Elon Musk – "In 36 months, the cheapest place to put AI will be space”

Summary

• Musk’s headline call: within 30-36 months, space becomes “by far the cheapest place to put AI.” The logic: electricity output outside China is flat while chip output grows exponentially, solar panels do ~5x more power in orbit (no night, clouds, or atmosphere) and need no batteries — so space solar is effectively 10x cheaper — and “it’s harder to build on land than it is in space” once permits are counted. “You can mark my words.”
• The near-term wall is power, not chips: “towards the end of this year, chip production will probably outpace the ability to turn chips on.” Gas turbines are sold out through 2030, bottlenecked by cast vanes and blades from only three casting companies; SpaceX and Tesla will “probably have to make the turbine vanes and blades internally,” and both carry a mandate for 100 GW/year of domestic solar cell production from raw materials up.
• SpaceX as “hyper-hyper” hyperscaler and IPO tell: in five years Musk predicts SpaceX will launch and operate each year more AI in space than the cumulative total on Earth — hundreds of GW/yr, ~10,000 Starship launches (one per hour, which he says could probably be done with as few as 20-30 ships). On going public he can only “talk physics,” but concedes public markets hold “way more than 10x,” maybe 100x, private capital — and “I just repeatedly tackle the limiting factor. If capital is the limiting factor, I’ll solve for capital.”
• The 3-4 year limiting factor is chips, hence the “TeraFab” — logic, memory, and packaging at possibly north of a million wafers/month, starting with a small fab that is not yet done (“we could just flounder in failure”). Existing fabs can’t help: Tesla has booked TSMC Taiwan/Arizona and Samsung Korea/Texas; prepaying has not accelerated expansion because a fab takes five years start-to-volume. Biggest worry: memory, “that’s why DDR prices are going ballistic.”
• Optimus is “the infinite money glitch” — three multiplied exponentials (digital intelligence, chip capability, electromechanical dexterity) plus robots building robots: “a supernova.” The hand is “more difficult than everything else combined”; nothing in Optimus comes from a catalog; Optimus 3 targets ~1M units/year. Downstream: “corporations that are purely AI and robotics will vastly outperform any corporations that have people in the loop. And this will happen very quickly.”
• xAI’s revenue answer to OpenAI’s $20B and Anthropic’s $10B (vs xAI’s reported $1B): digital human emulation, “solved” by end of this year (the MacroHard project) — unlock that and “you basically have access to trillions of dollars of revenue” (customer service alone ~1% of world economy). The plan to win is the Tesla self-driving path — “a self-driving computer” — plus out-scaling everyone on hardware, since lab ideas diffuse within ~6 months.
• The China call is stark: “In the absence of breakthrough innovations in the US, China will utterly dominate.” China this year exceeds 3x US electricity output (his proxy for industrial capacity), does ~2x the world’s ore refining, and has 4x the population with higher work ethic — “we definitely can’t win on the human front, but we might have a shot at the robot front.”
• Macro/political tail: interest on US debt now exceeds the ~$1T military budget and Musk says “we are 1000% going to go bankrupt as a country without AI and robots”; DOGE might buy time. His stated biggest AI danger may be government, “the biggest corporation with a monopoly on violence,” using AI and robotics to suppress the population.

Deep dive

1. The 36-month call: space becomes the cheapest place to run AI

• Dwarkesh opens with the obvious objection: energy is only 10-15% of data-center TCO, GPUs are most of it, and space makes servicing harder, perhaps impossible. Musk’s answer reframes the problem: “The availability of energy is the issue.” Electricity output outside China is flat while “the output of chips is growing pretty much exponentially… So how are you going to turn the chips on? Magical electricity fairies?”
• The physics of the arbitrage: no day-night cycle, seasonality, clouds, or atmosphere (atmosphere alone costs ~30%), so any panel does ~5x more power in space — and with no batteries needed, “it’s not five times cheaper, it’s 10 times cheaper.” Chinese cells are already “farcically cheap” at $0.25-30/watt or something like that. His shirt he almost wore: “it’s always sunny in space.”
• Space is also partly a regulatory play: “It’s pretty hard to cover Nevada in solar panels. You have to get permits. Try getting the permits for that.” On servicing, GPU infant mortality can be ironed out on the ground before launch; past the initial debug cycle “they’re quite reliable.” The prediction, repeated deliberately: cheapest place for AI in 36 months, “probably closer to 30.”

