xAI signed rental agreements worth a combined $2.17 billion per month with Anthropic and Google, and the cleanest way to read that number is not as an AI strategy but as a capital return mechanism ahead of SpaceX’s IPO. That is the argument Martin Alderson put forward on June 8 at martinalderson.com, and it is worth taking seriously.
The Anthropic deal, announced in early May, runs to $1.25 billion per month for 300MW of capacity at the Colossus 1 facility in Memphis, roughly 220,000 GPUs. Google’s agreement, disclosed last week, adds $920 million per month for 110,000 GPUs. Both contracts include 90-day cancellation clauses after an initial lock-in period. The cancellation terms matter: neither Anthropic nor Google is betting their infrastructure future on xAI. They are renting spare room in a building their rival happens to own.
The financial structure underneath this is what Alderson calls a REIT-style payback. At current contract values, and setting aside operating costs (power at Memphis runs to roughly $90 to $160 million annually on 300MW, which Alderson notes is under 1 percent of the Anthropic revenue alone), xAI would fully recover its datacenter capital expenditure in approximately 18 months. The capex recovery does not consume the remaining GPU inventory; xAI still holds hundreds of megawatts of additional capacity beyond what these two deals cover.
That math reframes what xAI built. Colossus 1 was constructed in 122 days, a build pace the hyperscalers cannot match. SpaceX’s project execution capability turns out to be the genuine competitive moat here, not Grok’s benchmark performance. The facility was apparently over-specified relative to actual Grok inference demand, which Alderson suggests may have been the trigger: monetize the excess while the training-run lottery continues.
Grok sits in an odd position inside this structure. Capacity that was notionally reserved for model development is now rented to direct competitors. Alderson does not write off Grok as a model, but he is clear that xAI’s positioning as a frontier lab is hard to sustain when its primary revenue source is serving the compute needs of rivals who are themselves positioning as frontier labs.
The broader capital story compounds this. Anthropic and OpenAI are both approaching public markets as AI research organizations, pitching trillion-dollar valuations on model capability and distribution moats. xAI and SpaceX, merged in February, are approaching the same institutional investor pool as infrastructure. Three structurally different pitches, all competing for the same capital allocation. The SpaceX IPO will price xAI’s GPU fleet as a durable asset regardless of where Grok ranks on future benchmarks.
Alderson is careful to acknowledge the deal has legitimate interpretations beyond financial engineering: compute remains genuinely scarce, Anthropic faced real capacity pressure severe enough to introduce peak-hour usage rationing this spring, and the geopolitical disruptions hitting competing datacenter builds (Iranian drone strikes have already damaged UAE facilities) make xAI’s US-based, rapidly built inventory more valuable by the month. He concludes that all three forces are real; the question is their relative weight.
The structural read still holds even if you accept the benign version of every motive. A company that generates $2 billion per month renting GPUs to its AI-lab competitors, recovers its full capital outlay in a year and a half, and retains hundreds of megawatts of unallocated capacity is an infrastructure business. The lab is the branding layer.
Operators evaluating long-term inference contracts should track whether Anthropic renews or exercises its 90-day exit clause at the first available window; that decision will tell you more about xAI’s future positioning than any Grok release note this year.
Reported by Martin Alderson at martinalderson.com, published June 8, 2026.
