SpaceX will win AI. Might sound like an absurd claim right now, but stay with me. For over a year, I’ve argued that frontier capability gains eventually stop mattering economically once “good-enough AI” is crossed in the agentic era. And models themselves stop being a unique edge. A even more absurd-sounding claim, but keep reading... What is “good-enough AI”? Agentic AI capable of performing nearly all computer work reliably, reaching or exceeding human parity at the same work. Once most frontier labs cross that threshold, two things happen. 1. Commoditization Once frontier companies cross it, open-source models eventually do too, particularly from China. Through distillation, anyone can download one, attach their brand, and become an AI provider. Models become a commodity rather than a unique edge. 2. Welcome to real-world economics. For four years, AI has operated under a future-capability premium that real-world economics could never sustain indefinitely. Once AI is good enough for most work and massively distributed through commoditized models, hundreds of providers compete primarily on cost and speed. And even as model efficiency and intelligence density improve, the way to win ultimately becomes scaling compute. The competition becomes: Who can scale compute fastest, at the greatest scale, with the cheapest energy and highest utilization? That leaves only a few players with credible paths through the energy constraint with these two main: – SpaceX’s orbital bet – China’s national energy scaling and abundance SpaceX’s potential solution is the most radical. If it wins, it would produce more energy than anyone combined at near-zero cost, giving it energy dominance. China obviously keeps competing, but is temporarily limited by its chip constraints, giving SpaceX a short window to build before China reaches its full energy potential. 3. This becomes even more important once AI becomes recursive. Even frontier models rapidly accelerating toward recursive AI eventually face a hard limit. The bottleneck won't be building cost-effective models that undercut Chinese open models. Their AIs eventually spawn smaller, highly efficient models that are better. The actual hard limit becomes energy. While everyone competes to serve users, they’re also competing to feed their AIs’ recursive improvement with compute. Every compute delay increasingly becomes a major setback against competitors. And once terrestrial data-center expansion starts hitting its wall of energy, infrastructure, and policy limits, with populist movements in the West massively slowing it, the challenge becomes building at scale. That pushes the industry toward one novel solution: Orbital compute. And if SpaceX solves it, it now controls the biggest constraint everyone else faces. That is the key to how SpaceX wins. 4. But how does that translate to winning the model race? It can then win the agentic market early with “workhorse models” 90% as good as the frontier at 10× lower cost. But here is the actual core edge: Use compute rent revenue to subsidize its own models. Competitor growth directly funds Grok’s cheap prices, to earn more market share. 5. And that creates an even bigger advantage. Beyond the obvious market-share benefits, this generates immense real-world data to help close the frontier gap. Ultimately, LLMs still come down to better data + more compute. Every dollar generated from competitors can therefore be turned back into compute and data. So we get the ideal flywheel: SpaceX’s energy advantage → compute-rent revenue → cheaper workhorse models → market-share gains → massive real-world data → endless scaling of compute + data → its own bigger recursive loop → frontier catch-up. SpaceX therefore doesn’t need to win the model race now. It needs to win the compute race. And the compute race ultimately comes down to energy. SpaceX solving orbital compute will be the reason it ultimately wins AI.