SpaceX just posted the biggest IPO in history - roughly $85 billion with the over-allotment - and the product carrying that story is a satellite that has not flown: AI-1, a 70-meter-wide solar structure feeding one full AI rack in orbit. The pitch is free sunlight and no grid queue. The thing that almost kills it is heat. In vacuum you cannot blow air across chips. You can only radiate infrared, and at room temperature the math fails.

Key Takeaways

• SpaceX raised about $85 billion in the largest IPO on record, then hung a huge slice of the narrative on AI-1, a first-generation orbital AI compute satellite still short of flight.
• Peak payload power sits near 150 kilowatts (about 120 kilowatts average) - roughly one NVIDIA GB300 rack: 72 GPUs in a liquid-cooled cabinet that draws around 140 kilowatts on the ground.
• Fully deployed, the bird spans about 70 meters - wider than a 747-8 - and carries up to 110 square meters of deployable liquid radiators with redundant loops and micrometeoroid shielding.
• At \~20°C a two-sided radiator sheds only about 633 watts per square meter; 110 m² dumps \~70 kilowatts while average compute is 120 - so the system only lives if radiators run hot, because radiated power scales with temperature to the fourth power.
• Google is already writing checks on the order of $920 million a month for SpaceX/xAI ground compute (through mid-2029); Anthropic is even higher at about $1.25 billion a month - so the lease-to-compute demand is real today, not a fantasy customer list.
• Independent cost stacks still put orbital compute at several times ground cost today, and Morningstar puts only about a 7% chance on the most bullish moonshot case against the IPO tape.
• Prototypes are targeted for early 2027, full AI-1 units late 2027, with volume production on Starship around 2028 and GigaSat-scale manufacturing in Bastrop.

A data center is a warehouse full of chips. The AI boom turned those warehouses into a fight over two scarce inputs: electricity and cooling. A modern training campus can pull 100 to 300 megawatts - a small city's worth of power in one building - while global data-center electricity use has been racing from roughly 415 terawatt-hours toward the 1,000 range. Grids are missing reliability targets. Interconnection queues stretch for years. That is the ground bottleneck AI-1 is trying to exit: continuous sunlight, no water bill, no neighbors arguing over a new substation.

In orbit the power story gets easier. The thermal story gets brutal. No air means no convection. The only path out for heat is radiation, and it is slow. Room-temperature panels cannot reject the average 120-kilowatt compute load with 110 square meters of radiator area. You have to push the liquid loops hot - 50, 60, 80°C or more - so Stefan-Boltzmann kicks in and rejection multiplies instead of crawling. Independent napkin work on a \~100-kilowatt class bird lands radiator area in a range AI-1's 110 m² actually covers. Physics can work. The open question is factory and cost.

I know how this sounds when you stack the bear case. Orbital compute still looks several times more expensive than ground, sometimes worse once you price short satellite life, relaunch every chip generation, and ground stations you still need. AI silicon refreshes every year or two. A dead GPU 500 kilometers up is dead mass. That is the real kill shot - not "can SpaceX build a big satellite," but "can Starship drive dollars-per-kilogram low enough that disposable, mass-produced birds beat warehouses." SpaceX already cut access costs hard with Falcon 9 reusability. The public-company bet is that Starship and a car-plant satellite line bend the curve again while the same lease-to-compute customers already paying tens of billions on the ground become the demand for sky compute. If that works even partly, intelligence stops waiting on the next power plant. If it does not, you get a very expensive solar kite with a short-lived rack on board. We get to watch the launches either way.