SpaceX Bets Public Markets on AI-1 Orbital Compute

A 70-meter AI satellite, an $85 billion IPO, and a thermal math problem that decides whether space compute ever undercuts Earth.

In June 2026 SpaceX went public in the largest IPO on record, roughly $85 billion with the overallotment, and almost immediately put the money next to a machine that has not flown: AI-1, a first-generation AI compute satellite wider than a 747. The pitch is simple enough to sound reckless. Earth is running out of cheap power and cooling for AI warehouses. Space has continuous sunlight, no interconnection queues, and no neighbors. Whether that is industrial history or an expensive dead end hangs on radiators that only work if they run hot, and on Starship driving launch costs into a band most of Wall Street still treats as fantasy.

Key Takeaways

• SpaceX raised about $85 billion in its 2026 IPO and tied the story to AI-1, a compute satellite that has not yet launched.
• A single AI-1 is rated at 150 kilowatts peak and 120 kilowatts average, roughly one NVIDIA GB300 rack with 72 GPUs in orbit.
• Deployed width hits 70 meters, wider than a Boeing 747-8’s ~68-meter span, with a ~20-meter unfolded height and about 110 square meters of liquid radiators.
• Global data-center electricity sat near 415 terawatt-hours in 2024 and is racing past 1,000; a modern AI campus can pull 100 to 300 megawatts alone.
• PJM, serving about 65 million people, came in roughly 6.5 gigawatts short of its reliability target in a recent capacity auction, the first miss of that kind.
• Google locked in about $920 million a month for SpaceX AI compute from October 2026 through June 2029 (~$30 billion, on the order of 1.1 million Nvidia GPUs). Anthropic’s larger monthly commitment runs through May 2029.
• Upfront orbital compute still looks roughly 3x ground capital in one public model (~$51/W vs ~$16/W), and full-cycle bear cases stretch toward 4–8x when short chip lives and relaunch are counted.
• Prototypes target early 2027; full AI-1 late 2027; volume production on Starship around 2028, after some compute rides early Starlink birds.
• Morningstar assigns roughly a 7% chance to the bullish moonshot path and marks fair value near $63 versus a ~$135 IPO reference, so more than half the print is space-AI optionality in that frame.

---

The Warehouse Problem AI Cannot Outrun

A data center is a building full of chips. Every search, stream, and model answer lives there, not on the phone. For years those warehouses were boring industrial real estate. The AI boom turned them into the most contested square footage on Earth, for two reasons only: power and cooling.

An AI training campus can demand 100 to 300 megawatts. Call 100 megawatts a small city in one building. Operators want dozens of those sites, then hundreds. Global data-center load near 415 terawatt-hours in 2024 is already climbing past 1,000, a real slice of U.S. electricity just to keep computers thinking. Grids are late. PJM’s miss on reliability capacity is the public proof. Interconnection queues stretch years. The bottleneck on intelligence is not algorithms. It is watts you can plug in this decade.

What AI-1 Actually Is

AI-1 is a satellite built to host a liquid-cooled AI rack in low Earth orbit. Peak generation about 150 kilowatts, average around 120. SpaceX’s own mapping is one NVIDIA GB300-class cabinet: 72 GPUs, on the order of 140 kilowatts on the ground, millions of dollars of silicon, now meant to fly.

The structure is mostly solar. Deployed, about 70 meters across and 20 meters tall. Cells SpaceX makes itself, around 250 watts per square meter, feed the payload. Inter-satellite laser links, reusing Starlink V3 heritage, point toward roughly a terabit per second class bandwidth between birds and down. No consumer phased-array antennas on the bus. Efficiency is quoted near 70 kilowatts of compute per ton, strong for spacecraft.

The payload bay is chip-agnostic. Nvidia, AMD, or in-house silicon later. That socket is a hedge in a market that rewrites the “best chip” every year. Two prototypes aim for early 2027. Full AI-1 birds late 2027. Volume production rides Starship around 2028, with some compute hitchhiking on ordinary Starlink broadband and mobile satellites first.

Simpler Than Starlink, Bigger Than Jewelry

For sixty years satellites were precious jewelry. Launch ran $10,000 to $54,000 per kilogram, so every gram got optimized. Few units. Tiny buses. Radiation-hardened parts. Triple redundancy. Fifteen-to-twenty-year design life because you never go up to fix them.

AI-1 throws that rulebook out. It is large, simple by spacecraft standards, meant for an assembly line, and treated as replaceable rather than repairable. Build it the way a car plant builds cars: obsess over the factory, not the gram. That only works if mass to orbit is cheap. Starship is the bet under the bet. When mass is cheap, you stop being clever and you make the solar and radiator area big.

