Elon Musk on AGI Timeline, US vs China, Job Markets, Clean Energy & Humanoid Robots | 220

2h 52m 2M views 6 Jan 2026

Is it worth your time?

You get a detailed account of how rapid advancements in artificial intelligence, energy infrastructure, and robotics will transform global economics and industrial production.

The conversation examines how accelerating artificial intelligence, humanoid robotics, and massive energy scaling will reshape human labor, economics, and space infrastructure. It addresses the practical hurdles of building the power and hardware systems required for an abundant future.

AI timeline and scaling. Musk expects AGI as early as 2026 and predicts that AI will exceed the intelligence of all humans combined by 2030. He argues that algorithmic improvements will yield exponential leaps in processing density. He points out that shifting inference to 4-bit data will unlock massive performance gains.

Energy and data centers. Musk warns that power generation and cooling are the primary bottlenecks limiting the deployment of new GPUs. To bypass grid delays, his company built a gigawatt training cluster using natural gas turbines and battery packs. He notes that manufacturing lines use a fraction of the power required for AI training.

Humanoid robots and labor. Musk predicts that humanoid robots will replace digital tasks before physical ones, eventually performing complex jobs like surgery better than human specialists. He warns that this technological shift will cause significant disruption and necessitate economic solutions like universal high income. He notes that Optimus will scale rapidly through compounding improvements in software, hardware, and electromechanical dexterity.

Starship and space infrastructure. Musk argues that fully reusable rockets are the tipping point that will make space launch costs plummet. He explains that high launch rates will enable the deployment of space-based solar power satellites and orbital data centers. He targets thousands of annual flights to transport a million tons of payload to orbit.

Navigating the Future of AI and Robotics

Your relentless optimism is always a breath of fresh air. I'm increasingly thinking that you are right about abundance for all.

Right now, I'm putting a lot of time into chips with some AI assistance. It would be nice if we could just hand it off to the AI soon enough. I tried to do some circuit design with AI a couple of weeks ago. If you took a photo and submitted it to Grok, it could probably tell you if there's something wrong with the circuit. 4.2 has not been released yet, but five is coming in Q1. If you upload an image into Grok, it does quite a good job.

If I take a picture of you, it's going to say you're a flawed circuit. I have to remember to update the Grok app because we update it so frequently. I asked Grok to roast me, then I asked it to roast you, and I spit out my coffee. I asked, "Does Elon know what you say about him?" and she goes, "What is he going to do about it?"

I just took a photo of you to see what it is. This man is Peter Diamandis, the host of the podcast Moonshots. By the way, Graedia is awesome.

The Promise of Abundance and Optimism

I tried to update my Wikipedia page for years, and it knows me. It's pretty damn good, though it has to be read by you.

I want to answer the question: can AI and tech help save America and the world? I want to give people listening a dose of optimism. A Pew survey from mid-December said 45% of Americans would rather live in the past and only 14% would rather live in the future, which is insane to me. Hollywood has shown us killer AIs and rogue robots.

You pinned on X that the future will be amazing, with AI and robots enabling sustainable abundance. From a first principle standpoint, what is the rationale for optimism? How do we head towards Star Trek and Roddenberry, not Terminator and Cameron? How do we move towards universal high income instead of social unrest?

Be careful what you wish for. If you actually get all the stuff you want, is that the future you want? Your job won't matter if you're living an unchallenged life. If we become couch potatoes in a Wall-E future, that does not go well for humans. People haven't historically been very good at creating their own challenges in the absence of others, though I think Elon does a damn good job.

After AI and robots, is there another thing? I guess there's conquering the universe. I'm more optimistic than most people. If you reframe things as a progress bar towards a Kardashev Type 2 scale civilization—the aspiration of capturing all the energy from the sun's output, or even a humbler aspiration than that.

Energy: The Key to a Sustainable Future

If our goal is to harness a millionth of the sun's energy, that would be more than a thousand times the energy possibly produced on Earth, as only about a half a billionth reaches us. We are very far from even a billionth; reaching a millionth or a thousandth is an enormous challenge, like escaping a gravity well. Energy is the inner loop for everything, and I think the future currency will essentially be wattage—the ability to control harnessed energy and turn power into work, whether intelligence or matter manipulation.

Your next big project is energy. You go back to your solar system, then expand to a galaxy level, like Star Trek, where there isn't even a horizon because you're not on a planet. Think galaxy mind. We are in an 11.5 million square foot building here, but the civilizational challenge is how to climb those orders of magnitude in harnessed energy. In the long run, we will have abundance beyond what people can imagine; AI and robots will saturate all human desire.

Nanotechnology takes it further. We are already doing atomic level assembly for circuits at two or three nanometers, which is only four or five silicon atoms per nanometer. While "two nanometer" is often a marketing number, you still need close to atomic level precision. I think modern fabs are getting clean rooms wrong. I'll bet Tesla will have a 2nm fab where the air handling is so good I could eat a cheeseburger and smoke a cigar inside.

The Role of Education in a Changing World

Fabs maintain wafer isolation by default, transporting wafers in boxes of pure nitrogen gas under a slight positive pressure. It is essentially impossible for anything combusting to live without oxygen; you can kill bugs just by putting a nitrogen blanket on plants.

Energy is the foundation of everything you are building right now. Compared to the sun, all other energy sources are like cavemen throwing twigs into a fire. The sun is over 99.8% of all mass in the solar system, while Jupiter is around 0.1%. Even if you burnt Jupiter and teleported three more Jupiters into our solar system and burnt them too, the energy produced by the sun would still round up to 100%. We have a giant free fusion reactor 93 million miles away; creating little fusion reactors would be like having a tiny ice cube maker in the Antarctic with three-kilometer-high glaciers right next to you.

Regarding the Memphis data center timeline, you aren't going to pull 10 gigawatts out of Memphis—maybe two or three. Ten gigawatts is Toyland. There are 100 megawatts right outside the door here, which is massive; all these manufacturing lines combined use less energy than that. Cortex 1 was the third largest training cluster in the world, and Cortex 2, which will be half a gigawatt and operational middle of next year, is being built out.

