How Elon plans to launch a terawatt of GPUs into space.
## Elon Musk plans to launch a massive computing power of 1 terawatt of GPUs into space to advance AI, robotics, and make humanity multi-planetary, while ensuring responsible use and production. ## ## Questions to inspire discussion.
Space-Based AI Infrastructure.
Q: When will space-based data centers become economically superior to Earth-based ones? A: Space data centers will be the most economically compelling option in 30–36 months due to 5x more effective solar power (no batteries needed) and regulatory advantages in scaling compared to Earth.
☀️ Q: How much cheaper is space solar compared to ground solar? A: Space solar is 10x cheaper than ground solar because it requires no batteries and is 5x more effective, while Earth scaling faces tariffs and land/permit issues.
Q: What solar production capacity are SpaceX and Tesla planning? A: SpaceX and Tesla plan to produce 100 GW/year of solar cells for space, manufacturing from raw materials to finished cells in-house.
Q: How much AI capacity could SpaceX launch to space within 5 years? A: In 5 years, SpaceX could launch hundreds of gigawatts per year of AI to space, surpassing Earth’s total AI capacity.
Q: How many Starships are needed to achieve 10,000+ launches per year? A: SpaceX will need 20–30 Starships to achieve 10,000+ launches/year for space AI, which is lower frequency than airlines.
⚖️ Q: What are the three requirements to launch a terawatt of GPUs into space? A: To launch 1M tons to orbit in 3–4 years at 100kW/ton, need to match mass to orbit, 100GW/year of solar power generation, and 100GW worth of chips.
Manufacturing Bottlenecks.
Q: What is the limiting factor for scaling power generation for space AI? A: Turbine blades and vanes are the limiting factor, with only 3 casting companies producing these components and they are massively backlogged.
Q: What manufacturing approach will SpaceX and Tesla take for turbine components? A: SpaceX and Tesla may need to internally manufacture turbine blades and vanes to scale power for space AI due to supplier constraints.
Q: What is the biggest concern for supporting logic chips in space? A: Memory is the biggest concern, with DDR prices skyrocketing, requiring sufficient memory to support terawatt-scale GPU deployment in space.
Chip Manufacturing at Scale.
Q: What new manufacturing concept is needed for space AI chips? A: Scaling AI in space requires terawatt-scale chip fabs (TeraFab) for cheap, high-volume chips to support massive deployment.
Q: Which companies must SpaceX partner with for chip manufacturing equipment? A: SpaceX will need to partner with ASML, Tokyo Electron, and KLA-Tencor (the few companies dominating fab equipment) to increase volume and modify equipment for higher production rates.
Q: What chip production volume is needed by 2030 for TeraFab? A: Musk’s TeraFab plan involves producing millions of advanced chips per month by 2030, requiring 100 million chips running at 1 kilowatt each to achieve 100 gigawatts of power.
Q: How will Dojo 3 chips be optimized for space deployment? A: Dojo 3 will be designed for space-based AI compute with chips that are more radiation tolerant and run at higher temperatures to reduce radiator mass by up to 50%.
Q: What makes lunar manufacturing of solar-powered AI satellites feasible? A: Lunar manufacturing is feasible due to 20% silicon and sufficient aluminum in lunar soil for solar cells and radiators, with chips initially sent from Earth but potentially made on Moon later.
AI Development and Alignment.
Q: What debugging capability is xAI planning to develop? A: xAI plans to develop debuggers to trace AI mistakes to the neuron level, identifying bugs or deceptions in AI reasoning and their origins in pre-training, mid-training, post-training, or reinforcement learning phases.
Q: What principle should guide Grok AI development? A: Grok AI should be rigorously truth-seeking to discover new physics and invent working technologies, ensuring correct, non-contradictory axioms with cogent conclusions adhering to critical thinking principles.
Q: When will digital human emulation be solved? A: Digital human emulation is expected to be solved by the end of 2026, enabling AI to do anything a human with a computer can do, amplifying human productivity until physical robots are developed.
Business Applications.
Q: What is the most immediate revenue opportunity for AI corporations? A: Customer service represents a trillion-dollar revenue stream that can be solved by emulating a human at a desktop, involving average intelligence and no barriers to entry.
Q: How will digital corporations outperform human-based ones? A: Pure AI and robotics corporations will vastly outperform those with humans in the loop by taking existing customer service apps and doing work at a fraction of the cost with no integration needed.
⚡ Q: What advantage do digital corporations have over traditional ones? A: Digital corporations will outperform human-based ones because AI can do anything involving moving electrons or amplifying human productivity, with no barriers to entry.
Optimus Robot Development.
Q: How many Optimus robots will Tesla build for training? A: Tesla plans to build 10-30K Optimus robots for self-play learning in reality, using physics-accurate simulators to close the sim-to-real gap and train the Optimus mind.
Q: What is Optimus 3’s production capacity? A: Optimus 3 can produce 1M units/year, but initial production will be a stretched S-curve due to custom-designed components not based on existing supply chains.
⏰ Q: What operational advantage do Optimus robots have? A: Optimus robots will be best for 24/7 continuous operations in homes and factories, increasing Tesla’s output per human without reducing headcount.
