SpaceX's 10GW Power Capacity Plan Could Generate $300B Annual Revenue, With Microsoft as Largest Customer

**The $300 Billion ARR Thesis: Why SpaceX's 10GW Datacenter Ambition Is Credible**

When Elon Musk announced on SpaceX's first-ever earnings call that he planned to build and deliver an incremental 6 to 10 gigawatts of AI compute capacity in 2027 alone, the reaction was predictably skeptical. At roughly $50 billion per gigawatt, that implies $300 to $500 billion in capital expenditure from a single company in a single year — a figure comparable to what Amazon Web Services and Google are expected to deploy, but coming from a company that lacks the balance sheets of entrenched hyperscalers. Yet a detailed review of SpaceX's construction pace, power infrastructure moves, and emerging customer commitments suggests the number is not fantasy.

The underlying economics make the ambition logical. Running a frontier AI model — think the architecture class of current GPT or Claude flagship models — on a one-gigawatt cluster of GB300 GPUs can generate over $100 billion per gigawatt per year in revenue for an inference provider selling API access. Even assuming conservative GPU rental costs of $3 per GPU-hour, the annual cost per gigawatt runs around $12 billion. The spread is extraordinary: inference gross margins for frontier models now exceed 60%, with some specific model-accelerator combinations running above 85%. A leaked DeepSeek investor call reportedly indicated a ten-month GPU payback period, consistent with this margin structure. For AI labs and the hyperscalers serving their models, securing large-scale compute as fast as possible is not merely a growth strategy — it is the highest-returning capital allocation available anywhere in the economy.

SpaceX is positioning itself as the supplier that can deliver compute faster than anyone else, and is pricing accordingly at $30 to $50 million per megawatt per year — a premium that reflects genuine scarcity. At that pricing, the capital outlay on a new cluster pays back in under twelve months.

**Microsoft's Awakening and the OpenAI Deal Restructuring**

The demand anchor for SpaceX's ramp is Microsoft, which has undergone a dramatic reversal after a well-documented pause in datacenter leasing activity that began in late 2024. Year-to-date in 2026, Microsoft has signed binding contracts totaling over 10 gigawatts across self-build construction starts, third-party leases, neocloud agreements, and large-scale power purchase agreements — representing more than $300 billion in total commitment value, excluding GPU costs.

The strategic logic behind this acceleration is rooted in a renegotiated relationship with OpenAI. A deal restructured in April 2026 eliminated the old revenue-sharing arrangement in which Microsoft surrendered roughly 20% of revenue back to OpenAI. Under the new terms, Microsoft retains the full economics of serving OpenAI models through its Azure platform and the Foundry API business. This changes the calculus entirely: every additional megawatt Microsoft can bring online for inference workloads now translates directly into high-margin Azure revenue rather than being partially handed back to its model provider.

The revenue opportunity is substantial. Microsoft signed a $250 billion infrastructure agreement with OpenAI in October 2025, estimated at roughly seven gigawatts of total capacity, but that commitment has consumed so much of Microsoft's available compute that it has been unable to fully exploit its access to OpenAI models for external API sales and Copilot applications — precisely the use cases that generate the highest revenue and margin per megawatt. Closing that gap is the imperative driving Microsoft's 10-gigawatt contracting surge. A reported 3-gigawatt agreement with SpaceX at $50 billion per gigawatt — totaling $150 billion — would be one mechanism for doing so. Crucially, SpaceX's standard 90-day cancellation clause eliminates balance sheet risk for Microsoft CFO Amy Hood, making the commitment structurally easy to authorize even at eye-watering headline numbers. The result could be Azure revenue growth accelerating from its current roughly 42% annual rate to above 100% in the next year.

**How SpaceX Circumvents the Constraints That Stall Everyone Else**

SpaceX's ability to execute a 10-gigawatt build in a single calendar year hinges on a construction philosophy that deliberately trades conventional efficiency metrics for raw speed — a trade that makes sense only in an environment where compute scarcity commands extraordinary premiums.

The evidence of this approach is concrete. The Colossus 1 build in Memphis delivered 300 megawatts in 122 days. Colossus 2 added 200 megawatts in six months. The onsite gas generation facility at Southaven, Mississippi expanded from 27 turbines — roughly 495 megawatts — in February 2026 to 69 turbines representing 1.7 gigawatts by July 2026, a near-quadrupling of generation capacity in five months. A new facility nicknamed MiniHard, upon reaching full vertical construction in March 2026, appears on track to hit 450 to 500 megawatts within approximately five months of that milestone.

SpaceX resolves the industry's two most acute bottlenecks through unconventional sourcing. Large power transformers are effectively sold out for two years and GE Vernova-class gas turbines face backlogs exceeding five years. SpaceX bypasses transformer constraints by importing power modules from China and delivering medium-voltage power directly to low-voltage transformers, which remain widely available — accepting some efficiency loss in exchange for speed. On the turbine side, the company draws on a broad set of over thirty manufacturers of gas generation equipment rather than concentrating on the marquee suppliers that face the longest queues. Secondary market turbines are another lever: equipment originally slated for Oracle's New Mexico data center project has reportedly become available, and SpaceX has the pricing power and willingness to pay secondary market premiums. Labor is managed through aggressive parallelization, reduced commissioning timelines, and prefabricated assembly — reportedly achieving peak daily worker counts at Colossus 2 of around 3,000, meaningfully below the staffing levels typical of other gigawatt-scale datacenter projects currently under construction.

The financing model to sustain this pace relies on two pillars: vendor financing from Nvidia — which likely explains Musk's declaration of Nvidia exclusivity on the earnings call, a notable reversal for a company that had actively evaluated AMD and Google TPU alternatives — and operating cash flow generated by the premium pricing that near-term delivery commands. At $30 to $50 million per megawatt per year, even partial occupancy of new clusters generates enough cash to fund subsequent construction without requiring the capital market access that traditional hyperscalers rely on.

If 50% of SpaceX's 2027 incremental capacity is monetized through external inference contracts — with the remainder reserved for Grok and Cursor training workloads — the resulting annualized revenue run rate reaches $300 billion by end of 2027. The number remains staggering, but the construction track record, the customer demand structure, and the financing mechanics make it a serious projection rather than a promotional one.