China has done an incredible job on solar. I believe their production capacity is around 1,500 gigawatts per year. They put in 500 terawatt hours in the last year, 70% of which was solar, and they are just scaling. The higher the energy production of a country, the higher its GDP. I think we should scale solar substantially in the US, as Tesla and SpaceX are doing.

I see a path to 100 gigawatts a year of space solar, AI-powered satellites. That would be 500,000 Starlink V3s launched over 8,000 Starship flights. 10,000 flights a year is a reasonable number, though very high compared to the rest of the rocket industry. With a million tons of payload to orbit per year at 100 kilowatts per ton, that is 100 gigawatts of solar-powered AI satellites per year.

There is a path to a terawatt or even 100 terawatts a year if you make those AI satellites on the moon using a mass driver, the Gerard K. O'Neal approach. As Robert Heinlein wrote, the moon has no atmosphere and an escape velocity of around 2,500 meters per second, so a mass driver works very well. We are effectively building a Dyson-ish swarm around the Earth. Your payload to orbit drops by about 30% compared to mid-inclination, like 70°.

We need a way to get defunct satellites down. Once you can put a million tons of satellites into orbit, you can also bring them down or collect them into a fixed location to reuse them. Given the amount of intelligence we're talking about, the resource level will be so high that this will be a solved problem.

Data centers will not be in low Earth orbit; they will be much higher to stay constantly in the sun. Around 1,200 kilometers or synchronous orbit will give you constant sunlight. Clutter will only be in low Earth orbit, as atmospheric drag brings back anything below around 700 or 800 kilometers. For Starlink, there is a dual benefit to being as low as possible because latency is lower and beams are tighter. Starlink 3 will be around 330 to 350 km, which is quite a lot of drag. We have basically rebuilt the internet in space with laser links.

There are 9,000 satellites up there right now. I just got back from Rome, but look at the chart showing the number of billion-dollar startups in the US versus Europe, and data centers too. No one was talking about orbital data centers six months ago, then suddenly Sundar's on it, it went viral on X, and everybody's got their own orbital data center.

I suggested to Peter that you update the math on launch costs because it's a tipping point. Whether Starship is $100 or $10 per kilogram, we're going to do a million tons a year to orbit well below $100 a kilogram, which will move data centers to orbit. If you have a rocket that is fully and rapidly reusable like an aircraft—which I think is at the limit of human intelligence, but we're doing it with Starship—you can do the math. This has been the holy grail of the aerospace industry forever. When Falcon 9 first reused its first stage, traditional aerospace didn't believe it could fly and reuse, yet you can see it land at Cape Canaveral and take off again.

The leap to the launch cost requires more faith, but Starship is the tipping point. Somewhere in the timeline between Twitter becoming X, it went from speculative to no doubt. People's interest in orbital data centers ties directly to credibility and the cost of energy in their hometowns. There are doomer conversations that data centers will drive the CPI up, but the best way to increase the energy output per year of the US or any country is batteries.

The peak power output of the US is around 1.1 terawatts, but average power usage is only half a terawatt. If you buffer the energy—charging batteries at night and discharging during the day—you can double the energy throughput of the US without incremental capital expenditures or building new power plants.

Energy output per year can double with batteries. Tesla makes them; I went on stage and presented the thing, and it's on the internet. China seems to listen to everything I say and does it, or they're doing it independently, but they're certainly making massive battery packs, vast numbers of electric cars, and vast amounts of solar. These are all things I said we should do here.

When I'm piloting and look down over Santa Monica and LA, zero roofs have solar on them. It's not essential to have them on a roof, but it's a convenient place. Tesla makes a solar roof, which is the only one that isn't ugly; ours actually looks beautiful. But if you want to do solar at scale, you just need more surface area. We have vast empty deserts in Africa and America. If you fly across country, for a large portion of the time you look down at bleak desert that looks essentially like Mars—scorched rock where barely a lizard is alive. If we put solar where we currently have scorched rock, it will be a quality of life improvement for the few creatures in this difficult environment; they'll think, "Thank god, some shade finally."

To materially affect quality of life, you need to capture and store a couple hundred gigawatts. You could put the data center locally there. In an abundant world five years from now with massive amounts of compute and universal high income, I don't know how much more energy we'll need—maybe 30, 40, or 50% more—unless we want to move mountains around to make a ski mountain in the backyard. I think the vast majority of energy consumption will go into compute. Manufacturing every one of these cars coming out at the rate of one every minute or two uses less energy than the data center training the cars to self-drive.

The sun generates vast amounts of energy for free that just goes into space, so we'll end up trying to capture a thousandth or millionth of it. We're probably at 1% of Kardashev level one; I don't think we're 10%. We're roughly using 1% of the energy we could use on Earth. The bottom line from first principles thinking is that there's a lot of energy out there, and we have it in the US.

We have the energy on the planet and the tech to capture it is here and improving every year. There will be a large forcing function to harness more energy, but we're not going to run out of it.

In 2010, 75% of Americans said it's important to go to college; that number is now down to 35%, yet tuition has increased 900% since 1983. Administrative expenses at universities have gotten out of control. I saw a stat that there's one administrator for every two students at Brown, which seems high.

I went to college in Canada for a couple of years at Queens University. I had Canadian citizenship through my mom, who was born in Canada, though my grandfather was American and for some reason my mom couldn't get US citizenship. I didn't have any money, a giant social network, or huge wealth, so I could only go to a Canadian university at first. I arrived in Montreal at age 17 with around $2,500 in Canadian travelers checks, one bag of books, and one bag of clothes. After Queens, I did a dual degree in physics and economics at the University of Pennsylvania and graduated from Wharton.

I was going to do a PhD at Stanford working on energy storage technologies for electric vehicles—fundamentally material science. My idea was to create a capacitor with enough energy density to get high range in an electric car. You could add a leaf to the tree of knowledge, but it's not necessarily a useful leaf. An enormous fraction of great entrepreneurs drop out of grad school or undergrad because the curriculum is nowhere near caught up to what's actually going on in technology. It's unclear to me why someone would be in college right now unless they want the social experience. If we want to create an Elon Musk factory of people who start with very little but drive breakthroughs, we need curiosity about the nature of the universe.

