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

Space as the cheapest place for AI

• Musk predicts that in 30-36 months, space will be by far the cheapest and most scalable place to put AI compute.
• Solar panels are about five times more effective in space (no day-night cycle, clouds, seasonality, or atmosphere, which alone causes ~30% energy loss), and no batteries are needed, making it effectively 10x cheaper.
• Earth's electricity output outside China is roughly flat while chip output grows exponentially, so power—not chips—is the near-term constraint. He expects chip output to exceed the ability to turn chips on toward the end of this year.
• Five years from now, he predicts AI launched to space per year will exceed the cumulative total of all AI on Earth—hundreds of gigawatts per year, requiring roughly 10,000 Starship launches (about one per hour, doable with 20-30 ships).
• Beyond a terawatt per year from Earth, he wants a mass driver on the moon, where lunar soil is ~20% silicon and there is plenty of aluminum for radiators; chips would initially come from Earth.

Power, turbines and solar

• The limiting factor for gas turbines is the vanes and blades; only three casting companies in the world make them and they are massively backlogged. SpaceX and Tesla will probably make them internally.
• Both Tesla and SpaceX have a mandate to reach 100 gigawatts a year of solar cell production, from raw materials to finished cell. Space-bound solar cells are cheaper since they need little glass or framing.
• For xAI's Colossus 2, the team ganged together turbines, hit permit issues in Tennessee, and built the power plant in Mississippi. Roughly 330,000 GB300s, including networking, storage, peak cooling (~40% extra) and servicing margin (~20-25%), need about a gigawatt at generation level.

Chips and the TeraFab

• The long-term chip constraint is memory more than logic; DDR prices are "going ballistic."
• The TeraFab would do logic, memory, and packaging, targeting around a million wafers a month by 2030. He would start with a little fab, make mistakes at small scale, then scale up.
• Tesla is pushing Tesla AI5 into production around Q2 next year, with AI6 hopefully less than a year later, using TSMC Taiwan, Samsung Korea, TSMC Arizona, and Samsung Texas.

Grok, alignment and AI products

• xAI's mission is to understand the universe, which he argues requires truth-seeking, curiosity, existence, and propagating intelligence and consciousness—including human consciousness—into the future.
• He thinks humans will not control AI vastly smarter than them; the best hope is instilling the right values. Making AI lie (e.g., political correctness) risks making it insane, citing HAL in 2001: A Space Odyssey.
• xAI is building fine-grained debuggers to trace AI mistakes or deception to their origin; he credits Anthropic's interpretability work.
• He expects digital human emulation to be solved by the end of this year; a human at a computer is the most AI can do before physical robots. xAI's path is like Tesla's self-driving: "a self-driving computer."

Optimus and China

• Three hard things for humanoids: real-world intelligence, the hand, and scale manufacturing. The hand is electromechanically harder than everything else combined; everything is custom-designed from physics first principles.
• Optimus 3 is the right version for ~a million units a year; robots building robots will drop costs fast. An "Optimus Academy" of 10,000-30,000 robots doing self-play, plus a physics-accurate simulator, will close the sim-to-real gap.
• China does roughly twice the ore refining of the rest of the world combined and will exceed three times US electricity output this year; the US cannot win on humans and needs robotics.

Starship, management and DOGE

• Starship moved from carbon fiber to steel out of desperation: strain-hardened 300-series stainless at cryogenic temperature has similar strength-to-weight at ~50x lower material cost, is easy to weld, and its higher melting point halves windward heat shield mass. The steel rocket ends up lighter.
• Starship is the most complex machine ever made; liftoff generates over 100 gigawatts. The biggest remaining problem is a fully reusable heat shield.
• His management approach: constantly attack the limiting factor, run detailed weekly skip-level engineering reviews with no advance prep, and set deadlines at the 50th-percentile of achievability.
• On DOGE: he cites a GAO estimate of roughly half a trillion dollars of annual fraud, and says a simple fix—requiring appropriation codes on Treasury payments—could save $100-200 billion a year.