Hook
SpaceX plans to add over 10GW of computing power by the end of 2027. That’s ten nuclear reactors—each dedicated to running AI inference. Elon Musk calls it a “conservative” target. Semianalysis’s model pegs annual revenue per GW at over $100 billion. The numbers are staggering. But they don’t survive a forensic audit.
Context
In late 2025, a SemiAnalysis report surfaced claiming SpaceX’s data center buildout is not only feasible but already underway. Musk stated that the company’s conservative target is 6-8GW of incremental compute in 2027, with upside exceeding 10GW. Capital expenditure: roughly $50 billion per GW. That means 2027 capex alone could hit $300–500 billion—more than the entire global semiconductor industry spends annually. The report further asserts that when OpenAI and Anthropic run API inference on GB300 clusters, each GW generates over $100 billion in revenue per year. At a rental price of $3 per GPU per hour, the annual cost per GW is about $12 billion. The implied margin: 88%.
Then comes the Microsoft connection. In October 2025, Microsoft signed a $250 billion infrastructure agreement with OpenAI, corresponding to about 7GW of compute. SemiAnalysis suggests Microsoft could sign a similar deal with SpaceX for roughly 3GW, valued at ~$150 billion. By end of 2027, SpaceX’s annual recurring revenue could reach $300 billion.
Core
Let’s crack open the model. The assumptions are seductive—but they collapse under scrutiny.
Revenue per GW: $100B/year?
That figure assumes 100% utilization at peak pricing. Real-world inference clusters operate at 60-70% utilization. Discounting by 30% yields $70B per GW. Still high, but the model also assumes no price compression. As more compute comes online, GPU rental rates fall. The $3/GPU/hour price is already below spot for H100s. By 2027, with 10GW+ flooding the market, that rate could drop to $1.50 or less. Revenue per GW then plummets to $35B.
Cost per GW: $12B/year?
That’s the rental cost. But building the infrastructure is the real capital sink. $50B per GW is a rough estimate. Let’s dig deeper. A 1GW data center requires 1 million square feet of space, 100,000 high-end GPUs, cooling systems, networking fabric, and 24/7 power. Current hyperscaler build costs hover around $30-40B per GW including the chips. SpaceX’s vertical integration (Starlink for networking, Starship for launching hardware?) could lower that, but the timeline is aggressive. The report assumes SpaceX can achieve economies of scale that no one else has. Based on my experience auditing infrastructure projections for crypto mining farms, I’ve seen similar optimism—always followed by delays and cost overruns.
The Microsoft deal: $250B for 7GW?
That’s $35.7B per GW. If SpaceX can build for $50B per GW, they’re losing money on the deal. Unless they plan to sell at a loss to capture market share. But Microsoft is a savvy buyer. They’ll negotiate hard. The $150B for 3GW from SpaceX implies $50B per GW, matching the capex. That’s zero margin. Why would SpaceX do that? Unless they expect to upsell inference services later. But the risk is asymmetric.
GPU supply chain constraints
Nvidia’s GB300 clusters are not off-the-shelf. They require custom interconnects, memory, and cooling. The lead time for a single 1GW cluster is 18-24 months. SpaceX would need to order 10GW worth of chips today to hit 2027. That’s 10 million GPUs. Nvidia’s total projected output for 2027 is around 15 million. SpaceX would consume 67% of the world’s supply. Not impossible, but it would require pre-ordering years in advance with no guarantee of demand. The contract with Microsoft might cover 3GW, but the rest is speculative.
Energy grid integration
10GW is 10% of the total electricity generation of Texas. Building that much capacity in a few years requires grid upgrades, permits, and substations. The average US data center takes 5 years from planning to operation. SpaceX would need to build 10GW in 2 years. The timeline is a fantasy.
Contrarian
What the bulls get right: Musk’s track record is real. SpaceX has revolutionized rocketry, Starlink is a global constellation, and Tesla’s gigafactories scaled fast. The vertical integration play could work. Starlink’s satellite network could provide edge compute in remote locations, reducing latency for AI inference. And the $100B revenue per GW might be conservative if we factor in near-term monopolistic pricing for high-end AI models. Additionally, the Microsoft deal provides an anchor customer, reducing demand risk.
But the blind spots are deeper. The report assumes perfect execution. It ignores the human capital bottleneck—there aren’t enough data center engineers to build 10GW in two years. It ignores the cooling water crisis—each GW consumes 1 million gallons per day. It ignores the political backlash—tech companies are already facing energy shortage protests. Most importantly, it ignores the crypto angle: if SpaceX hoards GPUs, it squeezes supply for blockchain mining and AI crypto projects. The decentralization narrative collapses when a single entity controls the majority of compute.
Takeaway
SpaceX’s 10GW plan is a high-risk bet that could reshape the AI infrastructure landscape—or explode into the largest capex bubble since the 2000 dot-com crash. For due diligence analysts, the signal is not the revenue numbers; it’s the silence around energy procurement, GPU supply agreements, and construction timelines. Metadata whispers what the contract screams. The silence in the logs is louder than any statement. The image is static; the provenance is a phantom.
If you’re a crypto investor, watch the GPU spot market. If prices don’t spike by 2027, the plan is falling apart. And if they do, the real winner isn’t SpaceX—it’s the energy producer and the chip manufacturer. My take: Musk is selling a vision to extract funding from Microsoft and OpenAI. The actual compute delivered will be closer to 3GW, not 10. The rest is noise.