SpaceX spent 13 times more on AI than on rockets last quarter

SpaceX spent $1.17 billion on its rocket business last quarter. It spent $15.83 billion on AI compute.

That ratio — 13.5 to one — is in the company’s own quarterly filing, and it is the single most useful fact about the largest listed company most investors still describe as a launch provider. Of $18.37 billion in total capital expenditure for the June quarter, 86.2% went to the AI segment. AI capex alone was more than twice the entire company’s revenue.

SpaceX is not a rocket company that dabbles in AI. On the evidence of where the money goes, it is an AI infrastructure company that also flies rockets.

Article Brief

Key Takeaways

5 Points30s Read

1. 01The ratio–SpaceX spent $15.83 billion of capital on its AI segment last quarter and $1.17 billion on Space — 13.5 to one, and 86.2% of all capital expenditure.
2. 02Who earns–Connectivity (Starlink) is the only profitable segment, at $1.66 billion of operating income. Space lost $542 million and AI lost $1.26 billion.
3. 03How fast AI is growing–AI revenue rose 247% year-over-year to $2.56 billion, with AI solutions and infrastructure revenue up sevenfold to $2.19 billion. Compute capacity reached 1.4 gigawatts.
4. 04The new contract–The CFO disclosed an AI compute agreement worth about $1.11 billion a month from December 1 — roughly $13.3 billion a year — from a customer he did not name.
5. 05How it is funded–The June IPO raised about $85.7 billion net and a $25 billion bond followed two weeks later. The quarter ended with $100 billion of cash and a $47.5 billion backlog.

The segment table nobody reads

SpaceX reports three segments, and they behave nothing alike. From the second-quarter results:

Space generated $962 million of revenue, up 29%, and lost $542 million from operations. It is the smallest segment — 12.3% of company revenue — and it is where the rockets are.

Connectivity, meaning Starlink, generated $4.29 billion, up 66%, and produced $1.66 billion of operating income. It is the only segment that earns anything.

AI generated $2.56 billion, up 247% year-over-year and 213% sequentially, and lost $1.26 billion from operations — while consuming $15.83 billion of capital.

Total revenue was $7.81 billion, up 92% — about $1 billion ahead of Wall Street. The company lost $143 million from operations and $541 million after everything else, an improvement on the $1.0 billion it lost a year earlier.

What the AI segment actually is

The detail underneath is more striking than the headline. AI solutions and infrastructure revenue went from $311 million a year ago to $2.19 billion — a sevenfold increase — driven by Cloud Services Agreements that the company says total $14.1 billion in contracted sales.

Compute capacity reached 1.4 gigawatts, up from 0.4GW a year earlier, with the build-out of Colossus II and more under construction. The segment turned its first positive adjusted EBITDA, $1.15 billion, against negative $276 million a year ago.

Two other lines in the same filing deserve more attention than they got. SpaceX announced an agreement to acquire Cursor for $60 billion, expected to close in the third quarter. And it released Grok 4.5 in July, described as incorporating a 1.5-trillion-parameter foundation model and trained alongside Cursor.

Meanwhile the advertising line inside that segment — the X business — shrank, from $426 million to $367 million. The AI segment is not growing because of advertising. It is growing because it sells compute.

The segment that pays for everything has a pricing problem

Starlink is the reason SpaceX is not burning far more cash than it is, so its internals matter.

Subscribers reached 12 million at the end of June, double a year earlier. Connectivity revenue grew 66%. Those two numbers cannot both be flattering, and the filing shows why: average revenue per user was $66, flat on the prior quarter and down from $85 a year ago — a decline of 22%.

That is what scaling a consumer broadband network looks like. Growth increasingly comes from cheaper markets and cheaper tiers, and each new subscriber is worth less than the one before. The segment is still compounding, and operating income grew 79%, faster than revenue, so the operating leverage is real. But the mix is shifting under it.

The offsetting strength is on the enterprise and government side, where SpaceX was awarded over $6 billion in multi-year US government contracts for Starshield, its secure satellite network, primarily from two Space Force contracts for communications and sensing constellations. Government revenue does not churn like consumer broadband, and it does not have an ARPU problem.

The money came from somewhere specific

A $15.8 billion quarterly capital programme requires funding, and SpaceX raised it in a fortnight.

