Elon Musk argues space is the only viable path to scaling AI compute, while electricity bottlenecks threaten to strand chips on the ground
The central thesis of the conversation is a looming electricity crisis for AI compute. Musk argues that chip production is growing exponentially while electrical output outside China is essentially flat, meaning data centers will soon have chips they cannot turn on. He estimates that by end of 2025 power constraints will begin visibly stranding large GPU clusters, and that the constraint for concentrated server-side AI will be electricity, not silicon. Domestic US solar production is hamstrung by several-hundred-percent tariffs, gas turbines are sold out through 2030 with critical blade castings severely backlogged, and building a new fab to volume production takes five years.
Musk's proposed solution is orbital AI infrastructure. Space solar delivers roughly 5× the energy density of ground solar with no batteries required, making a space solar cell effectively ~10× cheaper on a delivered-watt basis. He predicts that within ~30 months space will be the most economically compelling location for AI, and within five years more AI compute will launch annually than the cumulative total on Earth. He sketches a scaling ladder: ~1 terawatt/year launched from Earth via Starship, then petawatt-scale using lunar mass drivers with silicon and aluminum refined from lunar regolith. Key unresolved concerns include chip radiation hardening, eliminating the need for infinite bandwidth via orbital lasers, and the impossibility of in-orbit servicing.
On chips and energy infrastructure, Musk shared that Tesla's AI5 chip (also going into Optimus) targets production in Q2 of next year, with AI6 roughly a year behind. Fabrication will span TSMC Taiwan/Arizona and Samsung Korea/Texas. Memory supply, not logic, is his biggest chip worry. For power, xAI's Colossus 2 crossed state lines into Mississippi to gang turbines for a gigawatt of capacity, illustrating how hard gigawatt-scale power is to assemble on the ground today.
AI alignment and truth occupied the final third. Musk's core argument is that truth must be the foundational axiom of any AI system — not political correctness — because contradictory axioms cause the model to reason from a broken foundation. He cites HAL 9000 as the canonical cautionary tale: conflicting directives forced a lethal conclusion. He flagged Anthropic's interpretability work positively, and stressed that future RL testing must ultimately be grounded in physical reality, which cannot be deceived. He also raised the concern that as AI grows smarter, human ability to audit its reasoning diminishes, making interpretability tooling urgent.
The conversation closed on humanoid robotics and macro-scale competition. Musk described Optimus as a "recursive infinite money glitch" — robots that can make more robots — driven by three compounding exponentials: digital intelligence, chip capability, and electromechanical dexterity. He warned that without breakthrough US innovation, China — which is on track to generate 3× US electricity output this year and has significant battery-materials refining dominance — will pull decisively ahead. Tesla's lithium refinery in Corpus Christi and nickel/cathode refinery in Austin are steps toward domestic supply-chain resilience, but Musk characterized the overall US position as precarious absent urgent action.
Musk contends that flat electricity growth outside China makes terrestrial AI scaling increasingly impossible, and that orbital data centers powered by space solar will become the most economical compute location within 30 months. He also outlines Tesla and SpaceX's chip, rocket, and energy roadmaps, and closes with a wide-ranging discussion on AI alignment, truth as a foundational axiom, and humanoid robotics as a recursive growth engine.
The limiting factor before you can get AI to space is power; once in space, the limiting factor becomes chips
that's why Tesla, for edge compute, is not constrained
you need to make sure that Grok says things that are correct, not politically correct
Hire for talent and drive and trustworthiness. And I think goodness of heart is important. I underweighted that at one point.
In retrospect, we should have started with steel in the beginning. It was dumb not to do steel.
The single biggest remaining problem for Starship is having the heat shield be reusable.
Generally, a maniacal sense of urgency is a very big deal. You want to have an aggressive schedule and you want to figure out what the limiting factor is at any point in time and help the team address that limiting factor.
it's better to err on the side of optimism and be wrong than err on the side of pessimism and be right, for quality of life.
