SpaceX will deploy 10 gigawatts of compute by the end of 2027, according to Jeremy, a lead analyst at SemiAnalysis, speaking on the SemiAnalysis podcast with co-author Rake. The thesis rests on frontier AI labs extracting $100 million per megawatt per year from 2026-class hardware by selling inference tokens. OpenAI and Anthropic are adding close to $30 billion of monthly run-rate revenue, an annualized pace near $400 billion per year, without fundamentally changing API prices. A lab clearing 85% gross margins will buy any capacity it can switch on right now. The firm estimates SpaceX will monetize enough capacity to approach $300 billion in annualized revenue, with Microsoft as its largest single offtaker. The episode was recorded shortly after SpaceX's first public earnings call. Jeremy's estimate uses two methods: InferenceX, real-workload testing of open-source models on GB200 and GB300 hardware, and a hardware simulator. Chip generation forecasts show GB200 at $73.4 million per MW per year, GB300 at $99.7 million, and Vera Rubin, SpaceX's 2027 fleet, above GB300 level. Leaked lab financials report gross margins of 85%, implying roughly $100 billion of sellable output per gigawatt against a cost of about $15 billion per gigawatt. The analysts also examined a fifth link: whether models get shut down before the buildout pays off.