SpaceX’s own framing at the reveal was blunt: AI-1 is simpler than a Starlink satellite. Starlink carries huge phased arrays, dishes, and a dense laser mesh. A lot of AI-1 hardware is already flying as Starlink V3 tech. Hard relative to everyday industry. Soft relative to what SpaceX already ships by the thousand.

The Thermal Trap That Almost Kills the Design

Chips turn electricity into heat. On Earth you blow air or pump water. In vacuum there is no convection. Heat leaves only as infrared radiation from hot panels. A two-sided radiator at about 20°C sheds roughly 633 watts per square meter. That is more than a thousand times slower than water cooling the same silicon on the ground.

Napkin math at room temperature is ugly. About 110 square meters of two-sided radiator at 633 W/m² rejects on the order of 70 kilowatts. Average compute sits near 120 kilowatts. Cool radiators, the bird cooks. The fix is Stefan–Boltzmann: radiated power scales with absolute temperature to the fourth power. Double absolute temperature and rejection jumps by 16x. Run the liquid loops at 50–80°C or hotter, dump heat into scorching radiators, and rejection climbs past the compute load while coolant keeps the chips themselves in range. The machine lives only if the radiators run hot.

Independent thermal sizing for a ~100 kilowatt class satellite needing ~60 kilowatts of rejection lands in the 41–71 square meter radiator band. AI-1’s 110 square meters sits inside a physically sane envelope. Building a big satellite is not the open question. Paying for it is.

The Bear Case With Numbers Attached

By mid-2026, putting the same compute in orbit still costs on the order of 4x ground in many frames. Full-cycle analyses that price short satellite life, relaunch every chip generation, and ground clusters you still need push toward 7–8x. One public calculator lands near $51 per watt orbital versus about $16 per watt on Earth before a single FLOP is sold.

Obsolescence is the quieter killer. AI silicon jumps hard every one to two years. A GPU 500 kilometers up cannot be swapped when a better die ships. You relaunch forever or accept a fleet of antiques. Harsh credible work puts the relaunch business near $20–30 per kilogram launch cost against roughly $1,500/kg class pricing today. Launch bills alone climb into the hundreds of billions a year at scale under those curves. Verdict from that corner of the street: technically possible, economically non-viable on current cost curves.

Servicing makes it worse. A dead GPU in orbit is dead mass. Skeptics in the lab and research world have called orbital GPU farms peak insanity or worse. Morningstar’s 7% bullish probability and ~$63 fair value against a ~$135 IPO print is the cleanest public expression of that doubt. More than half the equity story, in that model, is space-AI smoke.

I know how that stack reads. It is the correct default until launch costs fall hard.

Customers Already Paying on the Ground

Demand is not a slide deck. After folding xAI into the stack, SpaceX already runs a lease-to-compute business for frontier labs. Google’s agreement is on the order of $920 million a month from October 2026 through June 2029, about $30 billion total, sized around 1.1 million Nvidia GPUs. Anthropic’s larger monthly commitment runs through May 2029. A smaller third deal adds further multi-year dollars. Three anchors, two of them direct competitors, writing 10-figure checks for ground compute today.

AI-1 is that same product, lifted. Who buys orbital FLOPs? The labs already buying terrestrial FLOPs from the same seller, if price and latency ever clear. The open variable is cost to orbit, not demand theater.

Starship, Wright’s Law, and the GigaSat Bet

SpaceX has already cut the price of reaching space by about 85% with reusable Falcon 9. Starship’s target band near $185–$200 per kilogram (and lower over time) is the threshold Google-class analyses treat as the point where orbital compute can start to work. Wright’s law is the production story: each doubling of cumulative output drops unit cost by a fixed percentage. The GigaSat plant in Bastrop, Texas, more than 1,000 acres and over 11 million square feet, is built to ride that curve from solar ingots to finished birds.

I think going public in this window is less vanity than capital for the factory and the flight cadence that make disposable 70-meter compute satellites rational. Maybe 25–35% chance I’m wrong and the thermal/econ wall wins for a decade. The path that works still looks like the one SpaceX has walked before: force launch cost down until mass production stops being a joke.

Intelligence Leaves the Rock

For the entire history of computing, every machine that has done real thinking has done it on the ground. SpaceX’s public-market story is that the act of computation can move off the planet, into continuous sunlight, with radiators that only work hot and a launch curve that has to keep falling. Prototypes, Starlink rideshares, then Starship volumes will show, in public, whether that is industrial scaling or a beautiful thermal paperweight. The IPO is a bet that the factory and the rocket bend the cost curve before the chips in orbit go obsolete.