SpaceX targets 10GW AI compute by 2027, Microsoft weighs $150B deal

SpaceX plans to build and deliver 6 to 8 gigawatts of AI compute in 2027, with upside above 10 gigawatts, requiring $300 billion to $500 billion in capital spending that rivals the outlays of AWS and Google. Elon Musk announced the target at SpaceX's first earnings call, calling the 6 to 8 gigawatts a "conservative" goal.

SemiAnalysis, which tracks datacenter leasing and power equipment quarter by quarter, said the target is achievable. "SpaceX will develop anything they can and bring it online as fast as possible," the firm wrote in a report, noting that large-scale near-term compute is scarce and priced at a premium of up to $50 billion per gigawatt per year.

The economics rest on AI inference margins. SemiAnalysis estimates OpenAI and Anthropic can generate more than $100 billion per gigawatt per year selling API inference on Nvidia GB300 clusters, against roughly $12 billion in annual rental costs at $3 per GPU-hour — gross margins above 60 percent, and more than 85 percent on flagship models. Microsoft, which renegotiated its OpenAI deal in April 2026 to drop a 20 percent revenue share, could monetize the same models at similar rates while paying none of the training costs.

SemiAnalysis sees a potential 3-gigawatt, $150 billion contract between Microsoft and SpaceX, which would help Azure revenue growth accelerate from about 42 percent to more than 100 percent. The firm called the opportunity "once-in-a-generation."

Microsoft's 10GW awakening

Microsoft has signed more than 10 gigawatts of binding contracts this year across leasing, self-build, and long-term power agreements, worth roughly $300 billion, according to SemiAnalysis. The reversal follows a pause in leasing activity the firm flagged in December 2024. A $250 billion infrastructure deal with OpenAI signed in October 2025, estimated at about 7 gigawatts, left Microsoft compute-constrained on its Foundry API business and Copilot applications — the services with the highest revenue per megawatt.

Speed over efficiency

SpaceX's edge is build speed. Colossus 1, its 300-megawatt cluster in Memphis, went from scratch to operation in 122 days. The Southaven power plant expanded from 27 gas turbines, about 495 megawatts, in February 2026 to 69 turbines, more than 1.2 gigawatts, by July. A greenfield project dubbed "MiniHard" is expected to reach 450 to 500 megawatts in about five months from March 2026. SpaceX has also bought Chinese power modules to bypass a two-year backlog in large transformers and sourced gas turbines on the secondary market, including units originally destined for an Oracle site in New Mexico.

Financing comes from two sources, SemiAnalysis said: Nvidia supplier financing, which likely explains Musk's commitment to Nvidia's Vera Rubin architecture exclusively, and premium pricing of $30 million to $50 million per megawatt per year that pays back capital spending in under a year. Assuming half of 2027's new compute is monetized for inference, with the rest reserved for Grok and Cursor training, SpaceX's annual recurring revenue could reach $300 billion by the end of 2027.

For investors, the buildout tests whether SpaceX can sustain a pace no hyperscaler has matched. Microsoft, trading on the strength of its OpenAI access, stands to gain the most if a deal closes; OpenAI and Anthropic face a new landlord with pricing power. Nvidia, meanwhile, locks in a customer that could absorb tens of gigawatts of Vera Rubin silicon.

This article is for informational purposes only and does not constitute investment advice.