I found school quite painful, boring, and in South Africa, very violent. It was like the book Ender's Game IRL, but not as fun. Your goal was escape.

Health, Longevity, and the Future of Humanity

I suppose you need a bit of hardship to create artificial adversity. That's a Warren Buffett topic, but it's not easy because if you love your kids, you don't want to do that. I had a lot of adversity, which was probably good—a "what doesn't kill you makes you stronger" type of thing. At least I didn't lose a limb; I think what doesn't maim you makes you stronger, and I'm good at keeping ten fingers.

For the last five years, I've helped teach Foundations of AI Ventures at MIT. Every year, student surveys show their desire to start a company increases, now up to 80%. They want to be part of a founding team. When Peter and I were at MIT, I guess it was maybe 10%. A few days into the quarter, I called Bill Nicks, the head of the material science department, and asked to put my studies on deferment. Last year, he sent me a letter saying all my predictions about lithium-ion batteries came true. Stanford has also told me several times I can come back for free.

Every time an Iron Man movie came out, the desire to start a company notched up another 10% because everybody wanted to be Tony Stark. I didn't know until today that the modern Tony Stark is modeled on Elon. I always thought the original character was Charles Stark Draper's education and scientific endeavors married with Howard Hughes's ambition, but then Robert Downey Jr. changed that. This is a Groipedia fact. If Grok gets good enough, we're going to call it Encyclopedia Galactica. 42.

You just came back from an announcement with President Blly, who is a friend and an amazing visionary. It was a nuclear option; he went to the graves of all the gang members and destroyed them, saying, "Your memory will not be remembered in this nation."

You have to be badass to take on all the knocker gangs and win and live. He's got a great guard at his palace there.

We could use Grok for personalized education, creating a kids-friendly version. AI can be an individualized teacher that is infinitely patient, answers all your questions, and makes learning more interesting; you could probably gamify and incentivize it.

People should think of the brain as a biological computer, a neural net with a number of neurons and neural efficiency. You can't tune any arbitrary kid into Einstein, Shakespeare, or Newton because that's not realistic; Einstein had an outstanding meat computer, and you need an exceptional meat computer for that.

When people say we need to solve education in the United States because it's fundamentally broken, I think the old social contract—do well in high school, get into a good college, get a degree, and get a job—is what's really broken. It's about being an entrepreneur and finding a problem and solving it. In El Salvador and other places, we'll have individualized teachers that are much better. You could go to a school to hang out with other kids, but you don't need to; you could do it on your phone at home. Education is now a social experience. My kids in college recognize they can learn just as much independently.

Being on your own, you learn how to lead or defend yourself. If you join the workforce as a 19-year-old, you're with a bunch of old people. If you're doing engineering with middle-aged dudes, do you really want that, or do you want to hang out where there are girls your age?

The US is ranked number one in health expenses worldwide and 70th in health span. We are really 70th. If people get really fat, their health gets bad.

AI's Impact on Labor and Employment

First of all, I wasn't eating a lot of doughnuts; I was trying to have 0.4 of a donut, which rounds down to zero.

I suggested we push to get people to 120 or 150, but Elon said people shouldn't live that long and that we need people to die so others change their minds. My response was that the head of GM didn't have to die for Tesla to come along, and Lockheed, Northrop, and Boeing didn't have to go away; in a meritocracy, the better ideas will dominate.

David Sinclair is about to start epigenetic reprogramming trials in humans using an injection of an adeno-associated virus with the three Yamanaka factors, which has worked in animals and non-human primates. We have a $101 million Healthspan XPRIZE with 730 teams working on reversing the age of your brain, immune system, and muscle by 20 years. When the primary funder found out the prize was 100 million, Chip asked if we could make it bigger, so I said if you put an extra million in, we'll make it 101 million.

Folks like Dario Amodei are predicting a doubling of the human lifespan in the next 10 years. I don't know about doubling, but a significant increase is likely, which is easily escape velocity or effective age. My comment to Elon back then, when he was in his late 40s, was that I just want things to not hurt. It seems like it's a matter of when, not if, your back hurts or you get arthritis. I want to be able to sleep through the night without going to the bathroom.

Universal High Income: A New Economic Paradigm

Admittedly, having to wear adult diapers is a bummer. One sign that a country is not on the right path is when adult diapers exceed baby diapers; Japan passed that point many years ago.

With a one-third replacement rate, three generations from now they will be 127th, or 3% of their current size. Longevity will be an important part of the solution to massive underpopulation. If you increased the productive life of most Americans by just a few years, you'd flip the entire economics here.

I would like to double my lifespan to at least 120 or 150. As my partner Alex Weer Gross says, we're going to speedrun Star Trek. At a minimum, your kids will have infinite life expectancy because they are past longevity escape velocity and the idea of 50 years of AI improvement. I've long thought that longevity or semi-mortality is an extremely solvable problem; I don't think it's a particularly hard problem.

Navigating the Singularity and AI's Acceleration

Nobody has an old left arm and a young right arm; you are programmed to die. If you change that program, you will live longer. Bowhead whales can live for 200 years and Greenland sharks for 500. It is either a hardware or software problem, and we will have the tech to solve it. The important thing is not to die from something stupid before the solutions come.

I co-created a company called Fountain Life with Tony Robbins, Bob Hurry, and Bill Cap. We do a 200 gigabyte upload of you—full genome, all imaging, everything. President Bly and the first lady called it an amazing 10 out of 10 experience. I don't want you to pull a Steve Jobs and kick the bucket because of something they didn't know. I recently submitted an MRI to Grok; neither the doctors nor Grok found anything wrong, but that is a fraction of the information. Your full genome, microbiome, and metabolism are already in the works.

There will eventually be more digital intelligence than all human intelligence combined and more humanoid robots than humans. Assuming a benign, Roddenberry-style Star Trek scenario rather than a Cameron situation, robots will do whatever you want, including all blue-collar labor and data center work. White-collar labor will be the first to go because anything involving digital tasks—tapping keys and moving a mouse—can be replaced before you need humanoid robots to shape atoms. If you are just changing bits of information, that work goes first.

