Goldman says SpaceX’s GPU rental boom exposes AI’s revenue gap

SpaceX built enormous data centers to train AI models. Now it rents out the spare horsepower, and business is very, very good.

Goldman Sachs thinks that may be exactly the problem. The bank says compute rental prices sit well above the revenue the AI industry needs to reach profitability, which suggests customers are paying a scarcity premium rather than a sustainable market rate.

The new landlord of the AI boom

SpaceX’s AI segment brought in $2.56 billion in Q2 2026. That marks a 247% jump from the same quarter a year earlier.

The growth comes from renting surplus GPU compute out of its Colossus data centers. Those facilities were originally built to train models for the company’s xAI segment.

The segment also posted its first positive adjusted EBITDA, coming in at $1.1 billion.

The customer list reads like a who’s who of AI. Anthropic is paying roughly $1.25 billion per month under a deal that runs through May 2029. Google is paying approximately $920 million per month. A third, undisclosed customer has signed a $1.11 billion monthly contract set to begin in December 2026.

That new deal pushes SpaceX’s total annualized compute commitments past $40 billion.

Why Goldman sees a warning sign in the rent checks

Goldman’s core argument is about the gap between what compute costs and what AI earns. Rental rates, the bank says, are running significantly higher than the revenue required for the industry to turn a profit.

Goldman attributes that premium to short-term demand and capacity constraints. In other words, prices are high because there simply isn’t enough supply right now, not because the underlying economics justify them.

Goldman notes that elevated rental costs create real challenges for AI lab profitability, particularly as infrastructure budgets tighten.

The bank also points to a much larger structural problem. Goldman estimates hyperscalers face a $230 billion annual revenue shortfall relative to their capital expenditure needs.

Goldman projects SpaceX’s AI revenue will rise from $3.2 billion in 2025 to $322 billion by 2030, roughly a 100x increase. Capacity is expected to scale alongside that. SpaceX had 1.4 GW of compute capacity as of June 2026, with plans to exceed 2 GW by year-end and expand to 5-10 GW in 2027.

A public company with a famous underwriter

SpaceX listed on Nasdaq in June 2026 under the ticker SPCX. Goldman Sachs served as lead underwriter on the offering.

What this means for SpaceX, AI labs, and investors

The most interesting tension in Goldman’s analysis is that SpaceX’s own growth plans could erode the premium it currently enjoys. If scarcity is what props up rental prices, then adding capacity, from 1.4 GW toward a planned 5-10 GW, works against that scarcity.

Customer concentration is the other risk to watch. A large share of SpaceX’s committed compute revenue rests on a small number of very large customers, so the financial health of those buyers matters a great deal.

For investors in SPCX, the near-term checklist is fairly concrete. Watch whether the undisclosed $1.11 billion monthly contract starts on schedule in December 2026, whether capacity clears 2 GW by year-end, and whether rental rates hold up as new supply comes online across the industry.