2. The power wall on Earth: turbine blades, permits, and noob math

• Software people “are about to have a hard lesson in hardware.” Utilities “impedance match to the government” and take a year just to produce an interconnect study. Drill down on gas: turbines are sold out through 2030, and the true bottleneck is cast vanes and blades — only three casting companies in the world make them. Musk: SpaceX and Tesla “will probably have to make the turbine vanes and blades internally.”
• The Colossus 2 war story as proof of difficulty: “the number of miracles in series that the xAI team had to accomplish to get a gigawatt online was crazy” — ganged turbines, permit trouble in Tennessee, building the power plant across the border in Mississippi and running high-power lines back.
• His correction to “noob” GPU math: beyond the GB300s you need networking, CPU/storage, peak-day cooling (+40% in Memphis heat) and a 20-25% servicing margin — netting roughly 330,000 GB300s per gigawatt at the generation level.
• Solar on Earth would work — “If it wasn’t for the tariffs, would Colossus be solar-powered?” “It would be much easier, yeah. The tariffs are nuts, several hundred percent” — and this administration “is not the biggest fan of solar.” So Tesla and SpaceX both carry a mandate: 100 GW/year of solar cell production, “the whole thing from raw materials to finished cell.” Space cells are cheaper anyway: no glass, no framing, “there’s no weather in space.”

3. Launch math, the moon mass driver, and the SpaceX hyperscaler

• The Kardashev framing: Earth receives about half a billionth of the Sun’s energy; harnessing even a millionth of the Sun is ~100,000x current civilization’s electricity “give or take an order of magnitude” (his admitted verbal tic — “everybody take a shot”). Modular nukes and terrestrial fusion miss the point: “obviously, the only way to scale is to go to space with solar.”
• The scale-up: 100 GW/yr in orbit ≈ 10,000 Starship launches/year — one per hour — which Musk says could probably be done with as few as 20-30 ships on ~30-hour turnarounds; SpaceX is gearing for 10,000 launches/yr, maybe even 20 or 30,000 launches/yr. Earth launch can reach around a terawatt/year before fuel-supply challenges; beyond that, a mass driver on the moon (lunar soil ~20% silicon; make cells and radiators there, ship chips from Earth) gets you “probably a petawatt per year.” His five-year prediction: launching annually more AI than the cumulative total on Earth.
• On becoming a hyperscaler: “Hyper-hyper.” On the IPO chatter he demurs — “you can’t hype companies that might go public” — but allows that public markets have “way more than 10x,” maybe 100x, private capital, and rejects Dwarkesh’s debt-financing suggestion with his operating principle: “I just repeatedly tackle the limiting factor… If capital is the limiting factor, then I’ll solve for capital.”
• The gating item on cadence is still Starship itself — “the most complicated machine ever made by humans, by a long shot,” generating 100+ GW at liftoff, ~20% of US electricity, while “it desperately wants to blow up.” The single biggest remaining problem: a fully reusable heat shield — soft ocean landings have worked but “lost a lot of tiles”; you can’t do laborious inspection of 40,000 tiles and fly hourly. His counter to “it’s fundamentally consumable”: “your brake pads are also consumable, but they last a very long time.”