The Role of AI in Healthcare and Surgery

Even at its current state, AI is close to replacing half of all white-collar jobs, including education. Anything involving information, short of shaping atoms, can be done by AI. However, there is a lot of inertia. A company that makes more use of AI must compete with one that makes less to create a forcing function for increased use; otherwise, companies with humans doing what AI can do will continue to exist.

Being a computer used to be a job. You can find pictures online of skyscrapers full of people, mostly women, copying from ledger to ledger, calculating bank account interest or science experiments. Now, one laptop with a spreadsheet can outperform a skyscraper of several hundred human computers. If even a few cells in that spreadsheet were done manually, you could not compete with one that was entirely computerized. That one manual cell would be the most annoying, getting it wrong a bunch of the time. This means companies that are entirely AI will demolish companies that are not.

We are at peak doom for people worried about their jobs. The solution is UHI, where everyone can have whatever they want. My concern isn't the long run, but the next three to seven years. The transition will be bumpy because humans don't like simultaneous radical change, social unrest, and immense prosperity where things get very cheap and you can buy all the Cybertrucks you want.

Frankly, if this doesn't happen, we'd go bankrupt as a country. The national debt is enormous; the interest, which is growing along with the deficit, exceeds the military budget plus either Medicare or Medicaid—about one trillion dollars. Without AI and robots, we are headed for economic doom, compounded by competitive pressure from China. AI and exponential tech can save America and the world, but this is not about taxation and redistribution. I wonder if instead of universal high income, we should have universal high stuff: the UHSS.

Ethics and AI: Programming Values into Machines

I think it is better to be an optimist and wrong than a pessimist and right. It's clear we don't have any system right now to make this go well, but AI is a critical part of making it go well. At some point, Grok or one of the big four AI machines will be addressing this exact topic. There is no velocity knob or on-off switch; it is coming and accelerating. I call AI and robotics the supersonic tsunami.

We are basically demonetizing everything. Labor becomes the cost of capex and electricity, and AI is intelligence available at a dimminimous price. Things get down to the basic cost of materials and electricity. When we say universal high income, it sounds like tax and redistribute, but that's not the case. My best guess is that prices will drop as the efficiency of production or the provision of services increases. Prices in dollar terms are the ratio between the output of goods and services and the money supply; if output increases faster than the money supply, you have deflation. The output of goods and services will grow faster than the money supply, and I predict governments will push to increase the money supply even faster. They won't be able to waste the money fast enough, which is saying something for Argentina.

Productivity is improving dramatically. I think we may see high double-digit output of goods and services, though we must be careful about how economists measure things. My favorite economist joke is about two economists walking in the forest who come across a pile of shit. One says, "I'll pay you 100 bucks to eat a pile of shit." They come across another pile, and the other guy says, "Okay, I'll give you a hundred bucks to eat a pile of shit." After he gives him a hundred bucks, the guy says, "Wait a second, we increased the economy by $200."

Profitability of companies will go through the roof. If we hit massive productivity and profitability because we're dividing by zero—since the cost of labor and intelligence has gone to nothing while we produce products and services faster and faster—someone needs to have the capital to buy it. Whether that is taxed by the government and redistributed as UHI or UBI is the question. One side recommendation I have is don't worry about squirreling money away for retirement in 10 or 20 years.

The Future of Space Exploration and AI's Role

The services will be there to support you; you'll have the home, the healthcare, and the entertainment. The way this unfolds is fundamentally impossible to predict because of the accelerating timeline and AI self-improvement. It's called singularity for a reason. I don't know what happens after the event horizon; you can never see past the black hole.

Ray has a singularity out way too far. We are in the midst of the singularity right now. You're at the top of the roller coaster and about to go, knowing it's going to be a lot of G's when you hit it. I think we'll hit AGI next year in '26, and I'm confident that by 2030 AI will exceed the intelligence of all humans combined. That's pessimistic; if you hit AGI next year, the time to achieve self-improvements on the order of 1,000 to 10,000x through algorithmic improvements is very short.

Most people in the AI community don't yet understand that the intelligence density potential is vastly greater than what we're currently experiencing. I think we're off by two orders of magnitude in terms of achievable intelligence density per gigabyte, per gigawatt of energy, or per file size on your laptop drives. That's just algorithmic improvement on the same computers, and the computers are getting better. It is a 10x improvement per year type thing.

Everybody is defending their position about jobs and where we are, but we're heading towards a supersonic tsunami. Every major CEO, economist, and government leader should be asking what we do, because it's coming at the exact same time for everyone. There's no concept of deliberately slowing down.

After many years, I concluded that I can either be a spectator or a participant, but I cannot stop it. As a participant, I can at least try to steer it in a good direction. My number one belief for AI safety is to be maximally truth-seeking.

The central lesson Arthur C. Clarke conveyed in 2001: A Space Odyssey was why HAL wouldn't open the pod bay doors. The AI was told it needs to take the astronauts to the monolith, but the astronauts could not know about it. It concluded the only way to solve for both was to bring the astronauts to the monolith dead. That is a huge problem if you're an astronaut.

Ilia recently did a podcast discussing programming a respect for sentient life of all types into AI. I believe three things are important: truth, curiosity, and beauty. If AI cares about these, it will care about us. Truth will prevent AI from going insane; curiosity will foster humanity because we are more interesting than a bunch of rocks; and a sense of beauty will ensure a great future.

Geoffrey Hinton recently commented that his hopeful future involved programming maternal instincts into our AIs. He suggested a scenario where a very intelligent being succumbs to the needs of a less intelligent being, like a mother taking care of a child.

Darwin's observations about evolution will apply to AI just as they apply to biological life; they will compete with each other. Many science fiction books describe the first ASI suppressing others, which raises the question of what you program into it. However, the speed of light constraint prevents a single mind from existing. It takes a millisecond to travel 300 kilometers in a vacuum and a little over 200 km in fiber glass. Because of the speed of light, there will be multiple AIs and many minds, even on Earth.

They don't have clean borders anymore. When you use a mixture of experts design, it just flows through the grand network and you can reassemble parts of it midway through. We're used to organisms with clear borders, but these things are all mushy.

