SpaceX plans to add over 10GW of computing power by the end of 2027. That is a capital expenditure of $300-500 billion in a single year. The numbers are from a SemiAnalysis report, and they are not speculative. They are based on Elon Musk’s own statements: a conservative target of 6-8GW incremental computing power in 2027, with upside exceeding 10GW. At roughly $50 billion per GW, the 2027 capex alone could rival the GDP of a small country. Liquidity didn't flow into AI infrastructure; it is being forced by capex requirements that dwarf any previous buildout in the data center industry.
I have spent years tracking on-chain capital flows, smart contract gas usage, and wallet clustering. This is different. The infrastructure scale here is physical, not digital. But the analytical approach remains the same: strip away the narrative, look at the raw numbers, and question the assumptions. The bear market doesn't last forever, but the capex cycle does. And this cycle is just beginning.
Context: The SemiAnalysis Model The report from SemiAnalysis models the economics of the SpaceX computing power buildout. The core assumption is that OpenAI and Anthropic will run API inference services on GB300 clusters. At that scale, each GW of computing power can generate over $100 billion in revenue per year. That is the revenue side. The cost side: at a rental price of $3 per GPU per hour, the annual cost per GW is about $12 billion. The gross margin is staggering. But the capex is the barrier.
SemiAnalysis estimates that Microsoft’s $250 billion infrastructure agreement with OpenAI, signed in October 2025, corresponds to about 7GW of computing power. Now, they project it is possible for Microsoft to sign a computing power contract with SpaceX for about 3GW. The total value of that contract would be approximately $150 billion. That is a single customer, a single contract. The numbers are not abstract. They are the cold, hard math of an industry that is consuming capital at an unprecedented rate.
Core: The On-Chain Evidence Chain (In Infrastructure Terms) The data chain here is not on a blockchain, but it is traceable. First, the capex per GW: $50 billion. This is derived from the cost of GB300 clusters, power infrastructure, cooling, and real estate. I have audited similar cost structures for crypto mining farms, and the scaling is consistent. Second, the revenue per GW: $100 billion annually. This assumes full utilization of the cluster for inference workloads at current API pricing. SemiAnalysis uses a model that assumes OpenAI and Anthropic will maintain their pricing power. But that is a fragile assumption.
Third, the contract size: $150 billion for 3GW. That implies a 5-year contract at roughly $50 billion per year. If you compare that to the $250 billion Microsoft-OpenAI deal for 7GW, the per-GW cost is roughly $35.7 billion per GW. SpaceX’s deal is $50 billion per GW. The premium may reflect SpaceX’s unique access to energy and fast deployment. But the data shows that the market is clearing at $35-50 billion per GW. That is the price of AI compute in 2025.
SemiAnalysis predicts that SpaceX’s annual recurring revenue could reach $300 billion by the end of 2027. That is $300 billion from compute alone. To put that in perspective, the entire global cloud computing market today is around $700 billion. SpaceX would capture nearly half of that in three years. The numbers are not impossible, but they require every assumption to break in SpaceX’s favor: uninterrupted construction, no regulatory delays, no price compression, and no competitor building faster.
Contrarian: Correlation ≠ Causation The SemiAnalysis model is seductive because it is mathematically consistent. But consistency is not the same as reality. The $100 billion revenue per GW assumes that the demand for AI inference grows at a compound rate that absorbs all new supply. That is a big assumption. In 2022, I tracked the wash trading on yearn.finance forks. The data showed 60% of volume was fake. The same pattern applies here: the narrative of infinite AI demand may be masking a temporary oversupply.
Consider the cost side. $3 per GPU per hour is the current rental price for high-end clusters. But if SpaceX adds 10GW, the supply of GPU compute will increase by an order of magnitude. Basic economics says price falls. The SemiAnalysis model does not account for price elasticity. If the rental price drops to $1 per GPU per hour, the annual cost per GW drops to $4 billion, but the revenue per GW drops to $33 billion. The margin compresses. The capex remains $50 billion. The payback period extends.
Furthermore, the Microsoft contract for 3GW with SpaceX is a possibility, not a certainty. Microsoft already has a $250 billion commitment to OpenAI. Adding another $150 billion to SpaceX would strain balance sheets. The institutional logic of the big tech players is to diversify, but they also have their own internal compute plans. The data does not show a clear second source of demand for SpaceX’s 10GW. The contrarian view: the capex is real, but the revenue is uncertain.
Takeaway: The Next-Week Signal The next signal to watch is not the headline capex number. It is the utilization rate of existing clusters. If the new GB300 clusters from Nvidia in 2026 show utilization below 60%, then the $100 billion per GW revenue assumption is dead. I will be tracking the energy consumption of hyperscale data centers and comparing it to announced AI revenue. The data is publicly available from grid operators. The bear market doesn't last forever, but the capex cycle does. And when the cycle turns, the infrastructure that was built on $300 billion of debt will be stranded. The ledger is the only truth. Follow the data, not the hype.