SpaceX Bets Its IPO Future on AI-1, a 70-Meter Orbital Supercomputer

A 747-wide, chip-agnostic compute satellite targeting 150 kilowatts in orbit sits at the center of the roughly $85 billion float, and the economics only close if Starship drives launch costs toward $200 per kilogram.

SpaceX went public in June 2026 as the largest IPO in history, raising roughly $85 billion with the over-allotment, and immediately put the whole raise behind a machine almost nobody outside a few engineering rooms has fully priced: AI-1, a first-generation AI compute satellite wider than a 747 that has not flown yet. I have been sitting with this design for weeks. The pitch is brutal and simple. Earth is running out of cheap power and cooling for AI warehouses; space offers free sunlight and a vacuum that can reject heat if you build the radiators right. Wall Street’s more cautious desks give the bullish moonshot something like a 7% shot. That tension is the whole story.

Key Takeaways

• SpaceX’s June 2026 IPO raised about $85 billion with the over-allotment, the largest public debut on record, and AI-1 is the product that float is effectively financing.
• One AI-1 bird peaks at 150 kilowatts of compute and averages 120, roughly one NVIDIA GB300 rack of 72 GPUs bolted into orbit.
• Fully deployed, the satellite spans 70 meters across and about 20 meters tall, wider than a Boeing 747-8’s 68-meter wingspan, mostly solar wing and radiator.
• Up to 110 square meters of deployable liquid radiators with redundant loops sit behind the thermal budget; independent napkin math puts that area inside a physically sane band.
• Prototypes aim for early 2027, full AI-1 birds for late 2027, and volume production on Starship around 2028, with some compute riding regular Starlink V3 craft first.
• Ground AI deals already book tens of billions: Google near $920 million a month from October 2026 through June 2029 (about $30 billion and on the order of 1.1 million Nvidia GPUs), Anthropic about $1.25 billion a month through May 2029, plus another multi-year deal north of $6 billion.
• Orbital compute still looks roughly 3–4× ground cost on capital alone in mid-2026 analyses, and the harshest models need launch near $20–$30 per kilogram versus about $1,500 today.
• Starship’s target band near $185–$200 per kilogram is the same threshold big buyers have said orbital compute needs to clear.
• GigaSat in Bastrop, Texas covers more than 1,000 acres and over 11 million square feet, vertically integrated from solar ingots to finished birds so unit cost can ride a manufacturing curve down.

---

Earth’s AI Grid Is Already Short

A data center is a warehouse full of chips. Every search, stream, and model answer lives there, not on your phone. The AI boom turned those warehouses into the most contested real estate on the planet for two reasons only: power and cooling.

A modern AI campus can pull 100 to 300 megawatts in a single site. Global data-center electricity sat around 415 terawatt-hours in 2024 and is racing past 1,000. PJM, which balances power for 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. Interconnection queues stretch for years. Intelligence on the ground is bottlenecked on electrons and heat rejection before it is bottlenecked on clever software.

The AI-1 Spec Sheet, Without the Fog

AI-1 is a flying power plant for chips. Peak compute load is 150 kilowatts; average sits near 120. In satellite language that is absurd. A typical communications bird sips 10 to 20 kilowatts. SpaceX maps 150 kilowatts to roughly one liquid-cooled NVIDIA GB300 rack - 72 GPUs, about 140 kilowatts on the ground, millions of dollars in one cabinet - now in vacuum.

Solar generation targets 150 kilowatts at about 250 watts per square meter from cells SpaceX builds itself. Bandwidth reuses the Starlink V3 laser mesh; per-satellite rates are not published, but a terabit-class link is the right order of magnitude, thousands of times a typical U.S. home connection. Efficiency lands near 70 kilowatts of compute per ton. The payload bay is chip-agnostic. Nvidia, AMD, or future in-house silicon can slot in. That hedge matters when the best accelerator changes every year.

Bigger Than a Jet, Built Like a Car

Deployed, AI-1 is 70 meters wide and about 20 meters tall. The structure is mostly solar array and radiator, not a precious jewel. For sixty years satellites were gold-plated because launch cost $10,000 to $54,000 per kilogram. You radiation-hardened everything, triple-redundant every board, and designed for 15 to 20 years because you only got one shot.

