In its first results since it listed on the stock market in June, SpaceX spent $2.35 on capital for every dollar of revenue it earned, far beyond the most stretched of the five big AI builders. The more interesting number is not how much it spent, but who is paying it.
| Line | Expected | Delivered |
|---|---|---|
| Revenue | ≈$6.8B (analysts) | $7.81B (+92% on a year earlier) |
| AI segment revenue | — | $2.56B (+247%) no growth line drawn: outside the five, so it counts toward no required pace |
| Connectivity revenue | — | $4.29B (+66%); operating profit $1.66B (+79%) |
| Starlink subscribers | analysts expected slightly more than was delivered | 12.0M, doubled on a year earlier; average revenue per user $66, down from $85 |
| Capital spending | ≈$48.7B a year (analysts) | $18.37B in the quarter; at that rate ≈$73.5B a year |
| Net result | — | net loss $541M (from $1,008M); adjusted EBITDA $3.54B (+191%), a profit measure explained below |
| Balance sheet | — | $100B of cash and investments, after ≈$85.7B raised from the share sale and a $25B bond |
| Contracted cloud sales | — | $14.1B signed in the quarter, counting only the part customers cannot cancel |
The three clocks above ask three questions: is AI revenue growing fast enough (demand), is the company still committed to the spending (capex), and are investors still willing to wait (patience). SpaceX is not one of the five biggest AI builders named by the Bank for International Settlements (the BIS, a bank owned by the world's central banks): Alphabet, Amazon, Meta, Microsoft and Oracle. Nothing on this page feeds the front page's required pace of revenue growth, its spending split or its gap chart. SpaceX is tracked beside the model, as evidence of how far the spending ratio can go. Figures are from the second-quarter 2026 results release filed with the SEC; the share reaction and the analysts' forecasts are from the press coverage listed below.
The BIS's own measure of strain, in the chart that opens the AI section of its report, is capital spending divided by revenue. It is the plainest statement of the problem: how much is being laid out now against how much is being earned now. Across the five big builders it has climbed steadily and, in the most extreme case, reached $0.86 of capital spending for every dollar of revenue.
SpaceX spent $18.37 billion against $7.81 billion of revenue. Inside the AI segment the numbers are of a different order: $15.83 billion of capital spending against $2.56 billion of revenue, up from $749 million of AI capital spending a year ago. Twenty-one times as much, in four quarters.
Apply the front page's revenue floor to that single quarter. The floor is the revenue the AI spending must earn every year to pay for itself (the front page's §2). On its standard assumptions, hardware that lasts five years, a 10% required return, and no allowance for running costs at all, $15.83 billion of spending adds about $4.7 billion a year to the permanent bill. The AI segment's entire revenue, at this quarter's rate, is about $10.2 billion a year. Keep building at this pace for four quarters, and one year of building will have committed more revenue each year than the AI business currently earns in total. That is the floor arithmetic at work inside a single company's accounts, in public, at speed.
The obvious objection to keeping a builder this size outside the model: if a sixth company is spending at a pace of $73 billion a year, is the front page's bill too low? Start with what is truly extra. When Alphabet rents computing capacity from SpaceX instead of building its own, the rent counts as one of Alphabet's running costs, not as capital spending. So these machines appear in nobody's capital spending among the five. The hardware is real, and it is not counted twice. On the spending side, the objection stands.
But the front page's bill is not worked out by adding up the five's budgets. The §2 slider follows the BIS's own scale: total AI capital spending committed across the whole sector from 2025 to 2030, between $2 and 4.5 trillion. The five's announced $840 billion only divides that bill up for the gap chart; it does not set it. A sixth builder's spending is part of how the sector climbs from this year's $840 billion toward a committed total in the trillions. So the right effect on the model is to argue that the slider belongs nearer the top of its range. Adding it on top as a new slice would count it against an assumption that already covers the whole sector. The arithmetic is on the slider. Keeping up this quarter's AI spending through 2030 comes to roughly $0.3 trillion, and moving the slider from $3.5 trillion to $3.8 trillion raises the yearly floor by about $90 billion. The front page treated Alphabet's July increase in spending the same way: the evidence moves the slider, and the slider was built to take it.
What truly cannot be added is the revenue. SpaceX's AI revenue is other builders' spending. Treat all the firms in the boom as one business and those payments cancel out, because in the end the floor must be paid by customers outside the build-out. Put a sixth revenue line on the scoreboard and the same dollar, going round in a circle, gets counted twice. That is the imbalance the top of this page points at, and it belongs to the model itself. The capital spending is real and pushes the bill up. The revenue, today, is largely the bill moving from one builder's pocket to another's.
Here is the part that matters more than the ratio. The AI segment's revenue nearly quadrupled, and the release says plainly where it came from: new cloud services agreements, $14.1 billion of contracted sales signed in the quarter, adding $1.6 billion of new revenue. The customers, as reported in the press: Anthropic, renting at least 300 megawatts of computing capacity at the Colossus site in Memphis for around $1.25 billion a month, and Alphabet, paying around $920 million a month from October.
Add those two together and you get roughly $26 billion a year of contracted revenue, against an AI business currently earning about $10 billion a year. On any ordinary reading this is the best growth story in the results. On the front page's reading it is something else. That revenue is not coming out of a company's software budget, and it is certainly not coming out of a payroll. It is coming out of other builders' capital budgets: the same $840 billion the five big builders plan to spend in 2026, which the front page charts as the bill this whole sector has to pay for.
