Why Elon Wants to Put Data Centers in Space | Ramez Naam

— Energy as the AI bottleneck: Grid connection wait times of 5–7 years have forced data center operators to explore behind-the-meter power solutions, from natural gas turbines to batteries and modular generators, because compute revenue ($20–$40 per dollar spent on energy) justifies premium power costs.

— Orbital and ocean data centers: Space-based solar requires launch costs to drop 4–10x (achievable with Starship if launched multiple times per day), while Pantalassa's floating ocean facilities in Antarctic waters use wave motion to generate power and ocean water for free cooling, bypassing grid permitting altogether.

— Bitcoin miners pivoting to AI: Miners have access to power infrastructure and can generate more revenue per kilowatt in AI compute than Bitcoin mining, making the shift economically rational and concentrating value in those who can route around grid constraints.

— Narrow superintelligence over general AI: AI excels only in formal, highly verifiable domains (math, coding, games) where infinite training data and instant feedback exist; most real-world tasks (writing, policy, business) remain messy and data-limited, making narrow, specialized AI more realistic than AGI.

— Data as the new moat: Proprietary, ongoing data—especially from biotech experimentation or industry-specific workflows—drives sustainable competitive advantage; synthetic data and reinforcement learning are becoming the secret sauce for model improvement rather than raw internet scraping.

— Supply chain and component shortages: Transformers, turbines, and switchboards are sold out 3–7 years in advance; companies like American Consolidated Electric and new entrants are capturing value by solving these bottlenecks, akin to selling picks and shovels in a gold rush.

— Power scarcity is the strategic advantage: If you operate or invest in infrastructure with access to cheap, reliable power (whether deregulated grids like Texas, behind-the-meter solutions, or alternative sources), you control a key bottleneck in the AI era and can capture outsized returns.

— Explore model routing and open-source stacks: As frontier models commoditize, efficiency gains come from routing queries to the right model (cheaper or faster for simple tasks, frontier for complex ones) and leveraging open-weight models like NVIDIA's Nemotron; custom solutions are emerging but the market is still nascent.

— Build proprietary data pipelines, not just consume data: One-time archive digitization has limited value; sustainable edge comes from continuous data generation (biotech experimentation, operational logs, medical workflows) fed back into AI training—companies like New Limit and Mercor are validating this loop.