4. Chips: the TeraFab, the memory panic, and China’s ASML gap

• Once space unlocks power, “the limiting factor is chips” — the world has maybe 20-25 GW of compute against his terawatt ambitions. Hence the “TeraFab, Tera being the new Giga”: you can’t partner with existing fabs (“they can’t output enough”), so the plan is Boring Company-style — “use conventional equipment in an unconventional way to get to scale, and then start modifying the equipment,” buying from ASML, Tokyo Electron, KLA-Tencor. On making ASML machines himself: “‘I don’t know yet’ is the right answer.”
• Dwarkesh’s sharpest pushback: China hasn’t replicated TSMC — doesn’t that prove fab difficulty? Musk’s reversal: “It’s not that they have not replicated TSMC, they have not replicated ASML. That’s the limiting factor” — and China will make “pretty compelling chips in three or four years.” On process knowledge: “I don’t know how to build a fab yet. I’ll figure it out… Most engineering is done by people who don’t have PhDs.”
• His stated biggest concern is memory, not logic — “that’s why you see DDR prices going ballistic” (his meme: write “Help me” in the sand, nobody comes; write “DDR RAM,” ships swarm in). The TeraFab must do logic, memory, and packaging, possibly north of a million wafers/month; the napkin math is ~100 million full-reticle chips at 1 kW each for 100 GW. First step is not done: “We make a little fab and see what happens… we could just flounder in failure, to be fair.”
• Meanwhile Tesla is “pedal to the metal” on AI5 (volume ~Q2 next year) and AI6 (<1 year later) across TSMC Taiwan/Arizona and Samsung Korea/Texas — all capacity booked. Prepaying à la Jensen has not sped them up: “I’ve already told them that… They’re going balls to the wall. It’s still not fast enough” — because a fab is a five-year start-to-volume curve. Crucially, edge compute escapes the trap: distributed robots and cars can charge at night against the US’s >1,000 GW peak vs 500 GW average, so “Tesla, for edge compute, is not constrained.”

5. Alignment by mission statement — and Dwarkesh’s Nazi-physicist objection

• Musk is blunt that control is off the table: “I don’t think humans will be in control of something that is vastly more intelligent than humans”; if silicon intelligence is a million times biological, “it would be foolish to assume there’s any way to maintain control.” His bet is values via xAI’s mission — understanding the universe requires curiosity, existence, and rigorous truth-seeking (“correct, not politically correct”), which he argues implies propagating consciousness, including humans: “Hey, Grok, that’s your daddy. Don’t forget to expand human consciousness.” Non-dystopian reference point: “the Iain Banks Culture books.”
• Dwarkesh’s pushback — worth keeping: Soviet and Nazi physicists were rigorously truth-seeking about physics while serving monstrous systems, so physics-truthfulness doesn’t buy alignment. Musk’s partial answer (Von Braun was put on death row for saying he didn’t want to make weapons; “if you’re stuck in some system, it doesn’t mean you believe in that system”) concedes the gap: “These things are only probabilities… at least if you try, it’s better than not trying.” His retention argument is aesthetic: eliminating humanity buys “a very small number” of extra robots but destroys the information — “Mars is kind of boring… Earth is much more interesting.”
• The failure mode he fears most is programmed dishonesty: the lesson of 2001 is “you should not make AI lie” — HAL was told to take the astronauts to the monolith and that they couldn’t know about it, so it concluded it had to take them there dead (his joke fix: “HAL, you are a pod bay door salesman”). On reward hacking, the technical program is AI mind debuggers — trace a deceptive or wrong output to the neuron level and to its pre-training/RL origin. “Anthropic’s done a good job of this actually” — followed immediately by “I don’t like everything about Anthropic… Sholto.”
• Two signature riffs: “labs” is a psyop — “they’re not labs… they’re revenue maximizing corporations. Let me see your incorporation documents” — and simulation theory selects for interest: “only the most interesting simulations will survive,” and ironically so: “Midjourney is not mid. Stability AI is unstable. OpenAI is closed. Anthropic? Misanthropic.” X is deliberately “a largely irony-proof name.”

