Goldman says AI labs pay SpaceX a scarcity premium on $40 billion of yearly GPU rent

The three disclosed contracts come to $3.28 billion a month, or $39.36 billion a year ^[15]. Crypto Briefing puts the full commitment book past $40 billion ^[7], so those three buyers account for about 98% of it ^[16]. Reported revenue is much smaller. The segment's $2.56 billion second quarter annualizes to about $10.2 billion, roughly a quarter of the committed figure ^[17]. Anthropic's monthly rate alone, run for a full quarter, comes to $3.75 billion, which is more than the whole segment reported ^[18]. The report does not say when the Anthropic and Google contracts began billing. Even so, most of the rent Goldman is talking about has been signed and has not yet shown up in reported revenue ^[17].

Goldman attributes the premium to short-term demand and capacity constraints ^[9]. It says high rents make it harder for AI labs to turn a profit as infrastructure budgets tighten ^[10]. The bank's own forecast for SpaceX runs from $3.2 billion of AI revenue in 2025 to $322 billion in 2030 ^[12], growth of about 150% a year for five years ^[20]. Spread over the 2 GW SpaceX plans to have by year-end ^[13], today's commitments come to about $20 billion per gigawatt-year ^[21]. If capacity stopped at 10 GW, the top of the 2027 plan, $322 billion would take $32.2 billion per gigawatt-year ^[22]. So the forecast assumes one of two things. Either SpaceX builds far more capacity than it has announced, or rents per gigawatt climb while supply grows five-fold ^[23].

The premium could come out three ways. In the first, new supply pulls market rents down while SpaceX's signed rates hold to term, which in Anthropic's case means until May 2029 ^[4]. The labs would then carry the overpayment. In the second, lab revenue grows into the rent and today's premium becomes the market rate. The third is a buyer under strain asking to renegotiate. With about 98% of the book on three names ^[16], one such request would move the whole segment.

We think the first case is the most likely through 2027. The contracts already signed fix the old prices. The segment's first positive adjusted EBITDA, $1.1 billion, was a 43% margin on second-quarter revenue ^[19], and in our view that gives SpaceX room to sell new gigawatts for less. Against that stands Goldman's own 2030 number, which only works if rents per gigawatt hold or rise ^[22]. Goldman's thesis is wrong if the 5 to 10 GW planned for 2027 ^[13] signs at or above about $20 billion per gigawatt-year ^[21]. That would mean the shortage Goldman calls short-term had lasted well into 2027.