Notice what Alphabet is doing in that sentence. It is one of the BIS's five. It is spending roughly $200 billion of its own on AI infrastructure this year. And it is also renting computing capacity from a sixth builder, at $11 billion a year, which that builder records as growth in AI revenue. Both figures are real. Both get counted. Neither comes from a customer outside the boom.
This is not an accusation of anything improper. Every step is disclosed, and renting capacity from one infrastructure owner to another is ordinary business. It is a measurement problem, and it is exactly the one the BIS raises when it warns about the loop. The loop is firms in the boom investing in, buying from and selling to each other, so that one firm's spending is another's revenue. The boom's scoreboard counts that loop as growth. The front page now draws the loop, with a reported figure on every arrow, and these results supplied three of them.
One number in the release deserves more attention than it got. The cloud services agreements were recorded as $14.1 billion of contracted sales, against contracts the press reports at tens of billions of dollars over their full terms. The release's own definitions explain the gap: contracted sales count only the period the customer cannot cancel and the contract can be enforced. Press reports of the terms show why that matters. The Alphabet agreement is described as one either side can leave on 90 days' notice after 31 December, with fees cut in proportion if the promised capacity does not arrive on time.
So the headline contract values and the recorded order book measure different things, and the company is recording the cautious one. Read that way, the disclosure is a credit to the accounting. Read against the company's valuation, it is a warning: the demand that makes the capital spending look justified is, in the form that can be enforced, a fraction of the number the story is told with.
Adjusted EBITDA, profit before interest, tax, depreciation and amortisation, rose 191% to $3.54 billion, and it is the figure the release leads with. Look at what is added back to get there: $2.85 billion of depreciation and amortisation in the quarter alone, the charge for hardware and other assets losing value as they age. That is up 87% on a year earlier and, at this quarter's rate, more than $11 billion a year.
For most companies, adding depreciation back is a reasonable way to see the underlying business through the accounting. For this one it removes the whole question. The front page's model exists because AI hardware is not railway track: it earns for a few years, not a century, so capital spending turns into a bill that comes back every year rather than a one-off outlay. Depreciation is that bill arriving in the accounts. A measure that adds it back shows the AI segment turning positive on EBITDA: $1.15 billion, from a loss of $276 million a year ago. In the same quarter the segment's operating loss is $1.26 billion and its capital spending is $15.8 billion. Both statements are true. Only one of them is about whether the bet works.
Take away the AI segment and there is a plainly good company here. Connectivity, which is Starlink, turned $4.29 billion of revenue into $1.66 billion of operating profit, a 39% margin, with subscribers doubling to 12 million. It is the profitable engine, and it is not the AI bet.
It also shows the usual sign of a consumer business buying growth: average revenue per user fell from $85 to $66, down 22% in a year, as growth shifted to cheaper plans in other countries. It is taking more customers at a lower price. That is a normal and often correct trade. But it means the part of the company that funds the story is growing revenue more slowly than it is growing customers, at exactly the moment the AI segment's need for capital is growing fastest. Six months of operating cash flow across the whole company came to $3.47 billion. Six months of capital spending came to $28.48 billion. The difference did not come from the business; it came from the June share sale and the bond sale.
Two more details from the balance sheet, offered without a theory attached. Debt stands at roughly $39.4 billion, of which $13.3 billion is owed to related parties, meaning lenders connected to the company, such as its owners or sister companies. $327 million of the quarter's $629 million of interest was paid to them. The release does not say who they are. In a sector whose central question for financial stability is who is funding whom, it is a line worth watching in the next filing.
The quarter also brought an announced agreement to buy Cursor, an AI coding tool, for $60 billion, expected to complete in the third quarter. Set aside whether the price is right. Ask instead the question the front page's §3 asks of every AI product: which pot of money does its revenue come from?
A coding assistant is paid for out of one of two things: a company's budget for programming tools, or its budget for programmers' wages. The first is one line inside a worldwide business software market of roughly $1.2 trillion a year; $60 billion is 5% of one year of that entire market, paid for one product. The second is the wage pot. Prices like this are the clearest evidence there is of which answer the buyers actually believe. They are not being paid on the assumption that programmers keep their jobs and simply get a better tool.
Three things to watch, in order of how much they would move the argument.
Who the customers are. If the next few sets of results keep growing AI revenue from cloud services agreements with other AI builders, the growth rate is not evidence of demand from end customers. It is evidence about capital budgets, and it is as fragile as they are. The number that would change the reading is AI revenue from customers who are not themselves building AI infrastructure. It is not disclosed today.
How much of the contracts becomes binding. $14.1 billion of sales that cannot be cancelled, against much larger reported contract values, is the honest version of the order book. If that recorded figure grows toward the headline numbers, the demand is firming up. If it does not, the gap is telling you something.
The growth rate the company has set itself. On the results call, management said its internal projections for reaching $1 trillion of revenue had moved forward from 2031 to 2030, with a chance of 2029. From this quarter's rate of roughly $31 billion a year, $1 trillion by the end of 2030 means growth of about 138% a year, kept up for four years. The growth rate the front page requires of the whole sector, the one it calls demanding, is about 40%. A stated target is not a forecast, and this page does not treat it as one. But it is the single clearest statement of what the people spending the money believe, and it is now on the record where it can be checked against each set of results.
The July US jobs report, the other side of the ledger and the front page's test of whether AI revenue and jobs are moving apart, landed on 7 August. Nvidia, whose sales show how much the builders are spending, and which sits at the far end of this quarter's loop, reports on 26 August. Oracle's next results are due around mid-September. Earlier this results season: Alphabet, Microsoft, Meta, Apple, Amazon.