SpaceX Spent $15.8 Billion on AI in a Quarter. Here's What Happens to the Stock if Orbital Data Centers Don't Work.

SpaceX allocated approximately $15.8 billion of $18.4 billion in Q2 capital expenditures to AI, versus $1.4 billion for Starlink and $1.2 billion for its space and rocket business. The company sees orbital data centers as a potential long-term solution to AI power and cooling constraints, but the technology remains unproven and received limited attention in its Q2 disclosures. Management's near-term emphasis appears to be terrestrial AI data-center infrastructure, limiting the likelihood that orbital data centers become a meaningful performance driver soon.

Analysis

The investable read-through is terrestrial AI infrastructure, not space computing. If the reported spending mix is real, the near-term revenue pool accrues to power-delivery, electrical-equipment, cooling, networking and accelerator vendors before it accrues to any orbital-data-center narrative. VRT, GEV, ETN and CEG have more direct 1-3 quarter sensitivity to incremental campus construction; NVDA benefits only if the build uses merchant GPUs rather than internally designed accelerators or a highly customized architecture.

The key second-order issue is funding intensity. A sustained buildout at this scale would raise depreciation, power-procurement and financing burdens well before utilization is proven, making the sponsor's valuation increasingly sensitive to AI-service revenue and external capital-market conditions over 6-18 months. The consensus risk is treating orbital computing as a scarce strategic asset today: launch cadence, radiation hardening, servicing, interconnect latency and unit economics make it an option value, not a base-case earnings contributor. Any premium assigned to that concept should be discounted until management discloses a deployed-MW target, customer contracts, compute utilization and returns on invested capital.

Company-reported capex and segment attribution require independent verification before a directional SPCX position. A confirmation of large terrestrial power commitments or named GPU/network procurement would be a stronger catalyst than additional conceptual discussion; conversely, a cut to AI capex, delayed grid interconnection, or evidence that spend is largely non-recurring development would unwind supplier enthusiasm quickly. The immediate trade is therefore a selective infrastructure basket, while avoiding extrapolation into satellite and launch peers whose earnings are not mechanically linked.

Key Decisions for Investors

• Initiate a 3-6 month basket long in VRT, GEV and ETN on confirmation of new AI-campus power or equipment orders; target 15-25% upside versus 8-10% downside, with a stop if management commentary points to customer project deferrals or order-book conversion weakens.
• Maintain NVDA as a watch rather than treating the reported spend as incremental GPU revenue. Upgrade only if procurement disclosures identify merchant accelerators and networking content; falsify a bullish read-through if custom-silicon adoption or accelerator supply commitments displace NVDA content.
• Do not underwrite a long SPCX solely on orbital-data-center optionality. Require disclosed deployed capacity, signed external demand, launch/maintenance economics and a credible return-on-capital framework; absent those data, any valuation uplift is vulnerable over the next 6-18 months.
• Pair a long VRT or GEV against a broad software/AI-theme ETF rather than a satellite peer if seeking expression: physical-infrastructure revenues are closer to construction spend, while software multiples remain more exposed to utilization disappointments and financing-rate shocks.