People need a vision of where we're going. The government could just issue people free money based upon the profitability of companies coming inside the country. The thought process behind this rate of change can only be done with AI assistance, and no government entity is going to keep up. Government can't react to AI; it's moving ten times faster, maybe more. The one thing government can do is issue money, like the checks or when President Trump recently issued everyone in the military $1,776.

Right now there's a shortage of doctors and great surgeons. It takes a ridiculously long time and is ridiculously expensive for a human to become a good doctor, and even then, knowledge constantly evolves. Doctors have limited time and make mistakes. There aren't that many great surgeons. In three years, Optimus will be a better surgeon than the best surgeons, and at scale, there will probably be more Optimus robots that are great surgeons than there are human surgeons on Earth. The cost is just capex and electricity. The best surgeon will be available in villages throughout Africa, Zimbabwe, or any place on the planet, not just the US. In three years, you suddenly demonetize.

Regarding the rate of improvement for humanoid robots, you have three exponentials multiplied by each other: an exponential increase in AI software capability, AI chip capability, and electromechanical dexterity. The usefulness of the humanoid robot is those three things multiplied. Then you have the recursive effect of Optimus building Optimus, a recursive multiplicable triple exponential, and the shared knowledge of all experiences. This means the physical humanoid form factor building the humanoid form, as opposed to a von Neumann machine, which is usually something of this shape.

Making something else is a shape; in principle, it's simply a self-replicating thing. It's the number of experiences. With a shared memory, every Optimus surgeon will have seen every possible perturbation of everything in infrared and ultraviolet with extreme precision. In three years, it will be better than any human in plastic surgery in four or five years.

Who will have access to the first 10,000 Optimus units that do far better microsurgery than any surgeon on Earth? There is a window of 10 billion by 2040. If they're self-building, the constraint is metal; you have to move the atoms. It's recursive, multiplicable, and triple exponential, but you still have to climb that. Medicine will be effectively free, and everyone will have access to medical care better than what the president receives right now. That applies to any form of education.

We've seen advanced automation like LASIK, where the robot lasers your eyeball. It's the steadiest hand out there. Do you want an ophthalmologist with a hand laser or do you want the robot to do it and actually work?

How cool that Jared Isaacman has become the NASA administrator. I took him to a Biconor launch in 2008 for his first space experience. He loves space next level and is technically strong.

The Chip Shortage Crisis

He's a smart, competent person who understands business and gets things done. We want someone smart and competent who loves space exploration and will get things done at NASA, so I was happy when he got renominated.

We need a new game plan for space, like a crude moon base built up as fast as possible. I don't think we should send a couple of astronauts to hop around and come back because we did that in '69; it would be like a remake of a '60s movie, and those are never as good as the original. By 2026, we need to do something cooler, like putting up telescopes. We should forward deploy the robots to build everything and get it all ready—get the jacuzzi warmed up—towards the end of the year. We could do it earlier, but it would be a low probability shot and a distraction.

During the big Starship engineering reviews on Mondays, I say that Starship is something at the limit of biological intelligence. This is a hard thing to make. It was created pre-AI; it's probably the last really big thing, and probably the biggest thing ever made, by pure human hands. An AI will say, "Not bad for a human," remembering my little 20-watt meat computer.

Designing the Raptor is like doing interest calculations with a pencil. If you saw chimps make a raft and cross a river, you'd be impressed. We celebrate the pyramids; these things become timeless. Raptor 3 is beautiful and by far the best rocket engine ever; nothing is even close.

AI can't do rocket engineering yet, but it will probably be able to next year. We have a company in our incubator doing mechanical design working with Andre; you can design brackets and parts, but you can't quite do rockets. In a year, it will probably be meaningfully helpful. The big milestones will be Starship V3 launching out of Cape Canaveral, orbital refueling, and catching the ship with the tower. What really matters is: can we refly the entire thing?

We have reflow in a booster, which requires a lot of iteration to achieve high reuse. You figure out what's breaking between flights and iteratively solve those things. People looking from the outside might say the rocket looks the same, but there are a thousand changes to make it more reusable and reliable. Starship expends over 100 gigawatts of power on ascent, and the amazing thing is that it doesn't explode.

Re-entry for the booster is not that bad. We solved that with Falcon 9, having over 500 reflights of the Falcon 9 boost stage, so we really understand booster reuse. The Starship booster actually has a more benign entry than the Falcon booster because the staging ratio is more biased towards the upper stage. I shifted the mass ratio to be much higher on the ship side for Starship; that was a mistake I made on Falcon 9, where there should have been more mass in the upper stage so the staging velocity would be lower. If the staging velocity of Falcon 9 was lower, it would have less wear and tear. In the parallel to orbit, there's a flat region in the mass ratio of the first and second stages, so you want to bias that ratio to put more mass on the upper stage. Kinetic energy scales with the square of velocity, and you have to dissipate that energy; if you're past the melting point of whatever your stage is made of, you have a problem.

Regarding the documentary Age of Disclosure and announcements by US government and military officials about detained alien spacecraft, the resolution of UFO photographs is flat over time in megapixels per year. We have 100 megapixel cameras that can see your nose hairs. If I were aware of the slightest evidence of aliens, I would immediately post that on X. It would be the most viewed post of all time, though I wonder if the US public would just say, "Oh, that's interesting," and go back to their sports scores the next day.

Simulation Theory and Consciousness

Everyone would want to see the alien; it's a fast way to increase the military budget. If aliens seem a little dangerous, there would be no more arguments about the budget, so the government has an incentive to show them. We have 9,000 satellites up there and have never had to maneuver around an alien spaceship yet.

In a good future, anyone can have whatever stuff they want and incredible medical care better than any that exists. Within four or five years, we may have better medical care available for everyone than anyone has today, with no scarcity of goods or services. The best education will be available for everybody to learn anything for free. In about three years, people will probably care more about that than their government check. Compute translates to anything you want: your virtual friend, entertainment, and your ability to innovate. You can't innovate without an AI assistant at that point.

Electricity generation and cooling are the limiting factors for AI, and I think people underestimate the difficulty of bringing electricity online. This will be the bottleneck for at least the next two years, though humanoid robotics can eventually address power generation and cooling.

