The idea sounds great in theory. SpaceX wants to launch a new network of 1 million satellites to tap into the sun’s power and the atmosphere’s naturally cool environment to enable rapid growth of AI compute power. If all goes according to plan, SpaceX said it’ll be able to deploy up to 100 gigawatts of new compute each year without straining the terrestrial energy grid.
But this plan is fool’s gold.
There are three huge problems with SpaceX’s plan: space junk, compute power and performance.
Cluttering the night sky
SpaceX’s plan calls for deploying up to 1 million satellites in various orbits to perform different functions for its data center network. Some satellites will be placed in high-altitude, “sun-synchronous” orbits to capture the solar power needed to serve “energy-intensive AI workloads.” Other satellites in lower orbits will help load-balance compute across the system. Redundant orbital paths across these levels will help ensure reliability, SpaceX said.
“These satellites will have continuous access to nearly unlimited solar power to meet energy demands without dependency on terrestrial grids, enabling scalable, reliable and sustainable AI growth,” SpaceX wrote.
To understand just how outlandish a constellation of this size is, it helps to know that today there are an estimated 10,000-15,000 active satellites in orbit today. In total. And many of those are Starlink satellites.
What SpaceX is now proposing for data centers is around 100 times that number – for a single constellation. Add in SpaceX’s Starlink expansion, the constellation Blue Origin has planned, Amazon Leo’s forthcoming network, China’s Thousand Sails project and other planned “megaconstellations,” and well, you can see how things could get very crowded — and dangerous — very quickly.
According to NASA data, there are more than 25,000 large pieces of debris and more than 100 million tiny fragments in orbit. These have primarily come from space explosions and collisions.
Adding more orbiting junk won’t help. So, for all SpaceX wants to claim its ultra-dense constellation will address reliability with redundancy, more space debris actually makes operations more risky, not less so.
Then there’s the risk of solar flares and other space storms, which have the potential to fry any electronics in orbit caught unawares. SpaceX, of all companies, should know this. It lost 40 newly launched Starlink satellites to a geomagnetic storm in February 2022.
It’s also worth noting that solar storms heat the atmosphere and increase drag, which can alter satellite orbits.
SpaceX said the proposed constellation would use its existing automated collision avoidance system and “agile and highly-reliable electric propulsion systems” for precise maneuvering.
But still, in the hypothetical world where there are millions of satellites – data center and otherwise – managing them all becomes a high-tech, high-stakes game of dodgeball. I highly doubt folks running business critical AI workloads will want to take that risk.
Compute questions
Then there’s the question of how much compute power each satellite would be able to deliver and – more importantly – how long that compute would be relevant for AI workloads.
SpaceX’s proposal calls for 100 kilowatts of compute power per tonne of satellite. Great. That’s 0.1 megawatt. SpaceX’s Starlink satellites are a little over 1 tonne each. Even if you double that to 2 tonnes, you’d still need roughly 500 data center satellites to deliver 100 megawatts of compute power. In a world where data centers are increasingly measured in gigawatts, it’s hard to square how this makes sense.
There is, of course, something to be said for the idea of distributed compute capability to enable a new generation of AI applications. But that’s where you start to run into additional questions around what kind of performance these satellites will be able to deliver, and how quickly the chips they contain will become outdated.
Sure, you could do some software optimization of older chips (as Nvidia has been keen on doing). But there’s only so much software can do, and software updates won’t be able to counteract any irreparable damage caused to the chips by the aforementioned storms.
Performance
Network performance is another issue.
When most people think of satellite connectivity today, they think of Starlink. Ookla data from December 2025 showed Starlink offered download speeds of 129.72 Mbps, upload speeds of 22.59 and latency of 39 milliseconds.
But Ookla Lead Industry Analyst Mike Dano told Fierce those broadband numbers aren’t a great proxy for what SpaceX is proposing. “The data centers could rely on Starlink's global laser mesh network, which currently supports the company's Community Gateway sites,” he said.
Indeed, that’s exactly what SpaceX said it plans to do: link its new constellation up with Starlink’s “petabit” laser mesh to transmit traffic to earth. Each Starlink satellite holds three lasers, each transmitting at up to 200 Gbps, according to the company’s website.
Speed, though, isn’t the only variable that matters. Packet loss can be a make-or-break factor, especially for AI workloads. Not only can packet loss stall or require the restart of training workloads, it can also reduce the accuracy of inferencing workloads.
Turns out that compared to fiber, satellite is considered a much more “lossy” technology, raising the question whether the proposed constellation is really the best solution for the technology it is purporting to be built for.
One other thing
One other item in the proposal that makes no sense to me: SpaceX tried to claim the proposed constellation is an environmentally friendly alternative to terrestrial data centers.
“These satellites will have continuous access to nearly unlimited solar power to meet energy demands without dependency on terrestrial grids, enabling scalable, reliable and sustainable AI growth,” the proposal reads.
While I’m not thrilled with where terrestrial data centers are headed in terms of resource consumption and siting, I’d argue that the atmosphere is also part of our environment.
And I’d much rather look up at the stars than data centers. Keep your mess to yourself, Elon.
Chill out, dude
Yes, doing hard things is good. We should absolutely strive to be better and make ambitious plans.
Could SpaceX's team of talented engineers solve some or all of the above-mentioned challenges? Perhaps. And space-based data centers could indeed have useful applications.
But when tackling a hard task, you don't start by launching a constellation of 1 million devices. Especially not for a technology like AI that is advancing and evolving at such a rapid clip. That’s just hubris.
Op-eds from industry experts, analysts or our editorial staff are opinion pieces that do not represent the opinions of Fierce Network.