The sharpest technical objection to SpaceX's orbital AI plan is not power or heat. It is that modern AI models are designed to run across hundreds of interconnected chips... so we model how a typical 72 GPU NVDA rack would perform as a node in comparison. Forcing a frontier model into one satellite forgoes up to 44% of achievable output. The penalty is real but not permanent, better software, slimmer models, and models designed to fit the satellite could reduce it to near parity over time.

SpaceX and xAI are building one of the largest compute fleets on the planet. We estimate about 1.3 million chips by mid-2029. This supply-side breakdown sizes the build, splits what gets kept from what can be leased, and shows the hyperscaler call is xAI's choice to make.

SpaceXAi's S-1 names Terafab, its partners, and the terawatt compute ambition, but avoids the cost question. The press has filled the void with figures of $55B, $119B and, at the top end, $5T; none of which come from SpaceXAi. We built the bottom-up model the filing leaves out, and the bottlenecks we found look different from the ones the press keeps chasing.