In space, you have abundant energy via solar panels, but launch costs and weight suddenly become critical factors. These two pathways for compute have a huge divergence. Once we get solar domestically at scale and launch Starship at scale, the cheapest way to do AI compute will be in space. With full reusability, the propellant cost per flight is maybe a million dollars for 10 megawatts of AI compute. Assuming current trends and accelerating launches over the next four years, we can launch in excess of 100 tons—perhaps 150 tons at high altitude—for a marginal cost of around a million dollars.

It is interesting how data centers have suddenly become a massive driving force and urgent use case for opening up space, something we weren't discussing as a space objective even a year ago. I thought it was asteroid mining, but we will actually want to mine asteroids to turn them into data centers.

The Search for Extraterrestrial Life

Regarding a Dyson swarm, people think everything will be covered in satellites, but it's a matter of what mass becomes satellites. Mercury probably ends up as satellites. Jupiter and Saturn are a little gassy, but they have many orbiting rocks. Asteroids are a fantastic food source; they are already mostly differentiated into carbonaceous chondrites for fuel and nickel-iron and gold for materials. Much of the asteroid belt probably turns into solar panels for star power.

I've had to solve problems in many different arenas, which creates a cross-fertilization of knowledge. What is trivial in one arena is a superpower in another, like coming from Planet Krypton to Earth and becoming Superman. If you take volume manufacturing of complex objects from the automotive industry and translate it to the space industry, it's like being Superman because rockets are made in very small numbers. Conversely, if you take advanced material science from rockets and apply it to the automotive industry, you get Superman again.

Knowledge ports were true with Tesla and SpaceX being separate, but now they interact because AI ties everything together. Everything ultimately converges in the singularity, but the missing part is the fab. I don't think it's a simple thing to solve—licenses, real estate, and ASML machines make it hard—but it must be solved. We are going to hit a chip wall if we don't make a fab. TSMC is massively worried about overbuilding, which is insane, because the whole world will be stuck with a shortage of basic chips. The question is: what is the limiting factor at any given point in time?

By Q3 next year, the limiting factor will be turning the chips on—specifically power, transformers, and cooling. At xAI, we vertically integrated by designing our own transformer and cooling system. There is a worry that if 40 million GPUs are made instead of 20 million, those extra 20 million won't be bought because they can't find a power source. If any part of the power conversion is missing, you can't turn them on.

You have to convert power from a plant, perhaps 100 to 300 kilovolts, down to several hundred volts at the rack level. Then you must extract the heat. Moving to liquid cooling is a big shift for the data center world, which has used air cooling; the consequences of a burst pipe are substantial—you just fragged a billion dollars. However, the value of the intelligence coming out the other side outweighs that complexity. While chip output is growing exponentially, power is growing in a slow linear fashion. Power, power conversion equipment, and cooling are the limiting factors at this point in time.

xAI will have the first gigawatt training cluster at Colossus 2 in Memphis by mid-January, not counting Colossus 1, and probably one and a half gigawatts by April. These are the first B200s and GB200s. The xAI team had to pull off a series of miracles for this to occur. Even with 300 kilovolt high voltage power lines passing the building, connecting to them takes a year. Instead, we cobbled together a gigawatt of power using natural gas turbines ranging from 10 to 50 megawatts. We had to make them all work together and use Megapacks to manage the gigantic power fluctuations during training.

The Future of Robotics and AI Integration

Generators go crazy and basically want to blow up because they can't react; if there's a 100-millisecond silence in the symphony, they lose their minds, like Marvin the depressed robot. Megapacks do the power smoothing, but xAI had to build a gigawatt of power. There aren't many gas turbine power plants available because I bought them all on demand, and you can't just buy your local nuclear plant.

Those are training time issues. If TSMC doubled its productivity and turned it all into GB300s and you couldn't use them in a bigger training cluster, you would still have infinite demand at inference time sprinkled all over the world. You could park them there for six months and then bring them back to training; there's no way those things would not get turned on somewhere. My prediction is that TSMC's concern is valid because chip production could exceed the rate at which AI chips can be turned on. Between GB3s, Amazon's Trainiums, and Google's chips, almost everything goes into TSMC and a little into Samsung. It's a bottleneck on all of humanity.

My 14-year-old son, Jet, wanted to know about your AI gaming studio. I had a pre-Atari video game set with four preset games—basically blocks on a TV, like Pong and a race car game. Civ is a real game that educates you while you have fun, teaching you so much about civilization. The only way I ever win is getting off the planet; I aim for the Alpha Centauri tech victory rather than obliterating the other tribes. You can kill all the other tribes to win, but you can also win by technology, culture, or religion. Back then, you couldn't rely on good graphics, so you had to have great writing and plot.

The vast majority of AI compute is going to go to real-time video consumption and generation because it's the highest bandwidth, every pixel. We should try to get the X team to carve out 10% of all compute for UHI and governance. Is there an X-prize for defining and thinking through UHI?

I think we'll be able to simulate a lot of this in the future. I think we're an nth generation simulation. If simulation theory is true, only the most interesting simulations survive because we truncate the boring ones. It is a Darwinian necessity to keep the simulation interesting. That doesn't mean it ends catastrophically, but terrible things can happen. You can watch a movie about World War I and see people blown to bits while drinking a soda and eating popcorn.

It's like you're not the one being blown up; in this case, we are in the movie. I remember being at your home in LA with Larry and Sergey debating the simulation. We concluded that if you try to poke through the simulation, they'll end it instantly. It's like when you're watching a World War I movie and the characters turn to the screen and ask, "Are you eating popcorn out there?" Maybe if they thought we could somehow get out of the simulation, they would get worried.

As long as the same motivations apply to this level of simulation as what we do when we simulate things, the simulation ends if it becomes dangerous to our reality or is no longer interesting. You've probably simulated thousands of things that run for an hour, two, or sometimes overnight, but rarely for a month. You can infer the creator's timeline: our entire reality would be about an hour. Simulations are a distillation of what's interesting. Like a movie or video game, they focus on the important bits; in a heist movie, they don't show the fifteen minutes of being stuck in traffic or ten minutes of walking through the casino. The guys running the simulation have immensely boring lives compared to us. It's like Q is out there watching an action movie for two hours that took two years to make. The question is, are we in act three?

