🚀 Space
SpaceX Spent 236% of Its Revenue on CapEx Last Quarter and the Market Blinked
The company's debut earnings beat every consensus estimate. Underneath, one Anthropic contract explains roughly three-quarters of the AI surprise, and Starlink's average revenue per subscriber is falling faster than new subscribers can compensate.
Two hundred thirty-six percent. That is how much of its quarterly revenue SpaceX poured into capital expenditures in the three months ending June 30, 2026, the first financial results it has ever reported as a publicly traded company, and it is neither a rounding artifact nor a restructuring charge. It is the consequence of a strategy that amounts to a $73 billion annual bet: burn through cash faster than any company in history to construct an AI infrastructure empire on top of a satellite internet monopoly, stack orbital compute on top of that, and hope the compounding revenue catches up before the cash pile collapses.
The headlines were bullish. Revenue of $7.8 billion beat the $6.93 billion consensus estimate by 12.5%, EBITDA of $3.5 billion crushed the $2.1 billion forecast, and the AI business, which swallowed xAI in February, generated $1.1 billion in EBITDA against a consensus expectation of negative $21 million. Starlink's operating income swelled 79%. Above the CapEx line, everything pointed to a generational company making a flawless debut.
Below it, the stock dropped 7.5% after hours.
The CapEx Ratio That Has No Peer
Context makes the number stranger, not more normal, because no major technology company in the current cycle approaches 236% of revenue in capital investment. Here is how SpaceX stacks against the hyperscalers it now fights for AI infrastructure dollars:
| Company | Q2 2026 CapEx ($B) | Q2 2026 Revenue ($B) | CapEx / Revenue |
|---|---|---|---|
| SpaceX | 18.4 | 7.8 | 236% |
| Microsoft | ~22 | ~66 | ~33% |
| Meta | ~13 | ~42 | ~31% |
| Alphabet | ~17 | ~94 | ~18% |
| Amazon | ~25 | ~155 | ~16% |
SpaceX's ratio is seven to fifteen times higher than any peer, driven almost entirely by the AI segment, which consumed $15.83 billion of the total, a twenty-one-fold increase from $749 million in the same quarter a year earlier when xAI was still a separate entity burning through Musk's personal capital. The remaining $2.57 billion covered Starlink satellite deployments and Starship development. CFO Bret Johnsen told analysts to expect "similar" spending for the next couple of quarters, which implies an annualized AI infrastructure run rate north of $63 billion, exceeding the annual capital budgets of every company on Earth except Microsoft, Amazon, and Alphabet, and roughly $56 billion more per year than the AI business currently generates in revenue.
The Anthropic Question
The AI segment's transformation from a $2.5 billion operating loss in Q1 to $1.1 billion positive EBITDA in Q2 is extraordinary by any measure, a $3.6 billion swing in ninety days that no tech business segment has ever matched at this scale. The obvious question: how did it happen?
Mostly one contract, according to Barron's, which reported before the earnings call that SpaceX's agreement with Anthropic is valued at $1.25 billion per month and was "ramping up in May and June," while the Google agreement had not yet started and compute contracts with Reflection AI contribute smaller amounts.
Apply the ramp timeline to the quarter and a rough estimate emerges. April revenue from the Anthropic contract: negligible or zero, since the agreement was still scaling infrastructure. May: partial ramp, call it $625 million at 50% utilization. June: full delivery at $1.25 billion. Total estimated Anthropic contribution for the quarter: approximately $1.875 billion.
SpaceX's total AI revenue, inferred from the "roughly 250%" year-over-year growth on a Q2 2025 baseline of approximately $727 million (per S-1 filings), comes to roughly $2.5 billion. If the Anthropic estimate is in the right neighborhood, a single customer generated approximately 75% of AI segment revenue in the quarter.
That concentration deserves scrutiny, because when you remove it the picture inverts. The AI business surprised to the upside by $1.121 billion against analyst consensus, but strip the Anthropic contract and the segment likely would have posted an EBITDA close to expectations, possibly worse. Brian Mulberry of Zacks declared on the earnings call that "AI is already monetizing itself," which is technically true in the way that a restaurant with one table occupied is technically open for business. X subscriptions, Grok enterprise licenses, and the consumer AI products contribute the remaining quarter of revenue, and at current scale those products alone do not cover the electricity bill at the Memphis data center complex, let alone justify the $15.83 billion in quarterly infrastructure investment.
The EBITDA Margin Looks Excellent. Until It Doesn't.
At roughly 44% EBITDA margin on $2.5 billion in revenue, SpaceX's AI business looks like a mature cloud infrastructure provider: AWS runs at roughly 35-37% operating margin, Microsoft's Intelligent Cloud hovers in the low 40s, and if you squint the comparison is flattering.
