Starmind: Elon’s Perhaps-Imaginary AI Megaconstellation Gets a Name

Well, it’s official. The proposed mega-constellation of AI data center satellites from SpaceX has a name: Starmind. That dovetails nicely with projects like Starlink and the Starship spacecraft, and it implies that real progress is being made toward the idea.

This is, unfortunately, nonsense.

The fact is that artificial intelligence companies (well beyond just Elon and xAI) have been reverse-engineering their approaches for a while now. First, the dream was directly profitable AI, but that was almost immediately shown to be impossible. So the solution had to be the same as before: centralization.

Thus, the dedicated AI data center was born, and eventually the gigawatt uber-AI data center, too. These weren’t dreamed up because they were directly proven to be good ideas; rather, they were arrived at out of necessity. It’s less that data centers could make AI profitable than that they had to.

But then they didn’t. Different companies are at different stages in accepting that ultra-large data centers just are not providing enough of an efficiency increase to make LLMs profitable at anything like a reasonable cost to the user. Since they have no other option to turn to, companies like OpenAI have to deny this fact as widely as possible, for as long as possible… but what about Elon Musk?

Tweet
(opens in a new tab)

Because of Musk’s association with the space industry and his enormous, preternaturally credulous fanbase, he was perfectly positioned to keep the game going. Data centers not working? Musk would simply put them in space.

Again, this idea was reverse-engineered from necessity, and the fact that space had to make AI profitable. At no point has it ever seemed as though it actually will.

And so, the game has continued in a Musk-only direction. Data centers in space make no sense? Well, then muddy the issue by talking about gigawatt data centers in space, as though that isn’t even more insane. Gigawatt installations in space aren’t realistic? Move the goalposts again by splitting the gigawatt data center into a gigawatt of data centers.

So, now we have the idea of a constellation of 10,000 data center satellites, each with 100 kilowatts of capacity, working together to form a gigawatt of networked compute. It quietly acknowledges what Musk directly denied until very recently: that generating a gigawatt of power and rejecting a gigawatt of heat are completely impossible in a single installation.

Here are some official, made-up numbers about the AI satellite they have the audacity to call “mini”..
Credit: SpaceX

So, as is tradition, Musk has moved on from such bald-faced falsehoods as though they never happened at all.

We could mention that 100 kilowatts is roughly the electrical and heat-rejection capacity of the International Space Station, which also happens to be by far the largest space project ever completed and a huge money sink for the countries involved. We could talk about how a globe-spanning network can’t all be in communication at once, and certainly can’t all be in sunlight at once. We could also talk about how, relative to terrestrial systems, a single gigawatt of capacity really isn’t a world-changing amount.

But let’s instead focus on the claims that are provably false.

Musk has said that space offers unlimited access to solar power, which is flatly wrong. In orbit, situated at a “sun-synchronous” position that allows the maximum number of hours of generation, a solar photovoltaic panel can generate 5-10 times more electricity than on Earth. That’s a great improvement and all, but many solar panels would still be needed to produce the required power.

A 100-kilowatt data center will spend about $1-2 million on power over 20 years. By comparison, even Musk’s rosiest projections for cost-to-orbit mean it will cost at least ten times that much just to launch 100 kilowatts’ worth of solar panels, all on their own.

Below, you can see one of the very few orbital options that could enable generation at 5-10 times terrestrial efficiency. Will SpaceX put all 10,000 satellites along this one ring?

Musk has also said that space-based data centers will negate the environmental impact of AI, but the reality is the opposite: Research shows that the environmental impacts of a satellite’s launch and eventual atmospheric deorbit are much greater than the operational impacts during its lifetime on Earth.

This is a more pressing concern than it might seem. The public sale of SpaceX is largely priced based on the alleged value of xAI as a component company, which is itself largely valued on Musk’s promises of space-based AI dominance.

The American and world economies are already so leveraged on AI that they fail to acknowledge the technology’s shortcomings, and now the single largest IPO in history has arisen from what is perhaps the least serious claim ever made about AI.

So, the next few years will be fascinating to watch play out. But one thing I can basically guarantee is that, whatever else does happen, Musk absolutely will not launch 10,000 International Space Stations into orbit, nor will he make a cent off of the launch of even a single one.

In all likelihood, he won’t even come close.