The Most Insane Partnership In History

Picture a brand new building packed with racks of AI chips, ready to go. Nothing turns on. The wires coming into the building can't feed it.

Every big AI company is hitting that wall right now. One voice on the All-In podcast, Chamath Palihapitiya, put it simply: "I think power will be the moat."

Nvidia's Jensen Huang went on the BG2 podcast to talk about working with Elon Musk. "One particular new market that we're going to go explore together is putting data centers out in space," he said. Their two engineering teams are working closely to create "a new type of server, new type of architecture."

On October 15, the day that episode came out, Musk posted that the Starmind computer being co-engineered by SpaceX and Nvidia is far better than almost anyone expects and will reach volume production next year.

Today, the chips are getting built faster than anyone can find the electricity to run them. The man who sells the most AI chips on Earth and the man who launches the most rockets both think the answer is in orbit. I covered the partnership itself in No One Sees What SpaceX and Nvidia Are Building. This one is about the physics.

Out of everything

On the same All-In episode, the investor Ben Horowitz shared a conversation with Andy Jassy, who runs Amazon, about how much computing Amazon has left. Horowitz told him "we're out of everything" till 2028. Jassy's reply: "Oh, it's beyond 2028, Ben."

Chips can show up in months. Every new power plant in America has to get in line to connect to the grid. Lawrence Berkeley National Lab tracks that line, and at the end of last year it held about 2,000 gigawatts of projects. Of all the power that asked to join between 2000 and 2020, only about 1 gigawatt in 8 ever got built.

Even when the power exists, getting it into a building takes giant transformers, the big gray machines that step high voltage down. The wait for a large one is now two to three years. Berkeley Lab also found data centers used a little over 4% of America's electricity in 2023, and it expects roughly 7 to 12% by 2028.

Musk said at Davos in January that "the lowest cost place to put AI will be space," and that it would be true "within 2 years. Maybe three. Three at the latest." If the slow part on Earth is everything between the power plant and the chip, go where the sun is the power plant and there's no line.

Every watt becomes heat

All the electricity a data center pulls in ends up as heat. Leave a laptop on a pillow and you can feel it. Every watt that goes into a chip comes back out as a watt of heat.

One rack of today's Nvidia AI chips, a cabinet about the size of a fridge, pulls around 120 to 140 kilowatts. An average American home draws about 1.2 kilowatts, averaged around the clock. So one fridge-size box uses as much power as about 100 homes.

Cooling takes fans, liquid running through metal plates on the chips, chillers, and cooling towers that let water evaporate the way sweat cools your skin. It's one item on a long list. On BG2, Jensen put it this way: "When you move from digital to the physical, there are a lot of obstacles." Land, permitting, power generation. Each has its own waiting line, and the lines stack.

On the ground, the computer is the easy part. Getting power in and heat out is the hard part, and both take years to line up.

A panel that never sees night

Take power above the air and it gets a lot easier. Up there, sunlight hits a panel with about 1,361 watts per square meter. On the ground, a perfectly clear noon gives you about 1,000. Count nights and cloudy days, and an American solar farm averages only about a quarter of that peak. So the same panel in the right orbit collects roughly five times the energy, the same figure Musk uses.

That orbit is called a dawn-dusk sun-synchronous orbit. It runs over the poles and follows the line between day and night, called the terminator. A satellite riding that line keeps its panels in sunlight almost all the time, apart from short stretches in some seasons.

On CNBC in May, Jensen said "Nvidia is already in space." He sees a lot of technical challenges but said "there's no laws of physics limitation." And he called the energy available in space practically infinite.

Space is a vacuum flask

Power is the easy half. Heat is where most people get space wrong, because they picture it as freezing. Think of a vacuum flask instead, the kind that keeps coffee hot all afternoon. Between its two walls is a layer of nothing, and that gap keeps the heat in. Space is that layer of nothing, and it goes on forever.

Jensen named the problem at Nvidia's GTC keynote in March: "in space, there's no conduction, there's no convection, there's just radiation."

That means glowing, the way a stovetop coil glows red. The rule for how much heat glows away is called the Stefan-Boltzmann law, and it climbs fast. Measured from absolute zero, double an object's temperature and it glows away 16 times as much heat. So space radiators run hot on purpose, and they have to be big.

When SpaceX revealed its first AI satellite in June, Musk's assumptions were 250 watts per square meter for the solar array and about 1,400 for the radiators, which are double-sided and oriented "knife-edge to the sun." He called 1,400 "a very achievable goal."

Run that radiator number back through the physics and it works out to a panel at about 69 degrees Celsius, roughly a cup of coffee you'd still blow on. Edge-on is like holding a sheet of paper sideways to a flashlight, so the sun isn't heating the very thing you're using to cool down.

