On August 4, Elon Musk sat down for the first earnings call in SpaceX’s history and made a claim that startled even the analysts who follow the company’s AI ambitions. He said SpaceX planned to build and deliver six to eight gigawatts of new artificial-intelligence computing capacity in 2027, called that figure “conservative,” and floated a tentative target of twenty gigawatts of power and cooling online by the end of next year. Even with delays, he said, the company would likely land close to fifteen.
Three days later, the research firm SemiAnalysis published a report arguing that the target was real. The firm projected SpaceX would reach roughly ten gigawatts of compute by the end of 2027, up from about 1.4 gigawatts today, and modeled an annualized revenue run rate of about $305 billion, with $235 billion of that coming from AI compute alone. At the firm’s assumed cost of about $50 billion per gigawatt, the build-out implies $300 billion to $500 billion in capital spending next year, on par with what Amazon Web Services and Google are expected to spend.
The report’s most striking claim involved Microsoft. SemiAnalysis said Microsoft could sign a roughly three-gigawatt, $150 billion lease with SpaceX, a contract that would make the software giant the largest buyer of compute from a company that, until this year, was best known for rockets and satellites. By August 9, multiple outlets reported that Microsoft plans to lease more than three gigawatts of AI data center capacity from SpaceX beginning in 2027.
Microsoft’s logic is straightforward. In April, it renegotiated its deal with OpenAI, dropping a 20 percent revenue share it had paid under the old arrangement. The company can now sell access to the same frontier models it helps host, earning comparable revenue per megawatt while paying none of the training costs. SemiAnalysis estimates that Microsoft has signed more than ten gigawatts of binding data center contracts this year, worth roughly $300 billion, a burst of activity that reverses a leasing pause the firm first flagged in December 2024. The driver, the firm says, is a $250 billion infrastructure agreement with OpenAI signed in October 2025, estimated at about seven gigawatts, which left Microsoft short of capacity for its own products, including the Azure Foundry API business and Copilot applications.
SpaceX enters that picture with an unusual advantage: it already holds the keys to xAI, which it acquired in February, bringing the Grok models and the Colossus data center campuses under one roof. Google has signed a compute agreement with SpaceX reported at about $920 million a month. Anthropic took over the entire footprint of the Colossus 1 data center, gaining more than 300 megawatts and over 220,000 Nvidia GPUs under a lease with a 90-day cancellation clause. SpaceX, in effect, is now the landlord for two hyperscale cloud providers at once.
The math that makes these contracts plausible runs through token economics. SemiAnalysis calculates that OpenAI and Anthropic can generate more than $100 million per megawatt per year, or over $100 billion per gigawatt per year, when they sell API inference on clusters built around Nvidia’s GB300 systems. That revenue potential dwarfs the cost of renting such a cluster at current neocloud prices, which is why labs keep signing leases that would have seemed reckless two years ago.
The arrangement also carries risks. A 90-day cancellation policy protects tenants more than the landlord; if token prices fall or models improve faster than expected, customers can walk away. Execution is the harder question: building ten gigawatts of compute in one year means securing power, cooling and equipment at a pace no company has managed at this scale. Musk compared the challenge to the engineering of SpaceX’s core launch business, and SemiAnalysis notes the company has the balance sheet and supply-chain relationships to try.
For the broader industry, the shift is structural. Hyperscalers can no longer build fast enough on their own, and a new class of intermediaries, neoclouds and now a rocket company, has emerged to sell capacity at a premium. SemiAnalysis described the Microsoft opportunity as “once-in-a-generation,” a phrase analysts rarely use. Whether the projections hold depends on assumptions about GPU availability, power access and the price of tokens a year from now.
What is clear is that the economics of AI infrastructure have redrawn the map of the industry. The company that lands rockets on drone ships is now, by SemiAnalysis’s reckoning, on track to become the largest owner of AI compute outside the hyperscalers themselves. For Microsoft, the deal would be a hedge against its own construction constraints; for SpaceX, it is a chance to turn a rocket company into the power broker of the AI boom. Neither side has confirmed the terms, and people familiar with the matter caution that negotiations remain fluid. But the direction is no longer in doubt: the landlord of AI has arrived, and its name is SpaceX.
