Will SpaceX Supply AI Computing Power to Microsoft? Aiming for 10GW by 2027 End, Targeting $300 Billion Annual Revenue

SpaceX (SPCX-US) is quietly transforming from a rocket and satellite communications company into a major player in large-scale computing, with expansion plans far exceeding market expectations. Research firm Semianalysis reports that SpaceX could add over 10GW of computing capacity by the end of 2027, potentially driving annual recurring revenue (ARR) to an astonishing $300 billion.

This ambitious goal isn't unfounded. Elon Musk himself stated at SpaceX's first earnings call that the company's "conservative" target is to complete and deliver 6-8GW of new computing capacity in a single year by 2027, with optimistic scenarios potentially exceeding 10GW. Industry estimates suggest that each GW requires approximately $500 billion in capital expenditure, meaning SpaceX's capital expenditure for next year could range between $300 billion and $500 billion - a scale comparable to investments by cloud leaders like Amazon (AMZN-US) AWS and Google (GOOGL-US).

Analysts note that for a company with far lower profitability than mainstream cloud service providers, this level of investment is enough to make the market take notice. Semianalysis has evaluated all of SpaceX's suitable development sites and is continuously tracking the progress of its gas turbine equipment procurement, believing the aforementioned goals are highly feasible.

The core logic driving this expansion wave lies in the astonishing profit margins of AI inference services. According to Semianalysis' proprietary Tokenomics model and inference simulator, OpenAI and Anthropic can generate over $100 billion in annual revenue per GW capacity when providing API inference services using NVIDIA's (NVDA-US) GB300 chip clusters. With a conservative GPU rental rate of $3 per hour, the annual operating cost per GW is only about $12 billion, resulting in inference business gross margins typically exceeding 60%, with some core models reaching 85% or higher.

This profit model also applies to Microsoft. The report indicates that Microsoft renegotiated its cooperation terms with OpenAI in April 2026, eliminating the previous 20% revenue split mechanism. This means Microsoft can now enjoy similar per MW revenue and profit margins to OpenAI and Anthropic when providing inference services through OpenAI models, without having to bear any model training costs. Semianalysis describes this as a "once-in-a-century opportunity".

Microsoft itself experienced a "computing power shortage" in late 2024, significantly reducing its data center leasing pace. However, Semianalysis points out that this situation has completely reversed between 2025 and 2026. In October last year, Microsoft signed a $250 billion infrastructure-as-a-service agreement with OpenAI, which Semianalysis estimates involved about 7GW of computing power, causing Microsoft's in-house Foundry API business and applications like Copilot to temporarily face production bottlenecks.

To fill the gap, Microsoft has signed commitments for over 10GW of computing power this year through leasing, self-building, and long-term power purchase agreements, with a total value exceeding $300 billion. Against this backdrop, Semianalysis believes that Microsoft negotiating a 3GW, approximately $150 billion contract with SpaceX is "not impossible" for two reasons:

1. Microsoft is currently in a peak period of computing power expansion and urgently needs new capacity; 2. SpaceX, following its contract model with Anthropic and Google, offers a 90-day cancellation clause, significantly reducing financial risk and making internal approval processes easier to pass.

Semianalysis also points out that Microsoft's potential demand for SpaceX's computing power will become a key variable in driving Azure's revenue growth rate from approximately 42% to over 100%.

Semianalysis particularly highlights that SpaceX's true competitive advantage lies in completely overturning the traditional data center "efficiency-first" construction logic, adopting a "speed-first" strategy. The report cites multiple examples:

SpaceX's Colossus 1 300MW facility was completed in just 122 days; The Texas Southaven power plant expanded from 27 gas turbines (approximately 495MW) in February to 69 turbines with a total capacity of 1.7GW by July; The new "MiniHard" project, which began construction in March, is expected to reach 450-500MW scale in about five months.

Facing supply chain bottlenecks, SpaceX has taken multiple unconventional measures, including using Chinese-made power modules to replace large transformers that had been backlogged for over two years, purchasing used gas turbines from the secondary market (some of which were originally scheduled for delivery to Oracle's (ORCL-US) site in New Mexico), and implementing highly parallelized construction methods with pre-assembled compression test cycles. Compared to data centers of similar scale that typically require much larger workforces, Colossus 2's peak daily workforce was only about 3,000 people.

Semianalysis believes that in the current environment of extremely scarce computing power and persistently high AI inference profit margins, a 500MW cluster that can be delivered in three months with a 90-day cancellation clause is among the most scarce assets in the market. Google's ultimate decision to sign a contract with SpaceX is the most powerful testament to this logic.

Regarding external doubts about whether SpaceX has the financial strength to support such massive capital expenditures, Semianalysis proposes two possible paths. The first is that NVIDIA may help SpaceX reduce initial cash pressure through supplier financing, seen as one of the important reasons Musk announced at the earnings call that SpaceX would fully adopt NVIDIA chips in the future, although SpaceX had previously evaluated alternative options like Google TPU and AMD chips. The second is that SpaceX, leveraging the industry's fastest delivery speed, can recover capital expenditures within less than a year through operating cash flow by adopting a premium pricing strategy of $30-50 million per MW per year.

Considering all the above factors, Semianalysis estimates that even if only half of SpaceX's new computing power in 2027 is used for commercial inference monetization, with the rest allocated to its Grok and Cursor teams for model training, SpaceX's annual recurring revenue could still reach the $300 billion milestone by the end of 2027.