SpaceX's $14.6B AI-fueled quarter meets Apollo's warning: compute demand isn't unlimited

The quarter that powered itself on AI

SpaceX's third quarter as a public company reads like an AI-infrastructure earnings report that happens to own rockets. The company reported Q3 2026 revenue of $14.6 billion, up 161% year over year, with net income of $4.4 billion against a $1.1 billion loss in the prior-year period. Super Intelligence segment revenue reached $7.6 billion, up from $813 million a year earlier, on newly activated compute capacity and the Cursor acquisition. Bloomberg's Surveillance podcast noted net income of $4.4 billion beat the $1.8 billion estimate, and revenue of $14.58 billion topped the $12.87 billion consensus -- with the AI and connectivity businesses delivering higher profitability than expected. m:…8862 m:…3907

The physical numbers tell the same story: SpaceX ended the quarter with 1.7 gigawatts of nameplate compute capacity, says it has line of sight to 2.5-3 gigawatts by the end of Q4, and expects fourth-quarter capital expenditures of $35-40 billion. Adjusted EBITDA hit $8.8 billion, up 306% year over year. Management now guides Q1 2027 revenue of $33-37 billion, citing visibility into compute contract pipelines. Elon Musk confirmed special committees are discussing a potential SpaceX-Tesla merger, though he said they have not reached agreement on terms. m:…8862

Beyond the AI segment, the connectivity flywheel keeps turning. Starlink reached 13.6 million consumer subscribers, adding 1.6 million in the quarter, and Bloomberg's discussion noted revenue growth outpacing cost increases as the constellation tops 11,000 satellites. SpaceX also announced a Louisiana spaceport with 20 or more planned launch towers, targeting weekly Starship launches by the first half of 2027, and holds a volume-production aspiration of Q3 2028 for chips at its Terafab facility in Gibbons Creek. On the Bloomberg podcast, Musk floated distributing SpaceX shares to US children via the Trump accounts program -- which the hosts noted set up roughly 70 million auto-accounts -- alongside the Tesla-merger discussions. m:…8862 m:…3907

Apollo's counter: the demand curve bends

Apollo's Torsten Slok is reading a very different demand curve. In an Oct. 20 analysis, Apollo argues the recent slowdown in token-spending growth shows Silicon Valley's premise -- that compute demand is unlimited -- is breaking down. Compute, Slok says, faces a flat, highly price-sensitive demand curve: even modest changes in token prices push buyers to throttle usage or substitute cheap open-source alternatives, and the marginal cost of producing compute now exceeds the marginal benefit it delivers to users. m:…8154

The pressure shows up in the balance sheet, not just the benchmark. Apollo notes that invested capital at the five largest hyperscalers -- Amazon, Google, Meta, Microsoft, and Oracle -- is growing faster than the earnings it generates, with return on invested capital declining. AI investments, pitched as growth engines, are turning into margin-draining cost centers because productivity returns are arriving slower than projected. In Apollo's framing, the number-one variable for data-center construction viability is not power or capacity but the price users are willing to pay for compute. m:…8154

Two stories, one bill

The tension between the two narratives is the executive question for the next two quarters. SpaceX's results show what happens when contracted AI demand meets deployed capacity: 1.7 gigawatts of nameplate compute converted into $7.6 billion of Super Intelligence revenue and tripled EBITDA. Apollo's warning describes what happens one step upstream, where uncontracted demand meets rising prices: buyers throttle, substitute, and hesitate. Both can be true at once -- contracted pipeline growing while marginal willingness to pay softens. m:…8862 m:…8154

SpaceX's forward commitments raise the stakes of Apollo's question. A $35-40 billion capex quarter and $33-37 billion Q1 2027 revenue guide require not just capacity but customers willing to pay for tokens at prices that justify the buildout. The company is betting the pipeline stays full; Apollo is betting buyers push back. Executives signing compute contracts into 2027 should watch which side the hyperscaler ROIC numbers land on -- and price their own AI bets accordingly. m:…8862 m:…8154

One more signal from the Bloomberg discussion underscores how long-dated this debate is: SpaceX's prospectus-level ambitions -- a $26.5 trillion total addressable market -- depend on orbital data centers, which in turn depend on Starship becoming rapidly reusable and economically viable. That is Apollo's question pushed to its limit: whether demand for compute ever justifies its production cost, on Earth or off it. m:…3907 m:…8154