The June IPO sold 638,888,888 Class A shares for net proceeds of approximately $85.7 billion. Two weeks later the company closed a $25 billion inaugural bond issuance of investment-grade senior notes across five tranches maturing between 2031 and 2056, at a weighted average interest rate of 5.855%.

That is roughly $111 billion raised in June. The quarter ended with $100 billion of cash and marketable securities and a $47.5 billion backlog.

So the capital is there, and this is not a company straining to fund its buildout. But it does clarify what the IPO was for. Investors who bought the largest offering in history were, in practice, funding a compute buildout — one that was 86% of capex in the first full quarter after listing.

TECHi made a version of this argument at the listing, when we described SpaceX as priced as an orbital infrastructure stack where only one layer earns. The Q2 numbers sharpen it: the layer that earns is Starlink, and the layer absorbing the capital is the one that was barely discussed in the prospectus coverage.

The customer nobody will name

The most recent development arrived on Thursday, and it is the largest single contract in the company’s history.

Speaking at the Goldman Sachs Communacopia and Technology Conference, chief financial officer Bret Johnsen disclosed a new AI compute hosting agreement. “We closed another hosting deal, and that translates into about $1.11 billion a month starting December 1st of this year,” he said, according to Benzinga’s account of the session. That is roughly $13.3 billion annualised.

Johnsen added that it gave him “even more conviction now about that $100 billion ARR target by the end of this year.” He did not name the counterparty or provide further terms. SpaceX’s existing compute agreements include Alphabet and Anthropic.

Set that against the reported numbers and the scale is obvious. A single undisclosed customer will pay SpaceX more per year than the company’s entire trailing twelve-month revenue of $23.0 billion — 58% of it, to be precise. The existing Cloud Services book already includes Anthropic and Google.

Concentration of that order is a legitimate thing to want disclosed. It is also, for now, the main bridge between $23 billion of trailing revenue and a $100 billion ARR target.

What the stock has done with all this

SPCX closed Friday at $151.21, up 2.0% on the day.

The path since listing has been violent. It priced at $135, closed its first session at $160.95, peaked at $211.39 on June 16, bottomed at $108.27 on August 5, and has recovered to current levels. That is a 49% drawdown and a 40% rally inside three months.

The event most expected to cause trouble was the September 9 share unlock, when restricted stock became tradeable. TECHi flagged this risk in July when the stock hit an all-time low a day after joining the Nasdaq-100. What actually happened was one uncomfortable session — down 3.9% on 120.6 million shares against a recent baseline nearer 50 million — followed by two days of recovery. Anticipated supply, like anticipated index demand, tends to be priced before it arrives.

At $1.95 trillion, the company trades at roughly 85 times trailing revenue. In TECHi’s ranking of 41 AI stocks this quarter, the enterprise software names that led did so by converting AI narrative into disclosed revenue. SpaceX’s AI segment is doing exactly that, faster than almost anything in that study — from a much smaller base, and at a capital cost nobody else is carrying.

What would change the read

Whether the compute economics hold. AI adjusted EBITDA turned positive this quarter for the first time, at $1.15 billion, while the segment still lost $1.26 billion at the operating line. The gap between those two numbers is depreciation on $23.6 billion of six-month capital spending. That depreciation does not go away; it arrives later.

Whether the unnamed customer is disclosed. A $13.3 billion annual contract from an unidentified counterparty is a material concentration. The same structure has been running through our reporting all week — Oracle’s $664 billion backlog is roughly half attributable to one AI lab. Counterparty risk in this cycle is not hypothetical.

Whether Starship changes the Space segment. A flight is expected from Starbase no earlier than September 15. SpaceX argues Starship will cut the cost to orbit by 99% or more, which would transform the segment currently losing $542 million a quarter.

The programme is further along than the segment’s economics suggest. Flight 12 in May completed the first V3 suborbital mission, lifting off from a new Starbase pad, landing the upper stage precisely and deploying modified V2 Starlink satellites. Flight 13 in July hit every objective: 20 production V3 satellites deployed, an in-space relight of a Raptor engine, and what the company calls the softest splashdown yet, returning views of an intact heatshield. Across the first half SpaceX flew 78 launches and put 1,041 metric tons into orbit — most of it its own Starlink hardware.

That last detail is the one to sit with. The launch business is largely flying for the connectivity business, which is funding the AI business. Vertical integration is the official term. It also means the segment losing money is the one that makes the other two possible.

That last one is the part the market still prices as the story. The filing suggests it is now the smallest of the three.