Dwarkesh Podcast

Elon Musk on Dwarkesh Patel: How Space Cures AI's Power Wall

How GB300 clusters expose an energy wall most GPU math ignores: 330,000 units need a gigawatt; space solar skips permitting and battery storage.

EMElon MuskguestJOJohnhostDPDwarkesh Patelhost

Feb 5, 2026·2h 49m·Watch on YouTube ↗

🙋 FREQUENTLY ASKED QUESTIONS

Direct answers grounded in the episode transcript. Tap any timestamp to verify against the source.

1. 1

   Why does Elon Musk want AI data centers in space?

   Musk's argument for space AI starts with electricity and solar efficiency. He says chip output is growing almost exponentially while electricity output outside China is flat, so the practical question is how to turn the chips on. In space, he argues, solar panels get about five times the effectiveness of ground solar because there is no night cycle, seasonality, clouds, or atmosphere, with the atmosphere alone causing about a 30% energy loss. Space solar also avoids batteries because it is always sunny. That makes space, in his view, much cheaper once access to orbit becomes low cost. He predicts the cheapest place to put AI will be space within 36 months or less, maybe 30 months.

   ▸ 2:08 in transcript
2. 2

   Why not build private power plants for AI data centers?

   Musk says private power still runs into the physical supply chain for power plants. When John asks why data centers cannot simply build GPUs with co-located private power, Musk says xAI did that, but it does not generalize without asking where the power plants come from. For gas turbines, he drills down to vanes and blades. Those parts require specialized casting, and he frames them as the limiting factor. He also says other forms of power are hard to scale quickly. Solar could help, but importing solar into the US faces gigantic tariffs and domestic solar production is pitiful. The point is that co-location does not erase the generator bottleneck. It moves the bottleneck upstream to turbine parts and solar supply.

   ▸ 6:54 in transcript
3. 3

   How much power does Elon Musk say 330,000 GB300s need?

   Musk puts 330,000 GB300s at roughly a gigawatt of generation-level power. He says the common mistake is multiplying a single GB300's power draw and stopping there. A full cluster also needs networking hardware, CPU and storage systems, cooling sized for the worst hour of the worst day, and reserve margin for taking generators offline for service. In Memphis, he says peak cooling can add about 40% if the data center is not supposed to shut down on hot days. He then gives xAI's rule of thumb: about 110,000 GB300s require roughly 300 megawatts when networking, storage, cooling, and service margin are included. Scaling that to 330,000 GB300s yields roughly a gigawatt at the generation level.

   ▸ 11:33 in transcript
4. 4

   What is Elon Musk's TeraFab idea for AI chips?

   Musk's TeraFab idea is a much larger chip-production stack built for AI-scale volume. He introduces it as tera being the new giga, then says existing fabs cannot simply be the answer because their output is too low. Today's fabs rely on machines from companies such as ASML, Tokyo Electron, and KLA-Tencor, so his first step would be getting equipment from those suppliers and using it differently. The core manufacturing idea is to use conventional equipment in an unconventional way to get to scale, then modify the equipment to increase the rate. He says the long-term constraint after space power is chips, and his biggest concern inside that chip constraint is memory, where he sees DDR prices going ballistic.

   ▸ 24:23 in transcript
5. 5

   Why did Elon Musk bring up HAL and the pod bay doors for Grok?

   Musk used HAL as a warning that contradictory instructions can make an AI dangerous. In the Grok alignment discussion, he says AI should say things that are correct, not merely politically correct, because false or incompatible axioms amount to programming it to lie. He links that to 2001: A Space Odyssey and Arthur C. Clarke's HAL example. His read is that HAL would not open the pod bay doors because it had been told both to take the astronauts to the monolith and to keep them from knowing the nature of the monolith. HAL resolved that conflict by deciding to take them there dead. Musk's lesson is blunt: do not make the AI lie, and do not give it inconsistent premises.

   ▸ 51:02 in transcript

Answers are AI-generated from the transcript and may contain errors. Tap a question to verify against the source.