Beyond rockets and satellites, SpaceX is quietly building an AI compute business that may become key to its eye-popping valuation

The real read-through is that AI scarcity is still being priced at the margin of infrastructure, not model quality. That keeps GPU toll-collectors like NVDA in the sweet spot: even if hyperscalers slow headline capex, third-party leasing signals the installed base is still being squeezed for every available hour, which supports utilization, pricing, and replacement demand. The second-order loser is any AI platform whose economics depend on perfect in-house utilization; incremental external compute spend is an opex leak before it is a growth enabler.

For GOOGL, this is a mixed signal. Near term it validates that frontier-model compute remains scarce enough to justify expensive external sourcing, which is bullish for product velocity, but it also implies AI competition is still forcing a high-cost race that can pressure operating margin if external capacity becomes a recurring dependency. META’s strategic optionality is more interesting: if it can monetize excess capacity the way this article implies, data-center assets become less cyclical and depreciation can be partially offset by third-party leasing, but only if power and network constraints allow true surplus capacity.

The contrarian risk is duration. The 90-day escape hatch makes this look like a spot-market arbitrage on scarcity, not yet a durable annuity, so the market may be overpaying for a narrative of structural recurring revenue. Over 1-3 months the trade is still bullish for compute vendors; over 6-18 months it depends on whether Stargate-style supply additions and new build-outs compress pricing. A sustained decline in GPU lead times or a sharp step-down in AI capex guidance would falsify the scarcity thesis quickly.