Can SpaceX Build 10GW of AI Data Centers by 2027?

SemiAnalysis concludes that SpaceX is on track to build roughly 10 gigawatts of AI compute capacity by the end of 2027. In a new report, the firm backs Elon Musk's claim on SpaceX's first earnings call that he "conservatively" plans to deliver an incremental 6-8GW next year, with potential to go well above 10GW. The case rests on a demonstrated construction pace no hyperscaler matches, inference economics that can pay back a gigawatt's capital cost in under a year, and a buyer, Microsoft, with an unusually strong incentive to sign for as many megawatts as it can get.

The dollar figures attached are enormous. At roughly $50B per gigawatt, the 2027 buildout implies $300-500B of capex in a single year, which SemiAnalysis notes is on par with what they expect from AWS and Google, from a company significantly less profitable than either. The report projects a path to $300B of annualized revenue for SpaceX by the end of 2027. The math underneath it is the clearest public articulation yet of why AI infrastructure spending keeps accelerating.

Why a gigawatt of inference capacity is worth $100B a year

SemiAnalysis estimates $100B per gigawatt per year. SemiAnalysis's Tokenomics Model and Inference Simulator estimate that both OpenAI and Anthropic can generate over $100B/GW/year of revenue selling API inference on a GB300 cluster, against roughly $12B/GW/year of cost, using a conservative rental rate of $3 per GPU-hour.

SemiAnalysis built an end-to-end simulator that models a frontier-class architecture executing on actual silicon, with timings for every operation and profiler-grade trace output, validated against a range of accelerators. The workload driving it is AgentX, their agentic coding benchmark built from real production coding traces, and the revenue figure blends input, cache-read, cache-write, and output token pricing at the ratios those real workloads produce. Agentic coding consumes tokens at rates interactive chat never approached, and the cache-heavy request shape (visible in any measurement of what a coding agent actually sends) changes the economics substantially compared to naive per-token math.

SemiAnalysis called inference gross margins north of 60% back in January, then published a June deep dive putting Opus 4.8 at 85%+ margins, a figure they note has since become the default number cited when analyzing Anthropic. A leaked DeepSeek investor call claiming a 10-month GPU payback period lands in the same ballpark. Serving frontier tokens at API prices is a high-margin business, and the market has taken a surprisingly long time to accept that.

$100B/GW/year says what a gigawatt can earn if demand absorbs every token at current API prices. Every downstream figure, including the $300B ARR path, inherits that assumption.

Microsoft is positioned as the largest buyer

Microsoft is, per the report, the third company in the world capable of printing those per-gigawatt economics, alongside OpenAI and Anthropic. It has full access to OpenAI's models, so it can earn the same revenue per megawatt while paying none of the training costs. The April 2026 rework of the OpenAI deal dropped the old 20% revenue share from the equation, which means every incremental megawatt Microsoft serves through its own products now accrues fully to Microsoft.

Much of Microsoft's current datacenter capacity goes to OpenAI at roughly $14M per megawatt per year under the $250B infrastructure agreement signed in October 2025, about 7GW of capacity by SemiAnalysis's estimate. That deal left Microsoft compute-constrained everywhere else: unable to fully serve its Foundry API business or applications like Copilot, which are its highest-margin, highest-revenue-per-megawatt services. Moving capacity toward those services means trading $14M/MW/year for a $100M/MW/year opportunity. SemiAnalysis models Azure revenue growth accelerating from around 42% to over 100% if Microsoft captures it.

After the leasing pause SemiAnalysis first flagged in December 2024, Microsoft has signed over 10GW of binding commitments year-to-date in 2026, roughly $300B in total contract value before GPU costs. Those contracts contribute capacity in late 2027 and 2028, leaving a near-term gap. SpaceX's 3-5 month lead times fill that gap.

A hypothetical 3GW Microsoft-SpaceX deal at $50B/GW/year is plausible. Microsoft's existing signing spree shows the appetite. A 90-day cancellation policy, similar to SpaceX's deals with Anthropic and Google, means the commitment carries little balance-sheet risk, which the report argues makes it an easy sign-off for CFO Amy Hood given the revenue on the other side.

How SpaceX finances a $300-500B buildout

SpaceX lacks a hyperscaler balance sheet, so the report expects the capex to be carried by two mechanisms.

The first is Nvidia vendor financing to lower the upfront cash cost. SemiAnalysis reads Musk's declaration on the earnings call that SpaceX would be Nvidia-exclusive as evidence: xAI and SpaceX had actively evaluated TPU and AMD alternatives, and the financing terms likely settled the question.

The second is operating cash flow, enabled by pricing power that comes directly from speed. Large-scale, near-term compute is scarce, and SpaceX can sell it with 3-5 month lead times. That lets them price at $30-50M per megawatt per year, which pays back the roughly $50M-per-megawatt capex in less than a year. A business that recovers its capital that fast can fund a large fraction of its own expansion at this scale.

Building 10GW in one year has little precedent

The report cites Colossus 1 reaching 300MW in 122 days and Colossus 2 reaching 200MW in six months, construction timelines conventional datacenter developers measure in years. On the power side, a Southaven, Mississippi generation facility reportedly expanded from 495MW in February 2026 to 1.7GW by July 2026. SemiAnalysis says it has mapped the candidate sites and tracked available gas turbine, engine, and fuel cell supply quarter-by-quarter across more than 30 suppliers, and finds the equipment exists.

SpaceX's compute capacity will be about 2GW at the end of 2026. Hitting 10GW by the end of 2027 means quintupling in twelve months, with the aggregate supply chain for generation equipment, grid interconnects, and GPUs becoming the constraint. Building 300MW in 122 days demonstrates the capability, but doing it continuously, in parallel, across many sites for a year has little precedent.

The $300B ARR projection assumes half the new capacity is sold externally

The revenue projection assumes only 50% of the 2027 incremental compute is monetized externally, with the remainder going to the Grok and Cursor teams for training and modeled at zero inference revenue. The monetized half is assumed to sell at the scarcity-premium rates of $30-50M/MW/year, and build completion, offtaker commitments, and token sales at API prices have to hold at once.

Secondary coverage has already garbled the figure: the piece's own page shows both "$500B ARR" and "$300B ARR" in different renderings of the title, and some aggregators ran with the larger number. The $500B figure belongs to the top end of the 2027 capex range; the ARR projection is $300B.

The 90-day cancellation terms make these deals signable because neither Microsoft nor the AI labs take balance-sheet risk, but they also mean the revenue behind the ARR projection can be handed back on 90 days' notice. Annualized run-rate built on cancellable short-notice contracts will hold as long as inference margins stay where SemiAnalysis measures them and token demand at API prices keeps growing fast enough for offtakers to renew at $30-50M/MW/year.