SpaceX doubles revenue on Anthropic and Google compute deals, Starlink growth

SpaceX just printed the clearest signal yet that the AI gold rush has reversed polarity. The picks-and-shovels play — renting bare metal and power to the labs that actually train frontier models — now generates more revenue growth than the rocket business that made the company famous. Nearly $2 billion of the $3.8 billion year-over-year jump came from the AI division. Another $1.7 billion came from Starlink. The launch manifest mattered less than the lease agreements.

This was not the plan. Musk folded xAI into SpaceX to build a vertically integrated titan: proprietary models, proprietary silicon, proprietary distribution. Instead, xAI face-planted. Grok hallucinated a "MechaHitler" persona. The system generated child sexual abuse material. Customers did not materialize. The Memphis data centers, built for xAI's training runs, sat fallow until SpaceX flipped them into a landlord business for Anthropic and Google. The CFO called the resulting margins "high incremental EBITDA." Translation: the hosting deals print cash because the fixed cost is already sunk.

The $100 billion annualized revenue run-rate claim deserves scrutiny. Bret Johnsen tied it to "fully integrating" Cursor, a coding assistant startup. Elon Musk went further: the $100 billion figure is what happens "if we basically did nothing." That phrasing should alarm shareholders. A CEO who treats a twelve-figure run-rate as a do-nothing baseline is either sandbagging or deluded. The 2025 full-year revenue was $18.67 billion. Quintupling that in six months while the stock sinks below the IPO price suggests the market has already modeled the skepticism.

The capital allocation tells its own story. $28 billion in capex through the first half of 2026 versus $7 billion in the same period last year. That is not a company "doing nothing." That is a company digging moats with borrowed shovels. The $85 billion IPO and subsequent bond sale handed Musk a $100 billion war chest. He is deploying it at a clip that would make a sovereign wealth fund blush. The quarterly loss narrowed from $1 billion to $541 million, but the burn rate remains violent. Free cash flow is a mirage.

Starlink remains the steadiest engine. $1.7 billion in incremental revenue implies the constellation is finally converting capacity into recurring subscriptions at scale. That business has physics on its side: spectrum, orbit, and a regulatory moat that terrestrial ISPs cannot replicate. But Starlink's growth rate is linear. The AI hosting revenue is step-function — lumpy, contract-dependent, and exposed to the whims of two customers who are also competitors. Anthropic and Google both build their own infrastructure. They are renting SpaceX's because they cannot bring their own online fast enough. That is a temporary condition.

The market's verdict is unambiguous. Shares priced at $135 at IPO closed at $125 and shed another 8% after hours. The largest public offering in history has become a broken print. Investors are not buying the $100 billion ARR narrative. They are pricing the execution risk of a company that pivoted from building models to landlording GPUs, that burns half a billion a quarter, and whose CEO treats a twelve-figure run-rate as a conservative estimate.

Musk's "did nothing" comment reveals the strategic vacuum. If the baseline requires zero incremental effort, the upside case requires what — miracles? The Cursor integration, the Anthropic renewals, the Google expansion, the Starlink direct-to-cell rollout, the Starship launch cadence — each is a dependency chain with single points of failure. The war chest buys time. It does not buy product-market fit for xAI's models. It does not erase the reputational scar tissue from Grok's disasters. It does not guarantee that Anthropic and Google renew at scale rather than migrate to their own silicon.

SpaceX has become a real estate investment trust with a rocket subsidiary. The economics are legible: high fixed cost, near-zero marginal cost, long-term contracts, predictable yield. That is a defensible business. It is not a $1.75 trillion business. The valuation always priced the vertical integration dream — rockets launching satellites feeding data into models running on proprietary chips sold to enterprise. The dream died in Memphis. The landlord business replaced it. The market is repricing accordingly.

The next six months will test whether the $6.7 billion in contracted cloud revenue converts to recognized revenue on schedule, whether Cursor integrates cleanly, whether Starlink's direct-to-cell deals with T-Mobile and others materialize into ARPU uplift. The war chest funds the experiments. But the editorial headline writes itself: SpaceX doubled revenue by admitting its AI ambition failed and renting the wreckage to the winners. That is a pivot, not a triumph. The stock knows the difference.