SpaceX's AI unit turned a profit. It also put $2.18 into buildout for every $1 it brought in

A new name, and a new number to go with it

Start with the footnote, because SpaceX did. "Super Intelligence (SI) refers to the operating and reportable segment previously reported as Artificial Intelligence (AI)," reads Note 2 of the third-quarter release published October 20.¹ Same segment, bigger name. And, for the first time, a number worth bragging about: SI revenue of $7.56 billion, up from $2.56 billion in the second quarter and $813 million a year earlier, which the company counts as 195% sequential and 830% year-over-year growth.¹ The CFO calls it "a clear inflection in our Super Intelligence business."¹

Inflection is the right word for the revenue line. The question is what is bending, and what it costs to bend it.

How much of the jump was bought

SpaceX closed its acquisition of Cursor, the AI coding company, on August 14, at an implied equity value of $60 billion. From closing through September 30, Cursor contributed $608 million of revenue to the company's consolidated results.⁴ The release lists three drivers for the SI jump: the activation of new cloud services agreements, Cursor, and wider Grok adoption.¹

Take Cursor's full $608 million out of the segment, on the assumption that all of it landed there, and SI still did about $6.95 billion, roughly 2.7 times the prior quarter.⁵ So the acquisition explains a slice, not the story. The bulk sits in one line, "SI solutions & infrastructure," which went from $2.19 billion to $7.02 billion in three months. Advertising, the old business, added $538 million.¹ That is the signature of selling compute to other people: cloud contracts that switch on, and revenue that arrives in blocks. It is real money. It is also lumpy money, and the release does not name the customers or say how much of the quarter one or two of them account for.

The profit, three ways

SI's income from operations was $431 million, against a loss of $1.26 billion in the second quarter.² On $7.56 billion of revenue, that is a margin of about 5.7%.⁵ The release adds a fair caveat in SpaceX's favor: the $431 million is after $1.2 billion of share-based compensation tied to the Cursor acquisition, the cost of keeping Cursor's people from walking.²

Then comes the number that will travel: SI Adjusted EBITDA of $4.95 billion. The reconciliation table shows how $431 million becomes $4.95 billion. Add back $2.47 billion of depreciation and amortization, $821 million of ordinary share-based pay, $1.22 billion of acquisition-related share-based pay, and $11 million of deal and restructuring costs.² Every add-back is disclosed, and the definitions are standard for the genre. But in an AI segment, depreciation is not a rounding item. It is what the data centers cost, spread over their useful lives. An EBITDA figure for a compute business is a figure from which the compute has been removed.

The bill

SI capital expenditures were $16.44 billion in the quarter, after $15.83 billion in the second. Over nine months, the segment spent $40.0 billion on capex while booking $10.9 billion of revenue.³ That is $2.18 of buildout for every dollar of sales in the quarter, and about $3.66 over nine months.⁵ Nameplate compute reached 1.7 gigawatts on September 30, from 1.4 GW at the end of June and 0.6 GW a year earlier.³

At the company level, the cash flow statement shows how the gap gets filled. Over nine months, SpaceX generated $13.1 billion of operating cash and spent $48.1 billion on property and equipment across all segments; the same statement lists $85.7 billion of net IPO proceeds and $51.8 billion of new debt and financing.⁶ The machine is being built with other people's money, which is what capital markets are for. It is not, yet, being built with its own.

The serious objection

Here is the best case, and it deserves its hearing. Capex is not an expense. A data center bought this year earns for several years, so comparing one quarter's capex with one quarter's revenue flatters nobody's thesis and proves little. Nor is the demand imaginary: deferred revenue rose by $4.9 billion over nine months, which is customers paying ahead for things not yet delivered.⁶ If the cloud contracts are long and the customers solvent, today's ratio is the shape of a buildout, not a verdict on it.

We accept most of that. It is why we would not call the segment a money pit on this evidence. But the objection proves demand, not returns. Whether $40 billion of nine-month capex earns its cost depends on contract length, pricing, utilization and how fast the chips age, and the release discloses none of those. The CFO's commentary promises "strong returns on invested capital."¹ That is a promise. The table that would turn it into evidence, return on the SI segment's invested capital, is not in the release.

Who carries the risk

The quarter's headline net income of $4.4 billion also deserves a second look: it includes $1.03 billion of other income, $873 million of interest income and a $390 million tax benefit, on top of $2.79 billion of consolidated operating income.⁷ None of that is hidden. All of it is below the line that measures the business.

So the honest reading is narrower than the press-release reading, and more interesting. SpaceX has proved that buyers will pay billions for its compute and its AI products, quarter after quarter. It has not yet proved that the compute pays for itself. Shareholders who just funded an $85.7 billion IPO are the ones holding that bet, and the scorecard they need is the one the release leaves blank.