Q: What capabilities is Optimus 3 designed for? A: Optimus 3 is designed for human-level intelligence and dexterity, enabling it to carry heavy objects for long periods without overheating or exceeding actuator power.
Q: What three factors drive humanoid robot improvement? A: Humanoid robots will improve exponentially through AI intelligence, AI chip capability, and electromechanical dexterity, with their usefulness being the product of these three factors, multiplied recursively.
Industrial Applications.
Q: How can Optimus robots help US manufacturing competitiveness? A: Optimus can build ore refineries in the US to increase refining capacity and competitiveness, as China currently dominates refining and manufacturing.
Q: How can Optimus robots enable space-based AI scaling? A: Optimus robots can help scale AI in space by building factories and launching 1M tons/year to orbit, enabling real-world AI and space-based economies.
Power and Compute Distribution.
⚡ Q: What is the key difference between edge and concentrated compute power constraints? A: Edge compute with Tesla’s AI5 chip in Optimus robots is not constrained by power because power is distributed over a large area, allowing more effective grid usage by charging at night, while concentrated compute in space is limited by electricity.
Management and Execution.
Q: What traits should you prioritize when hiring? A: Hire for talent, drive, trustworthiness, and goodness of heart as these fundamental traits are hard to change and more important than domain knowledge, which can be added later.
⏱️ Q: What deadline strategy maintains urgency? A: Set aggressive, achievable deadlines at the 50th percentile probability to maintain a maniacal sense of urgency and focus on addressing the limiting factor to keep the company moving fast.
Q: How should engineering reviews be conducted at scale? A: Conduct detailed engineering reviews weekly or bi-weekly with all relevant engineers, focusing on the limiting factor, to maintain deep understanding of progress and bottlenecks even as the company scales.
⚡ Q: When should you make drastic organizational changes? A: Make drastic changes when it’s clear that success is impossible without them, as Elon did in 2018 to fix a critical problem, based on thorough understanding of the situation.
Design Philosophy.
Q: What materials and processes enable rapid iteration for complex machines? A: Use simple materials and processes like stainless steel and welding for the most complex machines, such as Starship, to enable rapid iteration and scaling while maintaining focus on fundamentals.
♻️ Q: What design principles enable scaling of complex machines? A: Prioritize reusability and simplicity in design and manufacturing to enable rapid iteration and scaling, even for the most complex machines like Starship, which is the most complicated machine ever made by humans.
Government and Economic Impact.
Q: What does Elon believe is the only solution to the US debt crisis? A: Elon Musk believes AI and robotics are the only solution to the US national debt crisis, which exceeds $1 trillion in annual interest payments, surpassing the military budget.
Q: What simple government change could save $100–200 billion annually? A: The DOGE team implemented mandatory appropriation codes in payment fields, which could save the US government $100–200 billion annually by reducing fraud.
⚠️ Q: What is the greatest risk of AI misuse according to Musk? A: Elon Musk argues that the government, as the biggest corporation with a monopoly on violence, poses the greatest risk of misusing AI and robotics to suppress the population.
Q: What is Musk’s strategy to maximize good outcomes for humanity? A: To maximize the good outcome for humanity, Musk aims to ensure that anything within his control, including AI and robotics, aligns with pro-human values and limits government overreach.
## Key Insights.
Space-Based AI Economics.
Space delivers.
In this episode, John and I got to do a real deep-dive with Elon. We discuss the economics of orbital data centers, the difficulties of scaling power on Earth, what it would take to manufacture humanoids at high-volume in America, xAI’s business and alignment plans, DOGE, and much more.
Transcript: https://www.dwarkesh.com/p/elon-musk.
Apple Podcasts: https://podcasts.apple.com/us/podcast…
- Mercury just started offering personal banking! I’m already banking with Mercury for business purposes, so getting to bank with them for my personal life makes everything so much simpler. Apply now at https://mercury.com/personal-banking
- Labelbox can get you robotics and RL data at scale. Labelbox starts by helping you define your ideal data distribution, and then their massive Alignerr network collects frontier-grade data that you can use to train your models.
Spotify: https://open.spotify.com/episode/4nah…Mercury just started offering personal banking! I’m already banking with Mercury for business purposes, so getting to bank with them for my personal life makes everything so much simpler. Apply now at https://mercury.com/personal-banking.
Jane Street sent me a new puzzle last week: they trained a neural net, shuffled all 96 layers, and asked me to put them back in order. I tried but… I didn’t quite nail it. If you’re curious, or if you think you can do better, you should take a stab at https://janestreet.com/dwarkesh.
Labelbox can get you robotics and RL data at scale. Labelbox starts by helping you define your ideal data distribution, and then their massive Alignerr network collects frontier-grade data that you can use to train your models.To sponsor a future episode, visit https://dwarkesh.com/advertise.
0:00:00 — Orbital data centers.
0:36:46 — Grok and alignment.
0:59:56 — xAI’s business plan.
1:17:21 — Optimus and humanoid manufacturing.
1:30:22 — Does China win by default?
1:44:16 — Lessons from running SpaceX
2:20:08 — DOGE
2:38:28 — TeraFab