SpaceX spent $15.828 billion on capital projects in its AI segment during the three months ended 30 June 2026. That quarterly figure is 4.13 times as large as the $3.832 billion its entire Space segment spent across all of 2025.
The comparison comes from SpaceX’s second-quarter Form 10-Q and the segment note in its 2025 audited accounts. This is one company’s unaudited quarterly report, not a general measure of what AI infrastructure costs. It nevertheless shows how sharply SpaceX’s centre of capital investment has moved.
It also requires an accounting caution. Capital expenditure pays for assets expected to remain useful beyond the current quarter. It is not the same thing as an operating expense, and the full $15.828 billion did not pass through SpaceX’s quarterly profit-and-loss statement.
AI absorbed 86 per cent of the quarter’s capital spending
SpaceX reported $18.369 billion in total capital expenditure for the second quarter. AI accounted for $15.828 billion, or 86.2 per cent. Connectivity accounted for $1.367 billion and Space for $1.174 billion.
The pace had already accelerated. AI capital expenditure was $7.723 billion in the first quarter, meaning it more than doubled in Q2. Across the first six months of 2026, SpaceX put $23.551 billion into AI capital projects, 82.7 per cent of the company’s $28.476 billion total.
When we examined the AI infrastructure pivot buried in SpaceX’s IPO prospectus, it was still largely a statement of direction. The first public quarterly accounts put a price on the opening phase.
“AI compute infrastructure” is more than processors
SpaceX described roughly $15.8 billion as supporting the build-out of AI compute infrastructure. That wording matters. The figure should not be read as a single purchase of Nvidia chips.
Compute infrastructure can include accelerators, servers, high-speed networking, storage, buildings, electrical substations, backup generation, cooling equipment and construction. SpaceX also disclosed that it bought $295 million of Tesla Megapacks during the quarter. Together, these systems turn a collection of chips into a data centre that can reliably sell computing capacity.
The physical result was 1.4 gigawatts of nameplate compute capacity at the end of June, up from 1 gigawatt in March and 400 megawatts a year earlier. SpaceX said it expects to exceed 2 gigawatts by the end of 2026.
Capital expenditure is not a quarterly loss
The AI segment generated $2.561 billion in Q2 revenue, including $1.6 billion of new AI infrastructure revenue, while recording a $1.257 billion operating loss. Depreciation and amortisation in the segment reached $1.885 billion.
Those numbers cannot be compared as though they were entries in the same column. A server or data-centre building is recorded as an asset, then generally recognised as an expense through depreciation over its useful life. Capital expenditure tells us how much capacity SpaceX is building. Operating income tells us how the business performed under accounting costs recognised in that period.
SpaceX said its new compute deployments were achieving payback in less than a year, according to a published transcript of the earnings call. That is a management claim based on early contracts, not a result the 10-Q independently proves. The filing also warns that major cloud customers can generally terminate after an initial ramp period by giving 90 days’ notice.
The Space comparison is real, but not complete
Dividing $15.828 billion by $3.832 billion produces the headline’s “more than four times” comparison. Both figures are segment capital expenditure, so the arithmetic is like-for-like.
It still does not capture every dollar devoted to rockets. Space research and development was $1.076 billion in Q2 alone, much of it linked to Starship, and R&D is an operating expense rather than capital expenditure. Nor does the comparison imply that AI has become four times more important than Space. It says that, during this particular build-out, far more money was being capitalised into AI assets.
That distinction echoes our earlier report on where SpaceX said its IPO money would go: rockets, constellations and compute share one balance sheet, but they are at very different stages of construction.
Musk says the supply chain cannot keep up
SpaceX’s prospectus says orbital AI at scale requires “significantly more” AI chips than are currently available to the company and notes that GPUs and specialised components come from a small group of qualified suppliers.
On the Q2 call, Elon Musk put the constraint more starkly. He estimated that memory production was growing at about 20 per cent a year while demand was rising around 200 per cent. Those are Musk’s figures, not audited industry statistics, and forecasts from an interested buyer should be treated accordingly. His broader point is less controversial: semiconductor factories take years to plan, equip and qualify, while data-centre orders can expand much faster.
That is the rationale behind Terafab. When Musk announced the project in March, he said existing manufacturers were not increasing output quickly enough for the combined needs of SpaceX and Tesla. Terafab is intended to bring logic, memory and advanced packaging under one programme. It is also an enormous manufacturing risk, not an instant answer to today’s shortage.
The terrestrial build comes before the orbital one
Most of this quarter’s spending supported terrestrial compute. SpaceX’s longer-term plan is to deploy computing satellites, an ambition we placed in environmental context when examining its proposal for up to one million AI spacecraft. That orbital system does not yet exist at commercial scale.
SpaceX ended June with $93.522 billion in cash and $6.487 billion in marketable securities after its IPO and a $25 billion notes offering. The capital is available for an unusually fast build. The unresolved question is whether chips, memory, power equipment and paying customers can all scale at the same rate. Q2 shows SpaceX is spending as though it believes they will.