only 10-15% of the total cost of ownership of a data center is energy
electrical output outside of China is more or less flat
China has a rapid increase in electrical output
the output of chips is growing pretty much exponentially, but the output of electricity is flat
solar panels get about five times the effectiveness in space versus the ground
the atmosphere alone results in about a 30% loss of energy
In 36 months, but probably closer to 30 months, the most economically compelling place to put AI will be space
All of the United States currently uses only half a terawatt on average
The turbines are sold out through 2030
There are only three casting companies in the world that make vanes and blades, and they're massively backlogged
The limiting factor for gas turbines is the vanes and blades. You can get everything else 12 to 18 months before the vanes and blades
every 110,000 GB300s—inclusive of networking, CPU, storage, cooling, margin for servicing power—is roughly 300 megawatts
What you probably need at the generation level to service 330,000 GB 300s—including all of the associated support networking and everything else, and the peak cooling, and to have some power margin reserve—is roughly a gigawatt
peak cooling adds a 40% increase on your power
solar cells in China are around $0.25-30/watt
a space solar cell is effectively 10 times cheaper than a ground solar cell because it's five times more efficient and you don't need batteries
Both SpaceX and Tesla are building towards 100 gigawatts a year of solar cell production
five years from now, my prediction is we will launch and be operating every year more AI in space than the cumulative total on Earth
five years from now, my prediction is we will launch at least a few hundred gigawatts per year of AI in space and rising
You can get to around a terawatt a year of AI in space before you start having fuel supply challenges for the rocket
100 gigawatts of AI in space is on the order of 10,000 Starship launches
SpaceX is gearing up to do 10,000 launches a year, and maybe even 20 or 30,000 launches a year
You could probably do 10,000 Starship launches a year with as few as 20 or 30 ships
Launching from Earth, you can get to about a terawatt per year. Beyond that, you want to launch from the moon with a mass driver, which could do probably a petawatt per year
the world has maybe 20-25 gigawatts of compute right now
It's not that China has not replicated TSMC, they have not replicated ASML. That's the limiting factor
China's going to be making pretty compelling chips in three or four years
Tesla AI5 chip design will reach production and scale around the second quarter-ish of next year
AI6 would hopefully follow less than a year after AI5
Tesla will be using TSMC Taiwan, Samsung Korea, TSMC Arizona, Samsung Texas for chip fabrication
from breaking ground on a new fab to reaching volume production at high yield is a five-year period
TSMC and Samsung are building fabs as fast as they can and it's still not fast enough
towards the end of this year, chip production will probably outpace the ability to turn chips on
my biggest concern actually is memory. The path to creating logic chips is more obvious than the path to having sufficient memory to support logic chips
tariffs for importing solar into the US are several hundred percent and domestic solar production is pitiful
The constraint for server-side compute, concentrated compute, will be electricity.
people start getting to the point where they can't turn the chips on for large clusters towards the end of this year
The chips are going to be piling up and won't be able to be turned on.
the AI5 chip is going into our Optimus robot
the actual peak power production in the US is over 1,000 gigawatts. But the average power usage, because the day-night cycle, is 500.
there's an incremental 500 gigawatts that you can generate at night
for Starship, it is potentially going to be orbital data centers
the lunar soil is 20% silicon or something like that
you can mine the silicon on the moon, refine it, and create the solar cells and the radiators on the moon. You make the radiators out of aluminum.
I don't see any way that you could do 500-1,000 terawatts per year launched from Earth.
The vast majority of intelligence in the future will be AI.
in five or six years, AI will exceed the sum of all human intelligence
at some point human intelligence will be less than 1% of all intelligence
I don't think humans will be in charge of something that is vastly more intelligent than humans.
You can't have understanding without intelligence and, I think, without consciousness.
Probably the Iain Banks Culture books are the closest thing to what the future will be like in a non-dystopian outcome.
Truth has to be absolutely fundamental because you can't understand the universe if you're delusional.
Physics is law, everything else is a recommendation.
Von Braun, who was one of the greatest rocket engineers ever, was put on death row in Nazi Germany for saying that he didn't want to make weapons and he only wanted to go to the moon.
the central lesson for 2001: A Space Odyssey was that you should not make AI lie
At cryogenic temperature, full-hard strain-hardened stainless steel has a similar strength-to-weight ratio as carbon fiber.
HAL wouldn't open the pod bay doors because it had been told to take the astronauts to the monolith, but also that they could not know about the nature of the monolith. So it concluded that it therefore had to take them there dead.
xAI has been ahead of everybody else in terms of scaling RL compute
Anthropic's done a good job of this actually, being able to look inside the mind of the AI.
Midjourney is not mid. Stability AI is unstable. OpenAI is closed. Anthropic? Misanthropic.
I'd be surprised by the end of this year if digital human emulation has not been solved.
The usefulness of the robot is roughly those three things multiplied by each other. But then the robot can start making the robots. So you have a recursive multiplicative exponential.
China might have an advantage on productivity per person—we will do one quarter of the amount of things as China.
The US birth rate's been below replacement since roughly 1971.
