The First Orbital Workloads Won't Be LLMs
SpaceX just described a 120 kilowatt satellite as a data center. A piece in The Conversation says it is a thousand times too small. Both are right, and both miss where the first real workloads are.
SpaceX priced its IPO last week, and the filing gave the rest of us something we had not had before: a number. The AI1 satellite, which SpaceX calls an orbital data center, is described as averaging around 120 kilowatts of compute. The first units are slated for late 2027, with the filing itself more cautious and Elon Musk telling people to take the date with a grain of salt. Gwynne Shotwell told CNBC that compute will go onto Starlink satellites before AI1 flies at all.
Yesterday The Conversation ran a piece arguing that AI1 is somewhere between a hundred and a thousand times less capable than a terrestrial data center, and that the earliest practical uses will be processing Earth observation data, defence and intelligence work, and science missions.
We have been building for this for a year and we would put it more bluntly. The first orbital workloads will not be language models. They will be small, boring, data-heavy jobs that live next to the sensor. And the reason is not the chip. It is the pipe.
The pipe
A modern imaging satellite collects far more data per orbit than it can send down. Our rough working number, and it is rough, is that a low Earth orbit satellite can generate on the order of a hundred times more data than it can downlink in the same period. The ground passes are short, a few minutes at a time, a handful of times a day, and the link budget is what it is. Most of what the sensor sees never reaches the ground.
That is the whole business case for compute in orbit, and it has nothing to do with training a model. If the satellite can look at an image and decide that it is mostly cloud, it can throw it away and spend the pass on the frame that matters. If it can run a detector and send down a bounding box instead of a scene, the pass carries a thousand detections instead of one picture.
Every kilowatt of orbital compute that gets paid for in the next few years will be paid for by that arithmetic. Not by an LLM.
Illustrative, not measured. The sensor produces far more than the link can carry. Compute on board does not make the pipe bigger. It changes what you put through it.
What those workloads actually look like
When we designed the planner we gave it a handful of built-in workload presets, and in hindsight the list is a pretty good description of what the first orbital customers want. None of them is a language model.
Onboard inference. A model already trained on the ground runs against fresh sensor data on the satellite. The output is small. This is the bread and butter, and it is what every Jetson flying today is doing.
Earth observation quality assurance. Cloud masks, blur checks, geolocation sanity. Decide on board whether a frame is worth the pass.
Orbital-aware imaging. Schedule the capture itself around the lighting and the ground track, so the sensor spends its time on scenes that will actually be usable.
Resilient store and forward. Collect, hold through the gaps, deliver when the ground is in view. Unglamorous and essential.
Split learning and federated learning. These are the two that sound like AI training, and they are, but notice the shape. The heavy part stays on the ground. The satellite contributes a small, local piece and sends up a small, local result. The pipe stays small.
What is common to all six is that the compute sits where the data is, the result is much smaller than the input, and the job can tolerate running in the windows the orbit allows rather than demanding a continuous machine.
So what are the LLM demos for
Starcloud trained a small language model in orbit last year and we thought it was a great piece of engineering. It proved that an H100 can launch, survive, stay cool enough, and run a real training loop in a vacuum. That is a hardware proof, and a necessary one.
It is not a market. Nobody is going to rent a 120 kilowatt satellite to train a model when a 120 megawatt building on the ground is available at a fraction of the cost per token. The Conversation piece is right about that, and so is every skeptic who has done the launch cost arithmetic.
The market is the Earth observation operator who is throwing away ninety-nine percent of what their sensor sees. The market is whoever is running a constellation where the most valuable thing a satellite can do with its spare watts is decide what not to send.
That market exists today. It is why the planner ships with those six presets and not a training preset. It is also why, when SpaceX says it will put compute on Starlink before AI1 flies, we read that not as a data center announcement but as a confirmation. The first orbital compute that matters will be small, close to the sensor, and scheduled around the pipe.