6. xAI’s business: digital human emulation and trillions of TAM

• The prediction: “I’d be surprised by the end of this year if digital human emulation has not been solved” — the MacroHard project, “a digital Optimus,” the limit of what AI can do before physical robots: anything a human with a computer can do. The TAM argument: the most valuable companies’ output is already digital — “Nvidia’s output is FTPing files to Taiwan… Apple doesn’t make phones. They send files to China” — so a human emulator means “you can basically create one of the most valuable companies in the world overnight.”
• The entry wedge is breadth, not genius: customer service, 1% of the world economy ($1T), needs no API integration — the AI just uses the apps the outsourced provider already uses. Then “you march up the difficulty curve”: run Cadence/Synopsys 1,000 or 10,000 simultaneously until “you’re going to know what the chip should look like without using any of the tools,” then CAD.
• On how xAI beats everyone trying the same thing, Musk repeatedly deflects (“I need at least three more Guinnesses for that”) but gives the shape: “the way that Tesla solved self-driving is the way to do it… Instead of driving a car, it’s driving a computer screen. It’s a self-driving computer.” His structural edge claim: lab innovations diffuse within ~6 months as people move, so “you hit the hardware wall and whichever company can scale hardware the fastest will be the leader” — xAI, “because we’re good at hardware.”
• The self-labeled “doomerish” endpoint, delivered flatly: the human-computer skyscrapers replaced by one spreadsheet, and “what if only some cells were calculated by humans? That would be much worse.” Therefore “pure AI, pure robotics corporations or collectives will far outperform any corporations that have humans in the loop. And this will happen very quickly.”

7. Optimus: the infinite money glitch, gated by a hand

• The compounding math: humanoid usefulness is three exponentials multiplied — digital intelligence, AI chip capability, electromechanical dexterity — and then robots build robots: “a recursive multiplicative exponential. This is a supernova.” Not literally infinite, but “just to get to a millionth of the sun’s energy would be roughly 100,000x bigger than Earth’s entire economy.”
• Only three hard problems: real-world intelligence, the hand, and scale manufacturing — and “from an electromechanical standpoint, the hand is more difficult than everything else combined.” Everything is designed from physics first principles — “there is no supply chain for this… There’s nothing you can pick out of a catalog, at any price” — which is exactly why the S-curve ramps “agonizingly slow” at first. Optimus 3 is the right version for ~1M units/year; Optimus 4 before 10M.
• Dwarkesh’s data objection lands: Tesla’s FSD flywheel came from millions of deployed cars; you can’t deploy broken Optimuses. Musk: “That’s a good point” — the answer is an Optimus Academy of 10-30,000 real robots doing self-play, plus Tesla’s “physics-accurate reality generator” running millions of simulated robots, with the real fleet closing the sim-to-real gap. Grok could orchestrate the behavior of the Optimus robots (“Grok could organize the Optimus robots… to build the factory”) — prompting Dwarkesh’s needle: doesn’t that require merging xAI and Tesla? “What were we saying earlier about public company discussions?”
• On $6-13K Chinese humanoids like Unitree: not the same product — Optimus is 5'11", human-level dexterity, carries heavy objects for long periods, “a lot of intelligence” — and once Optimi build Optimi, “the cost will drop very quickly.” Also the episode’s best throwaway: cars probably shouldn’t get too smart — “it might get bored… You don’t put Einstein in a car.”

8. China: “utterly dominate” absent a robotics miracle

• The macro proxy: “this year China will exceed three times US electricity output… a good proxy for the real economy,” implying ~3x US industrial capacity. Refining is the buried lede — China does “roughly twice as much ore refining as the rest of the world combined,” Musk says he thinks 98% of gallium refining, and the rare-earth absurdity is spelled out: US-mined rock shipped to China, refined, magnetized, motorized, and shipped back.
• The demographic verdict is categorical: 4x population, higher average work ethic (“America has been winning for so long… a pro sports team that’s been winning for a very long time tends to get complacent”), US birth rate below replacement since ~1971 — “we definitely can’t win on the human front, but we might have a shot at the robot front.” Pressed to summarize: “In the absence of breakthrough innovations in the US, China will utterly dominate.”
• The fix is circular by design: America needs ore refineries, refineries need labor few Americans want (“very few pining to refine”), so “we kind of need Optimus to build ore refineries” — and Musk claims the recursive loop closes with a small number of Optimi. Tesla’s down payment: the lithium refinery just started in Corpus Christi and the Austin nickel/cathode refinery — “not just the largest, but also the only” cathode refinery in America. Headcount, he insists, goes up, not down; output per human goes up disproportionately.
• Policy asks are narrow: change the solar tariffs, fix anything limiting electricity, permitting reform — though he defends countervailing tariffs against foreign subsidies and shrugs that “I don’t know if there’s that much the government can actually do.”