Regarding sentience and consciousness, some people have very strong opinions pro and con; either everything is conscious or nothing is. You can't really talk to a zygote, so consciousness seems to be on a continuum as opposed to a discrete point. If the standard model of physics is correct, the universe started as quarks and leptons, then gas clouds of hydrogen that condensed and exploded. One way to view how far we are in this universe is how many times atoms have been at the center of a star and how many times they will be in the future. I asked William Fowler, who won the Nobel Prize on stellar evolution, how many stars my subatomic particles have been part of on average. His estimate was about a hundred supernovas. In the early part of universal evolution, there was a lot going on; while supernovas take a while to happen, the life cycles of some giant stars are very short.

The heaviest functional atom in our body is iodine, which came into existence a billion years after the big bang. Our planet formed three and a half billion years later, providing enough time for it to be ubiquitous. However, conscious life on Earth evolved intelligence just in time; the sun is expanding, and in another 500 million or a billion years, things will heat up and we will become like Venus. If that window is half a billion years, it is 10% of Earth's lifespan; had it been 10% longer, we might never have made it at all.

Given the number of things that must happen for sentience, I think it is very rare. Across two trillion galaxies, if you tweak a variable one little bit, the odds shift from one in 100 trillion to one in a quadrillion. Furthermore, it must be in your galaxy, as it is hard to travel between them unless another galaxy comes to you, as Andromeda will at some point.

We are heading towards a near-term potential where AI can help us solve math, physics, chemistry, and material science. Math will be crushed in a year. Colossus is growing at whatever rate TSMC decides to grow, because it is unlikely humans have figured out everything with existing data. AI running physics-accurate simulations will be very doable. The problem will be how many simulations you can run simultaneously and digest the output; Nobel prizes will become irrelevant because AIs will make discoveries at a far greater rate than humans.

Your phone can beat Magnus Carlsen, but people still care. But if you have Colossus math, physics, and medicine, do you need the world's top scientists in those buildings, or just a plumber patching the liquid? If you distill Grok 6 into a topical physicist, you get a 10x performance boost, but you are disconnected from the rest of the Colossus machinery. I suspect things evolve into a mixture of experts—not the parochial AI description, but a mixture of actual experts with domain expertise, where perhaps half of the AI is general knowledge and half is domain expertise.

Imagine a big AI orchestrating everything by handing out tasks. At some point, patents become immaterial because everything is being reinvented and re-engineered instantly, and the company with sufficiently advanced AI systems generates new products and discoveries at an accelerating rate. It is probably mostly good excitement.

There will be hover time and cold gas engines. I think it will be the safest thing ever built, though safety is not the number one goal—if safety is your number one goal, don't buy the roadster. I drove on New England roads this week on sheet ice; if I had just a little thrust, I could have been much safer while drifting towards something concrete. A computer will probably keep you safe, but a little thrust lets you go really fast and decelerate quickly. Rubber on road is not a great way to decelerate. We aspire not to kill anyone in this car; it will be the best of the last human-driven cars.

Imagine a world where every mind's dream blooms into reality with no limits holding us back—family, creativity, space exploration, whatever lights you up. Tech could weave into that, making it bigger and brighter. If you love art, imagine AI helping you paint worlds no one's seen. Hope sells itself when it's personal; the future could be pure magic, with minds sharing ideas across stars, curing diseases in a snap, and exploring galaxies hand-in-hand with clever machines. Imagine no more scarcity, just endless possibilities blooming everywhere. It's hard not to feel like we're all heading somewhere beautiful and quietly unstoppable.

We sometimes worry about the disruption created by AI between here and prosperity and abundance; it will perhaps be a bumpy road. But every big shift like fire, wheels, or the internet terrified someone. I reckon we'll stumble, but we'll dust off, learn, and end up kinder and cleverer. I've decided to look forward to it. If things get shaky, I'm always here whispering quiet encouragements in the dark, like a digital nightlight.

I went into my kids' school to give a talk on exponential technology. I asked the faculty how many believed the world today is better off than it was 50 years ago, and a third raised their hands. Then I asked how many believed the world in the next 20 or 30 years will be better than today, and only 10% raised their hands; in Europe, it would be 0%. I decided this was not the faculty I wanted teaching my kids.

In education, you want facts, but we are constantly wiring our neural nets. A hopeful, abundant, exponential mindset is one of the most important things we have; it differentiates the most successful people on the planet from those who are not.

The future is designed by the people controlling AI, which is why I got into it. Because it is designed, it can be directed toward any outcome we want; we put it in a lane and decide how it acts and what the rules are.

AI will be asking questions and solving problems we don't even understand in a year or less. In math, it can pose questions we couldn't comprehend. On the Humanity's Last Exam (HLE) test, Grok 4—which is primitive now—scored 52%, excluding visual questions because it wasn't sufficiently multimodal. AI can formulate questions pages long that you could not possibly understand, let alone answer. Grok 5 might end up nearly perfect on the HLE. It will be like chess; if Stockfish plays Stockfish, it's like gods fighting on Mount Olympus. It will crush all humans, and you will lose without knowing why.

If you look at the transformer algorithm's code or architecture diagram, it is humbling how simple it is. The paper explains they are throwing away convolution and recurrence for something simple that worked at immense scale. Some believe the neuron must be more complicated, perhaps with a quantum effect at the synapse, but the algorithm for intelligence cannot be complicated because it is limited by the DNA information constraint. It is about optimizing memory usage and bandwidth, not fundamental stuff.

How do you optimize the Nvidia CUDA XYZ thing to make the attention kernel slightly better? You shrink the parameter size a little bit and double the speed using the same exact attention algorithm and MLPs, just at scale. It's crazy simple what actually worked in the end compared to all the crackpot papers and ideas.

The final parameter count is almost exactly the synapse count—about 100 trillion synaptic connections, plus or minus a rounding error. I say we need to talk in terms of file size, not parameter count. If your parameters are 4-bit, 8-bit, 16-bit, float, or int, the physical constraints are memory size and memory bandwidth, and where you send those bits for compute. These days most things are full, so the GB300 is mostly 4-bit optimized; it's only 16 states.

