The New Product SpaceX Is Betting Its Entire Company On: What the Market Is Missing

June 2026. SpaceX becomes a public company in the largest IPO ever run, roughly $85 billion raised with the over-allotment. Right before that raise, Elon Musk pulls the curtain on the product the money is chasing: AI 1, a first-generation AI compute satellite wider than a Boeing 747 that has not flown yet.

One respected Wall Street shop gives the bullish version of this story a 7% chance. So which is it? The most important manufactured object SpaceX has ever designed, or an expensive moonshot dressed up as a roadmap?

Two bottlenecks, one satellite

A data center is a warehouse full of chips. Every ChatGPT answer, every stream, every search lives in one of those buildings. For most of the internet era they were boring. Then the AI boom hit, and those warehouses became the most fought-over real estate on Earth for two reasons only: power and cooling.

A modern AI campus pulls 100 to 300 megawatts. One hundred megawatts is enough to run a small city, inside one fence. Companies are building dozens of these campuses, then hundreds. Global data center electricity use sat around 415 terawatt hours in 2024 and is screaming past 1,000. That is a meaningful slice of U.S. electricity demand, spent on computers.

The grid cannot keep up. PJM, the largest U.S. grid operator, covering about 65 million people, came up roughly 6.5 gigawatts short of its reliability target in a recent capacity auction — the first miss. New AI campuses sit in interconnection queues for years just to get permission to draw power.

And every chip has the same problem your laptop has: electricity turns into heat. Leave the heat in place and the chip cooks itself. On Earth you blow air or run water. A data center is basically a building-size cooling system with computers inside. Getting rid of heat is the whole game.

What AI 1 actually is

Musk filmed the reveal at SpaceX's Bastrop, Texas factory. The pitch is simple. On Earth, more power keeps getting harder and more expensive. In the right orbit, the sun never sets on your solar panels. No electricity bill. No water bill. No neighbors. No interconnect queue. Just sunlight in a vacuum.

The compute payload on one AI 1 is 150 kW at peak and about 120 kW on average. A normal communications satellite runs on maybe 10 to 20 kW. SpaceX's translation for 150 kW: roughly one Nvidia GB300 rack — 72 GPUs in a liquid-cooled cabinet drawing about 140 kW and costing millions. Picture the most powerful AI box on the market sitting in a Virginia warehouse, then picture that same box flying.

Physical size is where brains start to break. The deployed satellite is about 70 meters wide. The 747-8 spans roughly 68 meters. AI 1 is bigger, and about 20 meters tall fully unfolded. An airliner-scale structure, mostly solar panels, feeding one box of chips. The array targets about 150 kW of generation at roughly 250 watts per square meter, using cells SpaceX makes itself.

The quiet part of the spec sheet is thermal. Up to 110 square meters of deployable liquid radiators with redundant pumping loops, plus micrometeoroid shielding so debris does not punch a coolant line. Efficiency is rated around 70 kW of compute per ton. Birds talk to each other and back to Earth on laser links. Bandwidth is unpublished, but on Starlink V3 hardware you are looking at something on the order of a terabit per second. The payload bay is chip-agnostic — Nvidia, AMD, or in-house silicon later.

Simpler than Starlink, on purpose

The natural reaction: a 747-size robot that unfolds in space and runs a supercomputer has to be insanely hard. Musk downplays it. Standing next to satellite engineering director Ian Doll, he said AI 1 is actually much simpler than a Starlink satellite. Starlink carries giant phased arrays, parabolic antennas, and a heavy laser suite. AI 1 strips most of that out. A lot of the hardware already exists on Starlink V3.

Two prototypes are targeted for early 2027. Gwynne Shotwell said full AI 1 birds arrive late 2027, with some compute riding first on regular Starlink broadband and mobile satellites, and volume production on Starship around 2028.

That timeline only makes sense if you throw out sixty years of satellite design. From Sputnik to GPS, every satellite was built like jewelry because launch cost $10,000 to $54,000 per kilogram. You optimized grams, radiation-hardened everything, added triple redundancy, and designed for 15 to 20 years because you only launch once.

AI 1 is big, by satellite standards dumb and simple, meant for an assembly line, and basically disposable. Replace it instead of repairing it. When mass is cheap, you stop being clever and you make it bigger. That only works if Starship makes launch dirt cheap. The satellite bet sits on the launch bet.