Bit flips and no repair crew

With no air overhead and less of Earth's magnetic shield, fast particles hit the chips and once in a while flip a bit, turning a one into a zero. Google tested its AI chips, called TPUs, with radiation. The memory only started acting up at almost three times the dose Google expects over five years in orbit. Google says the rare errors are fine for answering questions and a bigger worry for months of training. Distance is minor: light reaches a low orbit in about 2 milliseconds.

The harder problems are the ones physics can't fix. There's no repair crew in orbit, so these machines are built to fail in place. A dead chip gets switched off, and after about five years the whole satellite gets replaced. The dawn-dusk path is already a popular lane, so thousands of big satellites would need serious traffic control. Still, nothing in physics says no.

A gigawatt is a lot of anything

What's left is size. At 250 watts per square meter of solar panel, SpaceX's own figure, the math works out to about 4 square kilometers of panels for 1 gigawatt of computing in orbit. That's roughly 750 football fields, plus about seven tenths of a square kilometer of radiator. Close to 6,000 satellites of SpaceX's current design.

SpaceX lists that design at about 75 kilowatts of computing per metric ton, and its filing with the FCC aims for 100. So 1 gigawatt weighs 10,000 to 13,000 metric tons. Starship is designed to carry about 100 tons per flight, which makes roughly 100 to 130 launches per gigawatt. The same filing sketches a million metric tons a year adding 100 gigawatts every year. That's about 10,000 Starship flights a year, roughly 27 a day.

The price of the ride

Old throwaway rockets cost roughly $10,000 to $20,000 per kilogram to orbit. SpaceX's reusable Falcon 9 brought that down to around $3,600. Google's researchers think fully reusable rockets like Starship could get it to about $200 by the mid-2030s.

Jensen says Nvidia's share of a 1 gigawatt data center is now about $40 billion. Launching 10,000 tons at $200 a kilogram costs about $2 billion. Asked by Brad Gerstner about real computing in orbit within a few years, Jensen said the return on Nvidia's compute is so good "that you could put it on space and the launch cost is a small percentage of it." Then the catch: "The launch cost is not inconsequential." His conclusion: "You're going to put high-value computing out there."

The price per kilogram decides everything. At today's Falcon 9 price, launching that gigawatt would cost about $36 billion, almost as much as the chips. At $200 a kilogram, it's about a 20th of the bill. Google's researchers figure getting there takes about 1,800 Starship launches over 10 years. Starship has never flown more than five times in a year. The money only works if rockets fly like airliners and satellites come off a line like cars.

The first pieces

On June 8, Musk and Ian Dahl, SpaceX's director of satellite engineering, showed a satellite called AI1. It was first designed to run about 150 kilowatts of computing at peak and 120 on average, about one of today's AI racks. SpaceX has since made it bigger, and its newer spec sheet lists up to 250 kilowatts at peak and 175 on average, with a radiator of 160 square meters. 160 times 1,400 is about 220 kilowatts, right between average and peak. The physics checks out.

On August 24, Nvidia said SpaceX's planned first-generation Starmind satellite will be built around an optimized version of Vera Rubin, its newest AI server rack. In March, Nvidia also announced a space version of Vera Rubin. One of Nvidia's older Hopper chips, an H100, already flies on a satellite from a startup called Starcloud. And on October 1, Google launched a prototype TPU satellite on a SpaceX rocket.

The volume production in Musk's post is for the computer itself. Satellites full of them come after. His target for the first of those satellites to launch is late 2027, with significant scale in 2028, and no launch schedule has been published. SpaceX is building a gigasat factory in Bastrop, Texas, meant to make everything from the silicon wafers for its solar cells to the finished AI satellites, and it says thousands could start coming out as soon as late 2027. Musk's target is a gigawatt of orbital computing a year by then, and after that, aspirationally, 10 times more every year.

Starlink satellites already pass data to each other with more than 23,000 lasers, and SpaceX says its AI satellites will send results home through Starlink. On September 28, on flight 14, Starship reached orbit for the first time and dropped off 26 Starlink V3 satellites before coming home early after losing an engine.

Back to the wires

Every piece exists in some early form, and none of it exists at scale yet. The first satellites will probably have more downtime than a data center on the ground, since nobody can swap a broken GPU up there. The whole thing could take a decade or more. On Nvidia's February earnings call, even Jensen said the economics are poor today, though he expects them to get better. On All-In in March, he said it'll take years. On the other side is how fast SpaceX iterates, and the fact that no law of physics stands in the way.

On the ground, every new gigawatt waits in line for power plants, transformers, permits, and water. In the right orbit, the sun is almost always up, and there's no grid line to wait in.

My read is that Starship works like the railroad did, or the shipping container. Once moving stuff gets cheap enough, the question stops being whether you can afford it and becomes how fast you can build and launch.

Go back to that building full of chips, waiting on its wires. My bet is that once this works, AI growth ends up limited by how many satellites go up, more than by how much power Earth can spare. If that's right, the wall stops being the wires coming into the building.