Stop squinting. Depreciation kills the illusion.
EBITDA excludes depreciation by design, but $15.83 billion in quarterly CapEx creates an enormous charge that shows up below the line and compounds every quarter that spending continues at this rate. Assume a five-year useful life for AI infrastructure, the industry standard for GPU clusters, and the Q2 CapEx alone generates approximately $3.17 billion in annual depreciation going forward, while stacking Q1 and Q2 together at $23.53 billion cumulative pushes the run-rate to $4.7 billion per year, or $1.175 billion per quarter, by December 2026. That single line item would consume the entire Q2 AI EBITDA and then some, repeated every quarter for five years, regardless of how fast revenue grows in the interim.
This is not an accounting trick but rather the fundamental economics of infrastructure businesses that front-load capacity ahead of demand. Amazon did exactly this between 2011 and 2014, when AWS CapEx regularly exceeded AWS revenue, and the market punished Amazon's stock price for years before compounding growth validated the bet and produced a forty-fold return for investors who held through the pain. But here is the difference nobody has quantified: SpaceX's AI CapEx-to-revenue ratio is six times higher than Amazon's worst year during the AWS buildout, which means the magnitude of the wager dwarfs the historical precedent that people cite to justify it.
Starlink's Quiet Ceiling
While the AI narrative dominated the call, the Starlink business deserves its own dissection. It is the only segment generating real cash flow, the only business that actually works, and the math underneath it is quietly changing direction.
Subscribers doubled to 12 million, revenue rose 66%, and by every standard metric it sounds like a growth stock until you read the fine print. Average revenue per user declined 22% year over year, falling from approximately $85 to $66 per month, extending a slide from $99 in 2023. SpaceX is pushing into lower-income international markets and rolling out cheaper plans to drive adoption, a classic land-and-expand play that generates great subscriber charts right up until the moment ARPU hits a floor and subscriber growth decelerates simultaneously.
Here is what the crossover looks like if both trends continue:
| Period | Est. ARPU ($/mo) | Est. Subscribers (M) | Quarterly Consumer Rev. ($B) |
|---|---|---|---|
| Q2 2026 (actual) | ~$66 | 12 | ~$2.38 |
| Q2 2027 (projected) | ~$51 | ~20 | ~$3.06 |
| Q2 2028 (projected) | ~$43 | ~30 | ~$3.87 |
| Q2 2029 (projected) | ~$38 | ~40 | ~$4.56 |
Projections assume subscriber growth decelerates from 100% to approximately 40% by 2029, and ARPU decline slows from 22% annually to 12% as the pricing mix stabilizes. Consumer revenue only; total connectivity includes enterprise, maritime, aviation, and government.
Revenue keeps growing in this model, but the growth rate decelerates sharply from 66% to roughly 30% and then 18%, a trajectory that rhymes with every subscription business that has ever expanded internationally by cutting prices. SpaceX's entire financial architecture depends on Starlink generating enough surplus cash to subsidize AI losses and Starship development until those businesses stand on their own, and if ARPU compression accelerates or subscriber growth slows faster than modeled, the subsidy shrinks accordingly. It is a real subsidy: Starlink's $2.6 billion quarterly EBITDA against the company's $14.9 billion quarterly cash deficit on CapEx alone tells you exactly who is paying for the AI dream.
Shotwell's answer came during the call when she said SpaceX expects to snatch "quite a few" T-Mobile, AT&T, and Verizon customers by building ground-based infrastructure to complement the satellite network and capture developed-market mobile substitution where ARPU is higher. That sounds promising until you consider that it also means competing with entrenched terrestrial carriers on their own turf, with its own capital requirements, regulatory battles, and competitive dynamics, while simultaneously building the world's largest AI data center operation, developing the most ambitious rocket in history, and operating the planet's only commercial LEO constellation, all on $7.8 billion in quarterly revenue and a $541 million net loss.
Nvidia in Orbit
Almost buried in the release was a detail that reframes the entire CapEx conversation: SpaceX announced a partnership with Nvidia to power Starmind AI1 orbital compute satellites, and Musk added that SpaceX expects to receive "a significant percentage" of Nvidia's GPU output next year. Data centers in orbit sounds absurd until you trace the logic of a company that has already committed to spending $63 billion annually on AI infrastructure and needs to decide where to put it.
Orbital compute is not revenue-generating today, and the engineering challenges are real, from cooling GPU clusters in vacuum to powering them from solar arrays and managing thermal cycling across 90-minute orbital periods. But in space, power comes from the sun indefinitely, cooling radiates into the vacuum without chillers or water towers, and latency to every point on Earth equalizes because the constellation already exists overhead. Three hyperscalers have signed letters of intent to lease capacity.
