Space GPUs: Power Win, But 10K Launches for 100GW?
Orbital datacenters tap 100% solar capacity in sun-synchronous orbits, beating Earth's 25% factor, but demand 10,000 Starship launches yearly for 100GW amid chip costs and no repairs—viable if SpaceX scales massively.
Unlimited Solar Beats Earth Limits—If Launches Scale
Sun-synchronous orbits deliver near-100% solar capacity factor versus Earth's 25% due to no nights, clouds, or batteries, with 40% higher irradiance. Starship could drop launch costs to $100/kg from Falcon 9's $2,500/kg, making space energy cheaper than terrestrial power plants long-term. However, energy is just 15% of datacenter TCO—chips dominate at 70%, and launching them idle (after ground burn-in to weed out infant mortality) erodes gains. Reliability hurts too: Meta's 16,000 H100s fail every 3 hours on Earth, fixable by technicians; space offers no swaps until robots arrive. Radiation is manageable—Google TPUs endure 3x the 5-year ionizing dose before degradation.
Elon argues Earth can't scale: 1TW (2x US total power) faces grid queues, turbine lead times past 2030, solar permits/tariffs, and land needs (1TW solar = 30,000 sq mi or 1% US, but marginal sites spike costs with litigation/transmission). Critics counter Earth's scale—pave Nevada/UAE—but permitting/politics block it, with 1GW needing 45 sq mi solar+storage, supply curve steepening fast.
85W/kg Satellites Demand Extreme Launch Cadence
Target 100W/kg satellites (vs Starlink's 50W/kg) split mass across solar (200W/kg optimistic), compute (GB200 NVL72 at 1,452W/kg post-overhead), radiators (320W/kg at 60°C via Stefan-Boltzmann; hotter chips save mass), and 25% chassis. Rough math yields 85W/kg total, or 10MW per 150t Starship to LEO—100 launches/GW, 10,000/year for 100GW (1/hour). Elon eyes 20-30 ships cycling every 30 hours for 10k-30k launches/year.
Inference Fits Space, Training Doesn't—Latency Minor
Inter-satellite lasers hit 100Gbps (Starlink) to 10Tbps potential, near Infiniband's 400Gbps, but desyncing orbits limit clusters (Google's Suncatcher: 81 sats at same altitude = 15MW coherent compute). Training stays Earth-side; inference dominates future (RL as inference), beaming 100GW of 5T models (58B tokens/s, 230GB/s) easily via lasers-to-Starlink-to-ground. 50ms latency suits agent workflows needing minutes, not milliseconds.
Elon's Bet: SpaceX Monopoly Wins AI Race
Orbital shift needs Earth power ceiling, chip fab scaling (harder than solar), and Starship volume. If true, SpaceX/xAI monopolize unlimited power; others queue for grids. Reckless? Matches SpaceX playbook—Starlink fueled Falcon 9 reuse, datacenters could do it for Starship en route to Mars/Dyson swarms.