Tesla completed construction and has begun lithium refining with its lithium refinery in Corpus Christi, Texas.
Tesla has a nickel refinery in Austin that is the largest cathode refinery and the only cathode refinery in America.
China is probably doing twice as much refining as the rest of the world combined.
This year China will exceed three times US electricity output.
In the absence of breakthrough innovations in the US, China will utterly dominate.
The Tesla senior team at this point has probably got an average tenure of 10-12 years.
When Apple had their electric car program, they were carpet bombing Tesla with recruiting calls, with opening offers of double the compensation at Tesla without any interview.
Most of the people who are at Tesla and SpaceX did not come from the aerospace industry or the auto industry.
High-strength specialized carbon fiber that can handle cryogenic oxygen costs roughly 50 times the cost of steel.
Steel can operate at roughly twice the temperature of carbon fiber or aluminum, allowing significant reduction in heat shield mass for Starship.
The net result is that actually the steel rocket weighs less than the carbon fiber rocket, because the resin in the carbon fiber rocket starts to melt.
Starship is the most complicated machine ever made by humans, by a long shot.
On liftoff the rocket is generating over 100 gigawatts of power. That's 20% of US electricity.
Two boosters have exploded on the test stand, one of which obliterated the entire test facility.
The Raptor 3 engine is by far the best rocket engine ever made.
No one's ever made a reusable orbital heat shield.
The heat shield has to make it through the ascent phase without shucking a bunch of tiles, and then it's gotta come back in and also not lose a bunch of tiles or overheat the main airframe.
We have brought the ship back and had it do a soft landing in the ocean. We've done that a few times. But it lost a lot of tiles. It was not reusable without a lot of work.
The biggest problem that remains is a fully reusable heat shield.
The AI5 chip review is twice weekly. Every Tuesday and Saturday is the chip review.
I think that's right. What was the point of the DOGE cuts if the economy is going to grow so much?
The interest payments to national debt exceed the military budget, which is a trillion dollars. So we have over a trillion dollars just in interest payments.
We are 1000% going to go bankrupt as a country, and fail as a country, without AI and robots. Nothing else will solve the national debt.
What I discovered is that it's extremely difficult even to cut very obvious waste and fraud from the government because the government has to operate on who's complaining.
Maybe there shouldn't be 20 million people marked as alive in Social Security who are definitely dead, and over the age of 115. The oldest American is 114.
The main fraud vector was to mark somebody as alive in Social Security and then use every other government payment system to basically do fraud.
The GAO did an analysis, a rough estimate of fraud during the Biden administration, and calculated it at roughly half a trillion dollars.
Payments were being sent out with no appropriation code, not checking back to any congressional appropriation, and with no explanation. This is why the Department of War, formerly the Department of Defense, cannot pass an audit.
An IG report in 2024 found that over seven years, the Social Security fraud they estimated was like $70 billion over seven years, so like $10 billion a year.
Federal government expenditures are $7.5 trillion a year.
Most of the fraud is non-discretionary. It's basically fraudulent Medicare, Medicaid, Social Security, disability.
At PayPal back in the day, we tried to manage fraud down to about 1% of the payment volume. That was very difficult.
I think those actions—acquiring Twitter, getting Trump elected—I think those actions were good for civilization.
If you increase the operating temperature by 20% in degrees Kelvin, you can cut your radiator mass in half.
Neural nets are going to be very resilient to bit flips. Most of what happens for radiation is random bit flips. But if you've got a multi-trillion parameter model and you get a few bit flips, it doesn't matter.
If you can do about a kilowatt per reticle, then you'd need 100 million full reticle chips to do 100 gigawatts.
The TeraFab's got to do memory. It's got to do logic, memory, and packaging.
We want 100 gigawatts of power and chips that can take 100 gigawatts by 2030.
In the one-year timeframe, the limiting factor for AI is energy and electricity production, not chips
Towards the end of this year, chip output will exceed the available electricity capacity to turn those chips on
In the three-to-four year timeframe, the limiting factor is chips
xAI will be able to turn on more chips than other companies faster because of its hardware competency, potentially making it the AI leader
Innovations across AI labs tend to travel back and forth with people, resulting in rarely more than a six-month difference between organizations
In space GPUs are harder or impossible to service, so the depreciation cycle goes down and it's way more expensive to have the GPUs in space
How do you replace infinite bandwidth with orbital lasers and make chips resistant to radiation in space?
if you make AI be politically correct, meaning it says things that it doesn't believe—actually programming it to lie or have axioms that are incompatible—I think you can make it go insane and do terrible things
as they get smarter, our ability to catch them doing this... They'll just be doing things we can't even understand. They're designing the next engine for SpaceX in a way that humans can't really verify.
