🛰️ GPU in Space - Madness or the Future of AI?
Dwarkesh Patel wrote an analysis of the idea of orbital data centers (preparing for an interview with Elon Musk).
Arguments FOR:
- In space, the sun shines 100% of the time (on Earth ~25% due to night/clouds)
- No need for batteries - the main cost of terrestrial solar energy
- Earth simply won't be able to scale energy for AI: connection queues, permits, turbines booked until 2030+
Arguments AGAINST:
- GPUs account for 70% of the data center cost, energy is only 15%
- The Meta Llama 3 cluster (16K GPUs) crashed every 3 hours - and in space, there's no one to fix it
- To achieve 100 GW of orbital power, 10,000 Starship launches are needed per year (one per hour)
Elon Musk's bet: if orbital data centers work - xAI will get unlimited energy and win the AI race (SpaceX has a unique advantage in launches). If not - xAI will just become another lab.
Interestingly, is this a real bet or a hype before the IPO of the merged SpaceX + xAI?
🔗 dwarkesh.com/p/notes-on-space-gpus
P.S. I thought that this orbit (on the edge of light and shadow) must be the busiest (now or will be)
#ai #spacex #xai #elon_musk #dwarkesh_patel