AI-1 throws that rulebook out. It is large, relatively simple, meant for assembly-line volume, and designed to be replaced rather than repaired. When mass is cheap, you stop shaving grams and start sizing the factory. The claim from the program is that an AI bird is simpler than a full Starlink satellite - no giant consumer phased-array antennas, less of the RF complexity - and that much of the stack already exists on Starlink V3. Two prototypes target early 2027. Full AI-1 units aim for late 2027. Volume rides Starship around 2028. Some compute is supposed to hitch a ride on regular broadband and mobile Starlinks first.

Heat Is the Only Physics That Matters

Chips turn electricity into heat. On Earth you blow air or pump water. In vacuum there is no convection. The only path out is infrared radiation. A two-sided radiator near room temperature sheds about 633 watts per square meter - more than a thousand times slower than liquid cooling on the ground.

Napkin math on 110 square meters at that rate rejects about 70 kilowatts while the average compute load is 120. At room temperature the bird cooks itself. The fix is Stefan-Boltzmann: radiated power scales with absolute temperature to the fourth power. Run the liquid loops hot - 50, 60, 80°C or higher - and rejection roughly doubles or triples past the compute load while coolant keeps the chips themselves cool. Independent thermal checks for a ~100-kilowatt-class design land in a radiator area band that AI-1’s 110 square meters clears with margin. The physics can work. The question is price.

The Bear Case Is Ugly on Purpose

By mid-2026, putting the same compute in orbit still costs about four times the ground version on many full-cycle reads. Some stacks that bake in short satellite life, endless relaunch for new GPU generations, and residual ground clusters push the multiple far higher. One public calculator lands near $51 per watt in orbit versus about $16 on the ground - triple on capital alone before a single FLOP.

Obsolescence is the knife. AI silicon refreshes every one to two years. A chip 500 kilometers up cannot be swapped; it is dead mass until you launch again. The harshest credible launch math wants $20 to $30 per kilogram against roughly $1,500 today, with annual launch bills in the hundreds of billions if you try to keep orbital fleets current. Verdict from that side of the street: technically possible, economically non-viable on current cost curves. Research houses have called the idea peak hype. One major equity shop’s bullish moonshot gets a 7% probability, with fair value marked near $63 a share against a $135 IPO print - more than half the equity story treated as air tied to a space compute narrative they do not buy.

I know how that sounds. It should keep you skeptical. Building a big satellite is not the hard part for this company. Selling orbital FLOPs cheaper than Virginia is.

The Demand Side Is Already Writing Checks

The who-buys-this question is not hypothetical. SpaceX’s ground AI compute business already has the largest labs as customers. Google’s commitment runs near $920 million a month from October 2026 through June 2029 - about $30 billion total, on the order of 1.1 million Nvidia GPUs. Anthropic’s run rate sits near $1.25 billion a month through May 2029. A third deal adds more than $6 billion over four years. Direct competitors are writing ten-figure checks for the same leased intelligence.

AI-1 is that proven lease model lifted off the planet: sell compute to the richest AI labs, only with continuous solar and no grid interconnect queue. Demand is booked. Cost-to-orbit is the open variable.

Starship, GigaSat, and the Only Curve That Closes the Gap

The economic switch flips if launch collapses. Falcon 9 already cut the cost of reaching space by roughly 85% with reuse. Starship’s size and full reuse target a band near $185 to $200 per kilogram - the same neighborhood large buyers have said orbital compute needs. That curve is the bet under the bet.

Manufacturing has to match. GigaSat in Bastrop, Texas spreads over more than 1,000 acres and over 11 million square feet, integrated from solar ingots through finished satellites. Wright’s law does the rest: every doubling of cumulative output pulls unit cost a fixed percentage if the factory is real. Mass-produced, disposable, 70-meter compute birds only make sense when launch and build both behave like cars, not custom spacecraft.

Thinking Leaves the Rock

For the entire history of computation, every machine that has run a model has done it on the ground. Power plants, cooling towers, desert server farms, one rock. AI-1 proposes to move a slice of that work into continuous sun, with radiators as the dump and Starship as the truck. If it works even partially, the AI economy loosens its grip on terrestrial megawatts and the cost of intelligence keeps falling. If launch stays expensive or chips obsolete faster than relaunch can follow, it stays an expensive prototype class with pretty thermal slides.

I think the IPO is largely a public market wager on that stack - AI-1 and the generation after it - not on another incremental broadband bird. We get to watch the prototypes, the late-2027 units, and the Starship cadence in the open. The machine is either the product that justifies the largest IPO ever, or the clearest expensive moonshot on the board. The physics has a path. The spreadsheet only closes on a launch curve SpaceX has bent once and is trying to bend again.