SpaceX has quietly repositioned itself as a neocloud, the compute-rental business model pioneered by hyperscalers like AWS but now pursued by companies with stranded GPU capacity. AI revenue tripled year-over-year to $2.6 billion, surpassing the company’s legacy space segment at $962 million, driven almost entirely by selling compute to other AI firms rather than by its own Grok model. Starlink connectivity at $4.2 billion remains the only profitable segment. The company is still losing money overall, though the quarterly loss narrowed to $143 million. The SpaceX earnings picture<\/a> is stranger than it looks.<\/p>\n The interesting thing about SpaceX’s AI revenue number isn’t its size, it’s its origin. SpaceX built massive GPU clusters for Grok, Grok fell behind, and rather than write down the investment, the company flipped the capacity into a third-party compute business. That pivot from “internal AI infrastructure” to “rented AI infrastructure” is now SpaceX’s fastest-growing revenue line. Any enterprise CTO who has over-provisioned GPU capacity for a model project that didn’t scale should recognize the pattern, because the arbitrage is running in the other direction for buyers right now.<\/p>\n SpaceX losing $1.5 billion in its AI division while generating $2.6 billion in revenue signals that the unit economics of GPU rental at scale are still deeply underwater. This isn’t unique to SpaceX. CoreWeave, Lambda, and other neoclouds have all demonstrated that buying Nvidia hardware at list price and renting it out doesn’t naturally produce margins, especially when the hyperscalers are simultaneously dropping their own inference pricing. The companies renting from SpaceX are getting a deal today precisely because SpaceX needs the revenue to offset stranded capital. That arbitrage window closes as SpaceX optimizes pricing or finds its own model workloads to fill the capacity.<\/p>\n The pending Cursor acquisition for $60 billion is the more durable strategic signal. Selling raw compute is a commodity play. Owning a developer tool used by millions of engineers gives SpaceX a distribution wedge into enterprise software teams, a direct consumption surface for its own compute, and a feedback loop for training data. If you’re currently renewing a Cursor enterprise contract or evaluating AI coding tools for your engineering org, the question isn’t whether Cursor’s product changes immediately after close. The question is whether you’re comfortable with SpaceX, and by extension Elon Musk, sitting inside your development workflow. That’s a vendor risk conversation that procurement hasn’t fully priced yet.<\/p>\n A neocloud is a company that buys or builds large GPU clusters and rents that compute to AI developers, functioning like a datacenter landlord rather than a software vendor. Unlike AWS or Azure, which bundle storage, networking, and software services, neoclouds typically sell raw GPU hours at scale. They exist because Nvidia supply constraints made it faster to rent than to procure, and because AI labs needed burst capacity their own infrastructure couldn’t cover. The business risk is that margins depend entirely on utilization rates staying high.<\/p>\nWhat this means for your business<\/h2>\n
Concept deep-dive: Neocloud<\/h2>\n