SpaceX's 10GW Datacenter Ambition: How Elon Musk Plans to Build a $300B ARR Cloud Empire by 2027

Elon Musk unveiled during SpaceX's first-ever earnings call that he plans to build and deliver an incremental 6–10+ gigawatts of datacenter capacity in 2027 alone — a staggering claim that would put SpaceX's capital expenditure on par with AWS and Google at roughly $50B per GW. Despite SpaceX being far less profitable than those hyperscaler rivals, analysts tracking the buildout believe the number is credible, pointing to a pattern of construction speed that has consistently defied industry norms.

The financial logic underpinning the ambition is a crucial piece of the story. At current neocloud rental rates around $3/GPU-hour, a single gigawatt of GB300 cluster capacity costs roughly $12B per year to operate — but frontier AI model companies like OpenAI and Anthropic can generate over $100B per GW per year in inference revenue when serving API tokens. That's an extraordinary margin spread, and it's the economic engine driving every major player to scramble for compute as fast as possible. Inference gross margins on flagship models like Anthropic's Opus series have been estimated north of 85%, a figure now widely cited in analyst circles after being first surfaced via careful synthesis of leaked financials and microbenchmark data.

Microsoft emerges as the most consequential potential customer in this picture. After a dramatic and widely-noted pause in leasing activity in late 2024, Microsoft has since signed over 10GW in binding datacenter contracts year-to-date in 2026, representing more than $300B in total commitments. The catalyst: a renegotiated deal with OpenAI in April 2026 that stripped out the old 20% revenue share arrangement, giving Microsoft near-full economics on every token served through its Azure Foundry and Copilot products. With that restructured deal in place, Microsoft's incentive to accumulate megawatts is enormous — analysts estimate Azure revenue growth could accelerate from roughly 42% to over 100% if Microsoft can fill capacity with high-margin inference workloads. A 3GW deal with SpaceX at $50B/GW/year is floated as plausible, particularly given a 90-day cancellation clause that eliminates balance sheet risk.

SpaceX's ability to actually deliver at this scale hinges on its unconventional construction playbook. Where traditional datacenter operators spend years navigating large power transformer backlogs and switchgear shortages, SpaceX imports power modules from China and routes medium-voltage power directly to widely available low-voltage transformers, bypassing the bottleneck entirely. At the Southaven power plant, SpaceX expanded from 27 gas turbines (~495MW) in February 2026 to 69 turbines (~1.7GW) by July 2026 — a roughly 3.5x capacity increase in five months. A new facility dubbed 'MiniHard' is on track to reach 450–500MW within approximately five months of breaking ground. Peak daily construction labor at Colossus 2 reportedly ran to around 3,000 workers, which analysts note is significantly below what comparably sized projects typically require, reflecting SpaceX's emphasis on parallelization and prefabrication over headcount.

Financing a buildout of this magnitude without a hyperscaler balance sheet is the remaining open question. The most likely answer is a two-part structure: vendor financing from Nvidia — which analysts believe is the real reason Musk declared SpaceX Nvidia-exclusive on the earnings call, abandoning earlier evaluations of AMD and Google TPUs — combined with operating cash flow from SpaceX's premium pricing strategy. By charging $30–50M per MW per year with 3–5 month delivery timelines, SpaceX can theoretically recover its capital expenditure in under a year, a payback period no traditional datacenter operator can match. If 50% of SpaceX's 2027 incremental compute is monetized for inference (with the remainder reserved for Grok and Cursor training workloads), the math points toward $300B in annualized recurring revenue by end of 2027.