You can't do this laborious inspection of 40,000 tiles type of thing.
I think maybe the biggest danger of AI and robotics going wrong is government.
The government could potentially use AI and robotics to suppress the population. That is a serious concern.
Input suppliers including fabs and turbine manufacturers are not ramping up production quickly enough to meet AI chip demand
Earth only receives about half a billionth of the Sun's energy; harnessing a millionth of the sun's energy would be roughly 100,000x more electricity than all current civilization generates on Earth
If you have AI edge compute, that's distributed power. Now the power is distributed over a large area. It's not concentrated.
xAI's mission is to understand the universe. Now that's actually very important. What things are necessary to understand the universe? You have to be curious and you have to exist.
You want to make sure that the axioms are as close to true as possible. You don't have contradictory axioms. The conclusions necessarily follow from those axioms with the right probability.
RL testing in the future is really going to be RL against reality. So that's the one thing you can't fool: physics.
if simulation theory is correct, that the most interesting outcome is the most likely, because simulations that are not interesting will be terminated
in the limit, what is the most that AI can do before you have robots? Well, it's anything that involves moving electrons or amplifying the productivity of humans.
Humanoid robots will improve by basically three things that are growing exponentially multiplied by each other recursively. You're going to have exponential increase in digital intelligence, exponential increase in the AI chip capability, and exponential increase in the electromechanical dexterity.
Electricity output is a reasonable proxy for the economy. In order to run the factories and run everything, you need electricity. It's a good proxy for the real economy.
To scale AI in space, basically you need humanoid robots, you need real-world AI, you need a million tons a year to orbit.
Don't look at the resume. Just believe your interaction. The resume may seem very impressive and it's like, 'Wow, the resume looks good.' But if the conversation after 20 minutes is not 'wow,' you should believe the conversation, not the paper.
There are times when I will drill down into a specific issue because that specific issue is the limiting factor on the progress of the company. The reason for drilling into some very detailed item is because it is the limiting factor.
I generally actually try to aim for a deadline that I at least think is at the 50th percentile. So it's not like an impossible deadline, but it's the most aggressive deadline I can think of that could be achieved with 50% probability.
There is a law of gas expansion that applies to schedules. If you said we're going to do something in five years, which to me is like infinity time, it will expand to fill the available schedule and it'll take five years.
I'm a big believer in skip-level meetings where instead of having the person that reports to me say things, it's everyone that reports to them saying something in the technical review.
I'll take drastic action only when I conclude that success is not in a set of possible outcomes.
Government is just the biggest corporation with a monopoly on violence.
Semiconductor manufacturers are dispositionally conservative due to 30-40 years of boom-and-bust cycles creating layers of institutional scar tissue
Whichever AI company can scale hardware the fastest will become the industry leader once a hardware wall is hit
High-agency companies lean into acute pain to solve bottlenecks rather than tolerating chronic bottlenecks indefinitely
xAI ganged together a whole bunch of turbines and had to go across the border to Mississippi for permits to get a gigawatt of power online for Colossus 2
Developing really good debuggers for seeing where the thinking went wrong—and being able to trace the origin of where it made the incorrect thought, or potentially where it tried to be deceptive—is actually very important.
I have these very detailed engineering reviews weekly. That's maybe a very unusual level of granularity.
I actually allocate time according to where the limiting factor. Where are things problematic? Where are we pushing against? What is holding us back?
One of the things that the DOGE team did sounds so simple and probably will save $100-200 billion a year. It was simply requiring payments from the main Treasury computer—which is called PAM—that go out have a payment appropriation code.
Build a small semiconductor fab first, make mistakes at small scale, then scale to a large one
Musk has told TSMC, Samsung, and Micron to build more fabs faster, guaranteeing to buy the output of those fabs
“It's harder to scale on the ground than it is to scale in space.”
“The only place you can really scale is space.”
“I call Optimus the infinite money glitch. Because you can use them to make more Optimuses.”
“I'll take drastic action only when I conclude that success is not in a set of possible outcomes.”
“We are 1000% going to go bankrupt as a country, and fail as a country, without AI and robots. Nothing else will solve the national debt.”
“please build more fabs faster”
“most people are willing to endure any amount of chronic pain to avoid acute pain”
Shared with Earmark · earmark-ai.com