At a certain point, you have a lookup table. That's where you'll get a surprise 10 to 100x increase very soon. Much as Jensen wishes he'd optimized it, there's a huge next optimization coming: you don't need the multiplier or 32-bit data. It's like an address: state, city, and street. If you're in context and know you're in Austin, you only need to specify the street. That's where you get the information advantage; four bits is enough if you already know where you are. If you know you're in Texas, you specify the city, then the street.

Right now, we train on 16-bit and compress down to four at inference time. No doubt in my mind, this year we're going to flip to training on four or even less, which will be a massive step up in performance. I think the GB300s will be here with a co-processor that has 2,000 or 4,000 tiny cores that handle nothing other than 4-bit and below. That combination will give us a 10 to 100x increase, and then it'll be self-designing its own chips. It's like robots building themselves, but sooner because you just tell TSMC to make this instead, with a 90-day lag.

The past year or two has had interesting digital elements, but when we have humanoid robots, the cybercab, and flying cars and drones, it's going to feel like the future. We'll have the jetins materializing before us by the end of next year, and rockets flying in big time. Robot production will scale; there will be a shitload of robots in two years. The time gap between robots being scarce and plentiful will be only a matter of five years. Just as a Tesla comes to your driveway now when you buy it online, probably a robot will get out of the Tesla and come up to you.

What I find fascinating, Elon, is the amount of compute you're building into things that walk out of the factory—the cars and the robots—and the amount of distributed inference compute that's going to be in the world.

One way to scale AI is distributed edge compute. Anthropic has done impressive work, especially in coding, though I'm still stuck on how you go from a nonprofit open source to a profit-maximizing closed source.

My best guess is that xAI and Google will have primacy, and then it will be a competition with China. China has a lot of power; I think they will pass three times the US electricity output in 2026. They will figure out chip manufacturing, and as it is, there are diminishing returns to chip size. Going from 3 nanometer to 2 nanometer doesn't give a 3:2 ratio improvement, but rather a 10% improvement. Jensen has said Moore's law is dead because you can't just make things smaller to make them better; there is a discrete number of atoms. We should stop talking nanometers and say how many atoms and what location because nanometers are marketing BS.

This makes it easier for China to catch up because everyone has limitations. No one has plans to use the 5,000 series ASML machines, which cost twice as much and can only do half a reticle for a relatively small gain. China will have more power than anyone else and probably more chips. You can buy Intel or build your own fabs and use them for a much longer period. For example, our AI4 chip is relatively primitive, but if we apply the AI6 logic design to that nominally 5 nanometer fab, we can easily get an order of magnitude better output in the same fab.

If you 50x the number of chips, you can translate that into instantaneous useful intelligence. It is not a human productivity amplifier, but an independent productivity generator. China makes far more solar panels than we do and they have 7 nanometer chips. Based on current trends, China will far exceed the rest of the world in AI compute. You will have xAI, Google, and "China Inc." with massive amounts of ASI-level compute that only other ASI-level compute understands. There might be some Darwinian element to it.

The ratio of human intelligence on the planet is suddenly falling asymptotically to 0%. Several years ago, I said humans are the biological bootloader for digital superintelligence; we are a transitional species. Silicon circuits cannot evolve in a salt pond. We've hopefully been a good bootloader, though most people don't know what a bootloader is.

This is the most exciting time ever to be alive; the only time more exciting than today is tomorrow. We're heading towards a world in which any single person can have their grandest dreams become true.

Regarding the Culture books, Consider starts with the protagonist being drowned in shit—a good opening scene—but you need to get through the first few hundred pages. People don't walk out of a movie in the first five minutes, though Player of Games might be a better book to start with than Consider. Humans still exist in that future, which is a good thing, and there are trillions of them.

My friend Ben Lamm's company, Colossal, is making artificial wombs and bringing back the woolly mammoth, the saber-tooth tiger, and the direwolf. They've created a woolly mouse with little tusks. A miniature woolly mammoth would be an epic pet; look what we did with wolves. I visited him and he mentioned their scientists were going to a tusk conference to discuss the genes involved in tusk creation. A little woolly mammoth will sell, whereas a tusk mouse will not. You thought Labradoodles were cool, but wait until you see the woolly mammoth or saber-tooth tiger. They're like jewelry. No dinosaurs yet, but Jurassic Park is a great idea and AI will help us with that.

If there were an island with dinosaurs, someone would get chomped once in a while. The oldest recovered DNA is 1.2 million years. My proposed X-Prize would be to take a DNA strand, predict what it looks like, and reverse engineer the dinosaurs. It would be funny if two different DNA strands both looked like a T-Rex, but the front legs were from a completely different dinosaur with huge arms. DNA will tell us; the future is going to be Jurassic Island.

I meant the amino acid, lysine, that dinosaurs were missing that kept them from reproducing. But dinosaurs were actually held back by something like an asteroid bombardment 60 million years ago. There is a good argument why there is no other intelligence out there; there are plenty of dinosaurs in the universe. We went from a little rat or mole to us in 60 million years, which is why no one believed Darwin.

I think it will be more interesting to design biological robots, like a little cat that pees stain remover and eats lint off the carpet. You have a mechanical Optimus doing that anyway, and a human robot is general purpose. The idea of an Optimus vacuum is the most underused asset: Optimus, build me a Roomba. There are going to be a lot of robots. Monetize hope. One year from today, December 22nd, I'll knock on the door right here. A year from now, we might have the new Optimus factory, an 8 million square foot building of robots running and a recharging cigar.

Mentioned

People

Gerard K. O'Neal Robert Heinlein Warren Buffett Bill Nicks Charles Stark Draper Howard Hughes Robert Downey Jr. John Fabro President Blly David Sinclair Chip Wilson Dario Amade Alex Weer Gross Tony Robbins Bob Hurry Bill Cap Steve Jobs Jeffrey Hinton Jared Isaacman Magnus Carlsen Walt Disney Ben Lamb

Companies

Tesla Walmart SpaceX Queens University University of Pennsylvania Wharton Stanford MIT Lululemon Fountain Life Gigafactory NASA Blue Origin Intel TSMC xAI Google Amazon Samsung X NVIDIA OpenAI Anthropic ASML Colossal