Summary

  • SpaceX reported $15.828bn of AI capital expenditure in the second quarter, up from $7.723bn in the first quarter and $749m a year earlier.
  • AI investment represented about 86% of the company’s $18.369bn total quarterly capital expenditure, making compute infrastructure the dominant use of new capital.
  • The AI segment produced $2.561bn of revenue and a $1.257bn operating loss, while reporting $1.146bn of non-GAAP adjusted EBITDA.
  • Depreciation of $1.885bn and stock-based compensation of $516m help explain why positive adjusted EBITDA does not mean the business was profitable on an operating basis.
  • Nameplate compute capacity reached 1.4GW, up from 1.0GW in the first quarter and 0.4GW a year earlier, but the filing does not disclose utilisation.
  • SpaceX cited $14.1bn of contracted sales and $1.6bn of incremental Q2 AI-infrastructure revenue, alongside customer concentration and agreements generally terminable by either party on 90 days’ notice after initial ramp.

The balance sheet can fund the race; it cannot manufacture utilisation

At roughly $100bn in cash, cash equivalents and marketable securities, SpaceX has more room than most infrastructure builders to place capacity ahead of demand. Its reported backlog of $47.5bn adds another measure of future work. Those numbers explain how the company can more than double AI capital expenditure from one quarter to the next without an immediate financing crisis.

They do not answer whether each dollar earns an adequate return. AI capex of $15.828bn was more than six times the segment’s $2.561bn quarterly revenue and accounted for about 86% of all company capital spending. That ratio is not itself a project-return calculation: assets serve multiple periods, construction timing differs from revenue recognition, and some capacity may still be ramping. It does, however, define the size of the conversion task.

The company must transform hardware, power systems and facilities bought today into contracted, utilised service over several years. Cash lowers refinancing risk during that process. It does not remove technology obsolescence, customer churn or the possibility that installed capacity runs below plan.

Nameplate gigawatts are supply, not proof of demand

SpaceX reported 1.4GW of nameplate compute capacity, compared with 1.0GW in the first quarter and 0.4GW a year earlier. The increase demonstrates execution on physical deployment. Yet nameplate is the maximum designed capacity of the installed base, not a disclosed measure of workloads actually running, billable output or power drawn.

That distinction matters because utilisation determines the economic bridge from capital to revenue. A lightly used gigawatt still depreciates. It can still require staff, networking, cooling and power commitments. If demand arrives later than the infrastructure, fixed costs are recognised while customer revenue waits.

Investors therefore need a capacity waterfall: energised capacity, available capacity, customer-committed capacity and revenue-generating capacity. The filing supplies the top-line nameplate figure but not that progression. Until utilisation is disclosed, 1.4GW should be read as demonstrated supply capability and an operating exposure, not as proof of a fully absorbed market.

Positive adjusted EBITDA sits above a large operating loss

The AI segment reported $1.146bn of non-GAAP adjusted EBITDA while recording a $1.257bn operating loss. These two figures do not contradict each other because they answer different questions. Adjusted EBITDA excludes costs that remain economically important to an asset-heavy business.

Depreciation was $1.885bn and stock-based compensation $516m. Their scale explains much of the distance between the adjusted measure and operating profit. In a rapidly expanding compute fleet, depreciation is not an incidental accounting nuisance: it is the scheduled recognition of assets whose useful economic life can be challenged by faster chips and changing system designs.

Adjusted EBITDA can indicate that operations generate cash before certain charges and financing choices. It cannot establish that the deployed capital has earned its cost. The more useful sequence is revenue, direct operating cash contribution, depreciation, sustaining capital and then return on invested capital. SpaceX has disclosed enough to show early operating traction, but also enough to show that profitability depends on the future utilisation of a much larger asset base.

Contracted sales narrow the gap, but termination rights keep it open

SpaceX says it has several compute service agreements representing $14.1bn of contracted sales. Those agreements generated $1.6bn of incremental AI-infrastructure revenue in the quarter. The figures show commercial demand is not merely speculative. They do not mean $14.1bn has already become recognised revenue, and they do not disclose the margin, delivery schedule or capital attached to each contract.

The 10-Q also says the AI infrastructure business depends on a small number of customers and that agreements are generally terminable by either party on 90 days’ notice after the initial ramp. “Generally” matters: it describes the disclosed contract pattern, not necessarily an identical clause in every agreement. Even so, it creates an asymmetry. Infrastructure is long-lived and largely sunk once installed; a customer may retain a comparatively short route out after ramp.

This makes customer quality and switching economics central. Contracted sales provide visibility only to the extent that workloads are costly to move, service is differentiated and customers prefer continuity to cancellation. A backlog supported by cancellable commitments is valuable, but not equivalent to an irrevocable take-or-pay obligation.

Customer concentration transfers technical progress into counterparty power

A small customer base can accelerate an early build-out. Large buyers can commit workloads, help standardise systems and absorb entire sites. It also gives those buyers leverage over price, configuration and renewal. If one customer delays a model programme, adopts a different chip stack or internalises capacity, the revenue effect can be disproportionate.

The concentration risk is magnified by the pace of investment. Assets ordered for one architecture may be less attractive to another buyer, and specialised network or power configurations can reduce redeployment flexibility. The company’s economic moat therefore depends not only on owning compute, but on making that compute portable across customers and workloads.

Axios reported that Elon Musk said on the earnings call that the infrastructure was intended to use NVIDIA systems, including Vera Rubin. This is reported speech about intended architecture, not disclosure of a specific purchase volume, price, duration or exclusive supply contract. It signals concentration on a technology path; it does not quantify the commercial obligation to NVIDIA.

This quarter’s event is terrestrial economics, not an orbital announcement

SpaceX has previously attracted attention for plans involving compute beyond conventional terrestrial facilities. The material current action is different: the 4 August earnings release and 10-Q quantify second-quarter capital deployment, segment revenue, loss, capacity, contracted sales and contract structure.

Conflating the two stories would obscure the immediate investment case. The question raised by the filing is how a large, fast-growing installed base converts into durable customer payments. Whether future infrastructure is terrestrial or eventually uses another physical setting does not change the disciplines of utilisation, depreciation, customer concentration and contract enforceability.

The new evidence should therefore reset the monitoring frame. Announcements show ambition. Quarterly filings show whether capital, capacity and revenue are beginning to meet. Here, capital has moved far ahead; revenue has begun, but the disclosed operating loss and missing utilisation rate show that the conversion is incomplete.

Sources