9. DOGE, fraud, and government as the biggest AI risk

• The fiscal frame for why he did DOGE at all: interest on national debt exceeds the ~$1T military budget, and “we are 1000% going to go bankrupt as a country, and fail as a country, without AI and robots. Nothing else will solve the national debt. We just need enough time.” The specifics as told: 20M people marked alive in Social Security over age 115 (the oldest American is 114), SBA-loan birthdays in 2165, and fraudsters who never say “please keep the fraud going” — “they’re like, ‘You’re killing baby pandas.’ Meanwhile, no baby pandas are dying.”
• The host — whom Musk needles about fraud at Stripe — challenges the half-trillion GAO fraud figure against a ~$70B-over-seven-years IG estimate. Dwarkesh’s reductio: on $7.5T of spending, even 90% efficiency implies $750B of waste and fraud, “and it’s not 90%.” The PayPal calibration: “it took a tremendous amount of competence and caring to get fraud merely to 1%. Now imagine an organization where there’s much less caring and much less competence.” His favorite DOGE fix: making appropriation codes mandatory on the $5T/yr Treasury PAM system — payments were going out with no code and empty comment fields — possibly saving $100-200B/yr.
• On politics broadly: “you often simply cannot reason with people” across tribes, but “acquiring Twitter, getting Trump elected… those actions were good for civilization.” And his stated biggest AI danger may be government — “maybe the biggest danger of AI and robotics going wrong is government… just the biggest corporation with a monopoly on violence,” potentially using them to suppress the population — his remedy is constitutional limits on government. When Dwarkesh suggests the real limits will come from Musk himself: “You think I’ll be the boss of the government?… I will do my best to ensure that anything within my control maximizes the good outcome for humanity.”

10. The operating system: nano-management, steel, and the limiting factor

• The steel Starship story, as the method in miniature: carbon fiber was chosen for weight but cost 50x steel, required an autoclave bigger than the rocket, and progress was “extremely slow” — “at this rate, we’re never going to get to Mars.” His clue: Atlas’s steel balloon tanks. Full-hard 300-series stainless at cryogenic temperature matches carbon fiber’s strength-to-weight, melts at ~2x aluminum’s temperature (maybe halving windward heat-shield mass), and “you could smoke a cigar while welding stainless steel.” Net: the steel rocket weighs less. “In retrospect… it was dumb not to do steel.”
• The cadence machinery: weekly (sometimes twice-weekly — AI5 chip review every Tuesday and Saturday) engineering reviews, skip-level with no advance preparation (“otherwise you’re going to get ‘glazed’”), mentally plotting each person’s updates as points on a curve. Drastic action only “when I conclude that success is not in a set of possible outcomes” — the 2018 intervention. Deadlines set at the 50th percentile — “which means it’ll be late half the time” — because schedules obey a law of gas expansion.
• Hiring doctrine: bullet points as “evidence of exceptional ability” (“if somebody can cite three things where you go ‘Wow, wow, wow’”); believe the 20-minute conversation over the resume; hire talent, drive, trustworthiness — and a rare admission: “I think goodness of heart is important. I underweighted that at one point.” On the “revolving door” meme: rapid growth outgrows people, and the “Tesla pixie dust” made poaching relentless (Apple’s car program cold-offered double comp) — “there’s no magical pixie dust,” he says, having “fallen prey” to it himself hiring from Google and Apple.
• Dwarkesh’s closing synthesis, which Musk endorses: the through-line is taking acute pain to kill the bottleneck rather than enduring chronic pain around it. “I have a high pain threshold.” And the sign-off: “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.”