Whether the unit economics close before the ground-based burn rate forces a reckoning is the open question that will define whether orbital compute is a genuine second act or a PowerPoint slide that consumed billions in shareholder capital.
The Lock-Up Overhang
Timing matters here because SpaceX reported earnings two days before 911 million post-IPO shares unlock on Thursday, while another 455 million remain locked behind price-based thresholds that the current stock price has not met. The company raised $75 billion in the largest IPO in history and has since given back 8% of its market value, trading below the $135-per-share offering price.
Strong earnings might have provided cover for an orderly insider exit, but instead the stock dropped after hours, making Thursday's unlock more likely to produce selling pressure than price support. The 455 million price-gated shares are a small mercy: they will not unlock unless the stock rises past undisclosed thresholds, capping the supply overhang at the 911 million figure for now.
The Strongest Case Against
The bull case writes itself, and it is the same case that Amazon bulls made in 2012: build monopolistic infrastructure at a scale that looks economically irrational until the compounding revenue eventually makes it look visionary, because Starlink has no real competitor, the AI segment has signed enterprise contracts with the most sophisticated compute buyers on Earth, and Starship, if it works, restructures launch economics permanently.
The bear case is more specific and harder to dismiss. SpaceX's AI segment is not AWS. Jeff Bezos built AWS inside a company that already had $100 billion in e-commerce revenue to absorb the losses; SpaceX's total revenue is $7.8 billion, it lost $541 million in the quarter, and Starlink, the only cash engine, faces structural ARPU compression that no amount of subscriber growth has yet reversed. The AI business's profitability rests almost entirely on one massive contract that Anthropic could renegotiate, reduce, or replace with its own infrastructure as it scales toward an expected IPO this fall, and the $63 billion annual AI CapEx run rate will generate depreciation charges that make net profitability mathematically impossible for years, even if revenue triples from here.
Here is the arithmetic that the optimists have to answer: SpaceX raised $75 billion in its IPO, and at the current burn rate it consumes roughly $59 billion more per year on CapEx than it generates in EBITDA. That buys approximately fifteen months of runway, assuming no debt draws and no additional equity offerings. Fifteen months. After that, the company either cuts spending dramatically, raises again in a market that may be less accommodating, or achieves AI revenue growth that outpaces the most optimistic projections on Wall Street.
Limitations
The Anthropic contract contribution is estimated from disclosed contract value and ramp timing, not from segment breakdowns that SpaceX has not published. The 75% concentration figure could be higher or lower depending on the actual ramp curve and contributions from Google and Reflection AI contracts. Hyperscaler CapEx comparisons use approximate Q2 2026 figures from recent earnings reports and may shift with final filings. The Starlink ARPU projections extrapolate from a 22% annual decline rate; international market mix, enterprise pricing, and the proposed terrestrial buildout could alter the trajectory materially. The cash burn estimate does not account for debt capacity, working capital changes, or potential equity offerings.
The Bottom Line
SpaceX delivered a debut that beat every estimate, and the stock dropped because investors ran the numbers underneath the numbers. The company spends $2.36 to build every dollar of current revenue, its most impressive segment depends on one customer for three-quarters of its income, and its only profitable operation is a satellite internet service whose per-user revenue falls 22% annually while subscriber counts double.
None of this means SpaceX is wrong. Amazon's AWS bet looked equally irrational in 2012, and anyone who sold on CapEx ratios alone missed a forty-fold return over the next decade, because monopolies get built by companies willing to sustain losses at magnitudes their competitors cannot match, and SpaceX may be executing precisely that playbook at planetary scale.
But the market is not wrong either, because at $1.6 trillion SpaceX is priced for a future in which AI infrastructure, satellite internet, and reusable orbital launch all reach their theoretical potential simultaneously, and the Q2 numbers are a reminder that between here and there lies a canyon of capital spending that somebody has to fund. The list of somebodies is exactly three items long: Starlink subscribers paying less every quarter, Anthropic paying $1.25 billion a month for GPU time, and whatever Nvidia chips arrive next year. If any single leg buckles, the math changes faster than a Falcon 9 reaches orbit.
If you are a SpaceX shareholder: Thursday's lock-up unlock is the near-term risk, so do not panic-sell on after-hours moves but do set a price at which you would reduce exposure if the selling pressure materializes. If you are considering buying: wait for the first 10-Q filing, which will include full segment detail, balance sheet data, and the depreciation schedules the press release omitted, because the Anthropic contract concentration is the single highest-risk variable worth tracking and if Anthropic builds its own infrastructure or renegotiates terms at its own IPO, SpaceX's AI revenue trajectory changes in a quarter.