SpaceX claims it will add 10GW of computing power by end of 2027. The numbers are staggering. The logic is not. Let's trace the ghost in the compute contract state.
SemiAnalysis released a report asserting that SpaceX's goal of 10GW is feasible. Musk's conservative target is 6-8GW, with upside to 10GW. At $50 billion per GW, that's $300-500 billion in capital expenditure in 2027 alone. The revenue model: each GW generates over $100 billion annually from API inference on GB300 clusters, assuming $3 per GPU hour. The cost per GW is $12 billion per year. The gross margin would be 88%. Beautiful on paper. But paper is not a ledger.
Context: The Infrastructure Mirage
SpaceX is primarily a rocket company. Starlink generates revenue, but not enough to fund $500 billion in capex. SemiAnalysis estimates Microsoft's $250 billion deal with OpenAI corresponds to 7GW, and a possible SpaceX deal with Microsoft for 3GW worth $150 billion. But deals are not cash. They are commitments, often contingent on milestones. The crypto space taught me that a signed term sheet is not a confirmed transaction. I've audited protocols where $100 million TVL was claimed but on-chain liquidity was $10 million. The same principle applies here: announced capacity is not deployed capacity.
Core: The Capex Flood and the Debt Trap
Let's dissect the numbers. $50 billion per GW is the capex. For 10GW, that's $500 billion. SpaceX's current annual revenue is roughly $10 billion from Starlink and launch services. To fund this, they would need debt or equity markets to open a $500 billion pipeline. But the market is in a bear sentiment. Interest rates are high. The FTX collapse showed that even large crypto funds can evaporate when leverage is used to fund infrastructure. The same logic applies to compute: if the demand for AI inference slows, the capex becomes stranded. The revenue model assumes $100 billion per GW at full utilization. But utilization is a function of demand. If OpenAI and Anthropic are the primary tenants, what happens if their models commoditize? The margin erodes.
During the 2017 ICO boom, I identified a critical flaw in Parity Wallet's multi-signature implementation. The flaw was not in the code alone, but in the assumption that keys would be managed securely. Here, the flaw is the assumption that AI compute demand will grow linearly. The data shows that training compute costs have dropped, but inference costs are sticky. The SemiAnalysis model assumes $3 per GPU hour. That's a price point that can be undercut by competitors like Google's TPU or custom chips. The capex per GW does not include the full cost of energy, cooling, and network infrastructure. At 10GW, the power draw alone is equivalent to 10 nuclear reactors. The grid is not ready.
From my forensic analysis of the Lendf.me exploit, I learned to follow the transaction flow. Here, the flow is: capex → hardware → deployment → utilization → revenue. The bottleneck is not hardware supply; it's deployment. SpaceX's advantage is in launching hardware into orbit or remote locations? No, the report is about terrestrial data centers. The narrative is that SpaceX will build massive data centers powered by their own rockets? That doesn't add up. Rockets are for launching satellites, not for building data centers. The synergy is unclear. The ghost in the state is the lack of a clear path from the launchpad to the server rack.
Contrarian: What the Bulls Get Right
To be fair, the bulls might argue that SpaceX's vertical integration with Starlink provides a low-latency network backbone, and their manufacturing capabilities could reduce costs. The SemiAnalysis model also assumes that the $100 billion revenue per GW is from API inference, not from training. Inference is more predictable and less elastic. If the GB300 clusters are custom-designed for inference, the efficiency could be higher than general-purpose hardware. Additionally, the $250 billion Microsoft deal validates the narrative that hyperscalers are willing to pay a premium for guaranteed compute. The 3GW deal with SpaceX could be a hedge against supply chain constraints in Taiwan. The demand for AI compute is real, but the question is whether the infrastructure can be built faster than the hype cycle.
The Silent Log
Silence in the balance sheet is louder than the error. Without audited financials, the capex numbers are speculative. The $300-500 billion capex in 2027 is roughly 3x the entire global semiconductor capital expenditure in 2025. That concentration of risk is a systemic vulnerability. I've seen similar patterns in crypto where a protocol claims to have billions in liquidity but the underlying assets are illiquid. Here, the asset is compute capacity, which is perishable if not utilized. The annual cost of $12 billion per GW is a fixed cost. If utilization drops to 50%, the margin collapses.
Computing power is a warm lie if the capex leaks. The true test is not the announcement but the shovels in the ground. Based on my audit experience of infrastructure projects in crypto, the risk of over-leverage is high. The FTX forensics taught me that when you trace the flow of funds, you often find that the largest numbers are the most opaque. The 10GW claim is a forward-looking statement. The market should treat it as a hypothesis, not a fact.
Takeaway: The Exponential Curve Trap
The billion-dollar compute contracts are a bet on exponential demand. But exponential curves tend to break. Investors should audit the assumptions, not the press releases. The next step is to track actual deployment milestones: When will the first GW be online? What is the utilization rate? Who is the anchor tenant? Until then, the numbers are just variables in a model. And variables can be manipulated.