SpaceX clears the revenue bar, then trips over its own AI bill

SpaceX has spent 24 years being judged on whether its rockets land. On August 4 it was judged, for the first time, on a spreadsheet. The numbers were better than almost anyone expected, and the shares fell anyway.

Revenue for the quarter ended June 30 came in at USD 7.81 billion, up 92% from USD 4.1 billion a year earlier and roughly USD 900 million ahead of the USD 6.93 billion consensus.

The loss per share was nine cents against an expected 26 cents. Net loss narrowed to USD 541 million from USD 1 billion. Adjusted EBITDA almost tripled to USD 3.5 billion.

All three divisions beat their estimates. Connectivity, which houses Starlink, brought in USD 4.29 billion, up 66%, and delivered USD 1.66 billion of operating income. It remains the only profitable part of the company.

The AI segment, which now folds in Grok, the X platform and a fast-growing cloud rental business, posted USD 2.56 billion, up 247% on the year and 213% on the previous quarter. The Space segment, still the part most people picture when they hear the name, was smallest at USD 962 million.

There was plenty else to like. Starlink passed 12 million subscribers, double a year earlier and up 1.7 million in three months. The company ended June with about USD 93.5 billion in cash and equivalents, against USD 24.7 billion in March, and a backlog of USD 47.5 billion.

It has signed USD 14.1 billion of contracted cloud services agreements, with Google and Anthropic among the tenants, and won over USD 6 billion of multi-year US government Starshield work.

Then investors read the cash flow statement.

The bill for the AI ambition
Capital expenditure in the quarter was USD 18.37 billion. Analysts had modelled around USD 13.2 billion. Of that total, USD 15.83 billion went straight into AI infrastructure, roughly double what the market had assumed.

Management then said third and fourth quarter spending would run at broadly similar levels, implying something near USD 65 billion for the full year against a Street estimate closer to USD 50 billion. No formal guidance was issued, only long-term targets, leaving analysts to do the arithmetic themselves.

Debt and finance leases climbed to USD 36.8 billion from USD 22 billion three months earlier.

The AI unit still lost USD 1.26 billion at the operating line, though that was 49% better than the first quarter, and its adjusted EBITDA turned positive at USD 1.15 billion for the first time. Advertising revenue at X (formerly Twitter) fell 14% to USD 367 million as advertisers migrated onto a new system.

Chief financial officer Bret Johnsen pushed back, telling analysts that on the compute side the company is “getting less than a one-year payback”. The market was not convinced. Shares closed 9.4% higher at USD 125.33 before the release, fell around 8% after hours and slid again on Wednesday (August 5).

The uncomfortable detail sits underneath. Starlink is funding this build-out, and Starlink’s own unit economics are softening. Average revenue per user was USD 66, flat on the first quarter but down from USD 85 a year ago, as growth shifted towards lower-income international markets.

A monthly kit fee introduced in June has not yet flowed through, so the third quarter will be the first clean read on whether pricing action can arrest that slide.

Elon Musk used the call to double down rather than reassure. SpaceX will build its AI infrastructure exclusively on Nvidia silicon, he said, praising the Vera Rubin architecture and confirming the company is, in his words, “exclusive to Nvidia”.

He expects to end 2026 with over two gigawatts of compute and approach 10 gigawatts by the end of 2027. The two companies will also co-design the Starmind AI1 satellite compute payload, packing Nvidia Rubin GPUs and Vera CPUs into orbiting data centres. Nvidia shares rose on the news. SpaceX shares did not.

The roadmap, and why satellite count is a financial story
The next Starship flight, IFT-14, is planned for the end of August, and it matters more than a test number suggests.

For the first time the vehicle will attempt to place operational Starlink V3 satellites into orbit rather than on a suborbital arc, and, subject to regulatory clearance, SpaceX will try to catch the upper stage back at Starbase.

Musk said last month’s flight convinced him the heat shield problem is essentially solved, and suggested cadence could reach one flight a day within a year.

Satellite count is where the space story becomes a financial one. Each V3 satellite carries well over a terabit per second of downlink capacity, an order of magnitude beyond the current generation, and it is sized for Starship, not Falcon.

More capacity added per launch is the mechanism by which Starlink can keep signing subscribers in busy markets without degrading service, and it is the only credible path to stabilising ARPU.

He argued heavier capital spending is a feature rather than a bug this early in a company’s curve, pointed to the rising contracted run rate, and put a 70 per cent probability on a 2028 Moon landing rising to 95 per cent by 2029, which he thinks gives the shares a permanent bid.

JPMorgan’s Doug Anmuth was more measured. In a note published after the call he modelled capital spending approaching USD 200 billion in both 2027 and 2028, which he said would pressure free cash flow in a pattern now familiar across the hyperscalers.

He also flagged Thursday’s lock-up expiry, which frees up to 911.5 million insider shares against a float of roughly 639 million, a potential increase of 143%, though he thinks much of the positioning has already happened. JPMorgan nudged its target up to USD 240 from USD 225.

The rest of the Street scattered. Morgan Stanley’s Adam Jonas has stayed Overweight with a USD 300 target on the view that the AI business is undervalued. Piper Sandler trimmed to USD 140 from USD 156 and Wells Fargo to USD 215 from USD 230, while Bank of America turned more constructive and UBS reiterated a buy.

ALSO READ | Can Tesla afford its robot dreams? What the Q2 numbers really show

The consensus target sits near USD 223, with a spread running from USD 62 to USD 800, which is less a forecast than an admission that nobody has a settled model yet.

Melissa Otto of S&P Global’s Visible Alpha attributed the fall squarely to AI capital spending at more than double expectations, and Phillip Capital expects the group to move into a net debt position of about USD 3.3 billion by fiscal 2030.

What to watch
Four things over the next few weeks. Thursday’s unlock, and how much stock arrives. The IFT-14 outcome, which validates or delays the V3 capacity thesis. Any move towards formal capital expenditure guidance, the absence of which is damaging sentiment.

And the third quarter connectivity margin, the first period to capture June’s price rise. Short interest is near 34% of the float, so the market has already picked a side. The company now has to launch its way out of the argument.