Investors reading SpaceX’s chip-rental deals and Meta’s reported plan to resell compute capacity as proof that AI’s supply crunch is ending are drawing the wrong conclusion. That is the position Jamin Ball takes in his July 3 Clouded Judgement newsletter, and the reasoning matters more than the headline numbers.

The raw figures do look dramatic. SpaceX, operating through xAI’s Colossus data centers, is reportedly generating about $2.32 billion in monthly revenue by renting roughly 450,000 GPUs among Anthropic, Google and Reflection. Anthropic alone pays $1.25 billion a month for exclusive use of Colossus 1, part of roughly 325,000 chips split across both facilities. Reflection, an open-source model developer, pays $150 million a month for GB300 chips. Meta is reportedly building a cloud business to sell its own AI compute, though Ball notes it has not actually sold any capacity yet.

Read together, those deals support a bear case: if two of the largest AI buyers no longer need their compute, hyperscaler capex forecasts should come down and demand may be softer than the buildout implies. Mark Zuckerberg said this week that agent capabilities are developing slower than expected, which feeds the same narrative.

Ball’s rebuttal rests on contract structure, not dollar totals. Every SpaceX deal includes a 90 day exit clause available to both buyer and seller. A party facing genuine oversupply does not need an exit ramp that short; it needs buyers to commit for years. A 90 day window looks like a seller planning to reclaim the hardware, not release it permanently.

The company-specific detail explains why. xAI lost most of its team and saw usage of its own models fall sharply, per Ball, leaving compute earmarked for training with no inference business to fund it. Citing Anthropic co-founder Dario Amodei’s remarks on the Dwarkesh podcast, Ball frames this as a ratio problem: a company that dedicates all its compute to training generates no revenue to offset the spend, while one that dedicates everything to inference starves its own model development. Renting out idle training capacity while its team rebuilds lets xAI collect cash without giving up on training altogether.

Meta’s situation reads similarly, even with fewer public deal terms. Llama has fallen behind rival open-source models, especially from Chinese labs, and turnover on Meta’s AI team has been widely reported. Selling capacity could mean Meta is monetizing compute built for a roadmap that slipped, or it could mean Meta wants a foothold in the cloud business regardless of how its own models perform. Ball says Meta’s actual plans are not yet public, so this half of the argument stays speculative by his own admission.

The strongest evidence against a glut is pricing, not the contracts. SpaceX charged rates steep enough that Anthropic, Google and Reflection paid them anyway, because none of the three could source equivalent capacity elsewhere. Premium pricing under a short-notice cancellation right is not what a market flooded with idle GPUs produces. Ball’s read: these deals reflect two AI laggards trying to squeeze revenue out of idle infrastructure, not a system-wide surplus.

For operators tracking capex, the distinction is not academic. Ball’s framework suggests hyperscaler capex guidance should hold steady through this earnings cycle even as resale headlines multiply. Watch instead for whether the next wave of compute contracts drops the 90 day cancellation window or the price per GPU. Either shift, not another rental deal, would mark the actual arrival of the scarcity end the market keeps prematurely pricing in.

Analysis drawn from Clouded Judgement, July 3, 2026.