version 2 satellites.
We have just begun launching Version 3 satellites. Version 3 satellites have more than 10 times the capability of Version 2 satellites, and 10 times the number will be launched. So the amount of bandwidth Starlink will generate, I do not know, within two or three years, will be 100 times the bandwidth it currently provides. So, you say revenue per gigabit per second or something like that will be high. Then you say how many gigabits you transmit, and that is a very large number. With the enormous demand from AI and robots far exceeding human demand, literally a thousand times. And just the initial communication revenue alone will be astonishingly high. But if you look at the AI side, if we say 100 gigawatts of AI compute to space per year, let us say what is the value per gigawatt of compute?
Even if it drops to $10, that would be an additional $1 trillion per year. Cumulative, every year it would be another trillion dollars, because satellites do not evaporate. So it is like you are stacking a trillion dollars of annual revenue every year. Every five years... $10, if you say $50, you would have five trillion. So how do telecom companies meet the demand for services with their existing infrastructure? And if they own spectrum, for example, spectrum is a public good, why would the government allow this to continue when the capacity needed far exceeds what they could possibly provide?
Ron Baron:
Yes. Hmm... should we skip this question?
Elon Musk:
I think SpaceX with the spectrum it has already acquired will be able to go far. But at some point, if we can use spectrum better than, say, a traditional telecom company, we could simply acquire that traditional telecom company and use that spectrum more efficiently, 10 times more efficiently.
Ron Baron:
We have been thinking about data centers, so we are thinking about Starlink eventually at 10,000 to 100,000 satellites, and every 5 years when we replace existing satellites, we replace the ones we are de-orbiting with satellites that have 10 times the capability. So those are the enormous amounts of capacity we will provide.
Elon Musk:
Yes, we still see opportunities for order-of-magnitude improvement in satellite capability. I think Starlink Version 4 will have 10 times the capability of Version 3, especially with long-distance laser links. Our current laser links are about 100 gigabits per second. Our next-generation laser links will be 10 terabits per second. That is 100 times better. And then we will put 10 of them on. So you know...
Ron Baron:
We are a bit short on time, but a few more things. When people worry about data centers and AI, people in China are not worried about having too many data centers and too much electricity. In China, they have about three times our electricity. And the electricity they add each year, 300 gigawatts, compared to our total of 550 gigawatts. They add 300 per year.
Elon Musk:
I am not sure that is a steady-state electricity figure, but there is certainly growth, and I do not think it is that high. I think it is slightly lower than that. But it is an enormous number. China now, I think, is about three times U.S. electricity output, and the trend is toward four times, at least reaching four times total U.S. electricity output. Industrial capacity as a first approximation is proportional to electricity output. Then chips, they are behind...
Elon Musk:
And they are advancing quickly. Yes.
Ron Baron:
A few years ago, a SpaceX executive told me she thought the only thing stopping my grandchildren from speaking Mandarin was SpaceX. Do you believe that?
Elon Musk:
I think there is some merit to learning Mandarin, anyway. It is a very interesting language. There we can bring hundreds of gigawatts per year, and possibly one terawatt per year. But of course that is a very large number, that is saying the digital superintelligence we produce each year is twice the current average U.S. electricity usage. U.S. average electricity usage is 500 gigawatts, one terawatt is 1,000 gigawatts. If we do one terawatt per year, that would be sending twice the current average U.S. electricity usage into space each year in the form of digital superintelligence. These are basically astonishing numbers.
Ron Baron: Two more questions. Why do we need so much electricity? Why do we need so many chips? Where will the demand for that kind of intelligence, that kind of electricity and chips, come from?
Elon Musk:
The demand will be 鈥?this sounds a bit circular 鈥?but the demand for intelligence, for digital superintelligence, will come from digital superintelligence.
Ron Baron:
That is for medical purposes, that is for... what does that mean?
Elon Musk:
It means superintelligence will far exceed all human needs. There will be universal high income, universal extremely high income. Universal medical care, better than the best medical care available on Earth today. You could wish for a castle or anything you want. Digital superintelligence could manufacture it for you as a side project. If you zoom out a lot, as civilization progresses, you can think of a civilization's progress in terms of how much energy that civilization utilizes. Currently, to show you how much room for growth there is in our solar system: even if we increase the energy civilization utilizes by a million times, we would still be using less than one-millionth of the sun's energy. That is a lot of room for growth.
You would obviously completely satisfy anything anyone on Earth could possibly wish for. At that point, it really is digital superintelligence serving digital superintelligence. Our economy will seem quaint, our monetary system will look like trading shells for currency. That is what our monetary system will look like to superintelligence. Technically that once happened. Money used to be shiny metal that people gave each other. I give you this shiny silver metal, then you have shiny gold metal, and you have to move these shiny chunks of metal from place to place to trade, which frankly was really cumbersome and easy to rob, if you had gold or silver coins. But that used to be how people paid for things. Buying things over long distances must have been extremely difficult, because you had to transport all your silver and gold across the ocean, the ship sank, and you did not even know what happened. It is not like you could call them on a mobile phone or something.
Ron Baron:
Obviously it would be fun to talk with you forever, but I am out of time, so I just have one last question. When you are 83 like me, how do you want people to see you? You are working these crazy hours, you are earning, no one would work like you just to accumulate money, I think. You think anything you have will eventually be heavily taxed, but you have a lot of money that will eventually go into a foundation, your foundation. How do you want people to see you? When you are 83, how do you want people to see you?
Elon Musk:
I want... there to be cognition and reality. I want reality, regardless of what the cognition is, but I want the reality to be that I made a major contribution to the advancement of civilization. That, you know, I... yes.
Ron Baron:
Thank you very much for coming here today. Thank you.
Comments