Research Feed

AI companies have announced roughly 96 gigawatts of data center capacity. About 15 gigawatts run today. Mach33 tracks the gap facility by facility: ~240 sites and 65 supply deals, each project discounted by how likely it is to get built, against a price tape that turned in early 2026 and now sits ~40% above its October 2025 low. Six buyers have signed build programs that together exceed the entire installed base operating today.

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.

The radiation that is supposed to wall off orbital AI resolves into an operating condition: a 1,000-satellite reference fleet absorbs roughly 55,000 glitches a day and still posts 99.9984% availability. We built a physics engine, graded against 25 published flight and beam records, to price the damage, and 99% of the resulting tax is not the errors but the launch cost of carrying the shield. Radiation economics is launch-mass economics. The model behind the analysis is attached for members.