SpaceX Has Three AI Customers Paying $27.8 Billion a Year. How Big Can This Revenue Stream Get?

SpaceX's new AI compute business has signed three deals totaling about $2.32 billion in monthly revenue, or $27.8 billion annualized, more than doubling the company's 2025 revenue base of $18.7 billion. The contracts with Anthropic, Google, and Reflection AI monetize underutilized data center capacity and could become a high-margin growth engine, with gross margins potentially near or above 70%. The article is constructive on SpaceX's long-term earnings power, though it notes the stock still screens expensive at roughly 40x sales.

Analysis

The real read-through is not “SpaceX got a new revenue stream,” but that frontier AI infrastructure is beginning to separate into three layers: chip ownership, power/site control, and customer aggregation. If a high-profile platform can lease out stranded capacity at scale, the winner set broadens from GPU manufacturers to the owners of scarce, already-permitted power and cooling footprints. That is mildly negative for pure self-build hyperscale economics and supportive for any asset-light compute broker with balance-sheet flexibility, while keeping demand structurally positive for NVDA because every new leased contract still pulls through chips.

Second-order, this is a margin story before it is a revenue story. The market is likely underestimating how much of the capex cycle can be monetized once utilization improves; incremental gross margin on unused capacity should be far superior to greenfield builds, so the first inflection is likely to show up in FCF and valuation multiples rather than reported earnings. That creates a potential rerating path for GPU cloud names if investors start treating them less like cyclical infra and more like contracted utility-like cash flows.

The main risk is not demand evaporation but supply response: hyperscalers can compress pricing once they decide renting is temporary, and any acceleration in sovereign or private buildout could cap medium-term lease rates. Time horizon matters—near term, this supports AI capex sentiment for months; over 12-24 months, it could become a competitive pressure point if capacity becomes commoditized. The consensus seems to be focused on revenue magnitude, but the more important variable is whether this proves that idle AI infrastructure can be turned into recurring, high-ROIC annuity value; if yes, the whole AI infrastructure stack deserves a higher multiple, but if no, this is just a one-off monetization event.

Key Decisions for Investors

• Long NVDA on 3-6 month horizon: the business model shift increases terminal GPU demand without requiring a change in end-customer economics; use any post-rally consolidation as entry, with upside tied to a longer-than-feared AI capex cycle.
• Long GOOGL vs short MSFT, 1-3 months: GOOGL has more direct optionality from outsourced compute and appears better positioned to arbitrage build-versus-rent economics; MSFT faces more self-build capex pressure and less obvious incremental benefit.
• Long CRWV vs short a basket of hyperscaler capex proxies, 3-6 months: if the market starts rewarding contracted compute capacity, the multiple expansion on leased-infrastructure cash flows can outrun consensus estimates; risk is pricing compression if supply floods in.
• Sell downside puts on NVDA or buy call spreads 6-12 months out: the best risk/reward is exposure to persistent pull-through demand with defined downside, rather than outright chase after a sentiment-driven move.
• Avoid shorting the AI infrastructure complex here: the catalyst is utilization monetization, not demand destruction, and the more likely near-term error is underappreciating how quickly stranded capacity can be repriced into cash flow.