Starship Flight 13, targeting July 16, carries its first operational payload, 20 Starlink V3 satellites, six of them fitted with camera suites that will scan the heat shield in flight, making the deployment double as a vehicle diagnostics pass. The Space Development Agency put $1.75 billion on contract for Tranche 3 of its Tracking Layer, roughly $955 million to L3Harris and $798 million to Sierra Space, 18 satellites each, feeding the Golden Dome architecture by end of 2028. China caught a Long March 10B first stage with a net-rigged recovery ship, a reusability milestone for its program. Rocket Lab ran a full-duration hot-fire of its Archimedes Vacuum engine ahead of Neutron's late-2026 debut.

If you haven't already, sign up for this week's webinar, back at our usual time of 11 AM EST after last week's one-off. [Sign up here.]

This week's analysis, "The Orbital Inference Containment Tax," prices the sharpest technical objection to SpaceXAI serving AI from orbit: GPUs computing one answer must sync within microseconds and so the team of chips ends at the satellite's hull, 72 GPUs. Forcing a frontier model into that box forgoes up to 44% of throughput, roughly $5 million per satellite-year at an illustrative $0.50 per million tokens. The tax is real and it decays: chip-specific software, the rewrite we modelled last issue, cuts it from 1.80x to 1.40x, and a model co-designed to fit the satellite takes it to 1.15x. The full analysis is public and the underlying model is exclusive to Mach33 Premium Research members.