The heat problem nobody markets

In a vacuum there is no air, so there is no convection. You cannot blow a fan across a hot chip. The only way to dump heat is to radiate it as infrared. A two-sided radiator at about 20°C sheds roughly 633 watts per square meter — over a thousand times slower than running water across the same chips on Earth.

Napkin math: 110 square meters, both sides, at 633 W/m² rejects about 70 kW. Average compute load is 120 kW. At room temperature the satellite sheds 70 and produces 120. It cooks itself.

Radiated heat scales with absolute temperature to the fourth power — Stefan-Boltzmann. Double the absolute temperature and you get sixteen times the heat dump. So you do not run the radiators cool. You run the liquid loops hot — 50, 60, 80°C or more. Rejection doubles or triples, you clear past 120 kW, and the coolant keeps pulling heat off the chips into those hot panels. Run the radiators cool and the machine dies. Run them hot and it lives.

Independent analyst Mark Barra ran separate numbers: a roughly 100 kW satellite shedding about 60 kW needs somewhere between 41 and 71 square meters of radiator. AI 1 carries 110. Thermally, they are inside a sane range. Building a big satellite was never the hard question. Economics is.

The bear case at full strength

As of mid-2026, putting compute in orbit costs about four times the same compute on the ground. Analyses that count short satellite life, relaunch every time a GPU generation flips, and the ground stations you still need push that multiple as high as 78 times. Engineer Andrew McCalip's public calculator lands roughly $51 per watt for orbital compute versus about $16 per watt on the ground — triple the upfront capital before a single calculation runs.

Obsolescence is the real killer. AI chips jump a generation every one to two years. A chip you launch today is already behind the moment a better one ships on Earth, and you cannot swap it at 500 km. You are not launching once. You are relaunching forever.

Barra's harshest credible version says the relaunch business needs launch costs around $20 to $30 per kilogram against today's roughly $1,500. His verdict: technically possible and economically non-viable at current cost curves. Sam Altman has said GPUs still break a lot and that will not matter at scale this decade. Gartner called the idea "peak insanity." Short seller Jim Chanos called it "AI snake oil." Morningstar gives the moonshot scenario a 7% chance and pegs fair value around $63 a share against a $135 IPO price — more than half the price tagged to a space story they do not buy. That skepticism is fair. Do not paper over it.

Why the demand side is not a guess

Who buys this? People are already buying the same product on the ground. SpaceX, after absorbing xAI, already runs a compute business for the biggest AI labs on the planet. Google agreed to pay SpaceX roughly $920 million per month from October 2026 through June 2029 — around $30 billion total, about 1.1 million Nvidia GPUs. Anthropic committed about $1.25 billion per month through May 2029. Add a smaller third deal and you have tens of billions a year in recurring revenue from anchor customers who are direct competitors writing 10-figure checks.

The business model is lease compute to the world's richest AI labs. That model is already proven on Earth. AI 1 is that same business lifted into orbit. Customers are not hypothetical. They are already paying. The open question is whether SpaceX can drive cost to orbit low enough that the orbital version stops being a vanity tax.

Starship being enormous and reusable is the mechanism. The curve SpaceX wants runs from roughly $1,500 per kilogram down toward about $185 to $200 per kilo and lower. Google itself has pointed at that band as the threshold where orbital compute starts to work. Falcon 9 already cut the cost of reaching space by about 85%. SpaceX is building the Bastrop satellite factory — over 1,000 acres, more than 11 million square feet, vertically integrated from solar ingots to finished birds — to ride production costs down the same way.

What is actually being bet

For the entire history of computing, every machine that has ever thought has done its thinking on the ground. SpaceX is proposing, with a straight face and the most valuable IPO ever behind it, to move the act of computation off the planet — into a sky where the sun never sets and heat leaves as light.

Prototypes in early 2027. Full birds late 2027. Volume on Starship around 2028. We get to watch whether it works in real time, with launches you can track and satellites you can see. The physics can close. The customer checks are already clearing. The stock price is mostly a bet that Starship bends the launch curve again before chip obsolescence and radiator reality eat the spreadsheet.

That is the product. That is the company. That is the fight.

Check the video here.