Software Can't Fix Radiators. It Can Fix Utilisation.
The thermal argument against orbital data centers got sharper this summer. We agree with most of it. Here is the one lever that is ours.
We wrote about thermal management in vacuum in January. The physics has not changed since. The numbers attached to it have got a lot sharper, and they have become the strongest argument the skeptics have.
In April a researcher at NASA’s Jet Propulsion Laboratory posted a preprint working through the spacecraft constraints on a one-megawatt orbital data center. The radiator area comes out around 2,500 square metres, with more than twice that in solar array. In June the World Economic Forum ran a piece built around the same problem, with the observation that running the chips hotter helps the radiators but costs you in leakage and error rates, so you cannot simply turn the dial. And Forethought published a careful piece on the economics whose conclusion is that the whole case depends on launch cost, with space solar reaching parity with off-grid terrestrial power at roughly 250 dollars a kilogram and orbital compute competing with a terrestrial buildout somewhere near 100.
We think all three are broadly right. We do not make radiators and nothing we write will change how much heat a square metre can reject into the dark. So this is a post about the part that is ours.
Two different fleets
The thermal argument is really an argument about one kind of satellite: a purpose-built compute platform in a dawn-dusk sun-synchronous orbit, sized in megawatts, that sees the sun continuously and is limited by how fast it can dump heat. For that satellite the levers are hardware and launch cost, and the software layer has very little to say. We would rather admit that than pretend otherwise.
But that is not the fleet most orbital compute will run on for the next several years. The near-term fleet is compute hosted on satellites that already exist or are already designed: Earth observation platforms, communications buses, hosted payloads. They are in ordinary orbits. They spend a third of each pass around the planet in shadow. They see a ground station for a few minutes at a time. Their power budget is tens or hundreds of watts, not a megawatt, and the thing that limits them is not the radiator. It is how much of each orbit the compute spends doing something worth paying for.
That is a utilisation problem, and utilisation is software.
The same hosted compute node over one orbit. Top: work runs whenever it arrives, so it idles through most of the sunlight and drains the battery in the dark. Bottom: the same work packed into the lit windows, downlinks placed in the passes, and the eclipse spent idle on purpose. Minute counts are illustrative; the ratio is not unusual.
What the planner actually does with this
The planner does not know anything about radiators. What it knows is that a satellite’s power follows the sun, that its thermal margin follows how hard and how long it has been running, and that its contact with the ground follows geometry. All three are computable ahead of time from the orbital elements. So it treats them as windows, and it places work into the windows rather than letting work land wherever the clock happens to be.
The effect on the node is the bottom row of the diagram. The sunlit portion of the orbit gets used. The passes carry downlinks rather than being discovered after the fact. The eclipse is spent idle, which is not waste, it is the one time the hardware should be doing nothing. And a job that would have run on battery because it arrived at the wrong moment instead runs twenty minutes later in the light.
None of that makes a megawatt satellite viable. It makes a hundred-watt one earn its keep. For the fleet that actually exists, that is the difference between hosted compute being a line item and being a demonstration.
Where that leaves the argument
The skeptics are asking whether it is worth launching the hardware, and the numbers they are using to answer it are good numbers. That is a question about kilograms and square metres and it will be settled by launch prices and radiator engineering, not by us.
We are asking a smaller question about the hardware that has already launched: is it busy. Most of it is not, and that is a problem with a software answer. We would rather be honest about which of the two questions we can help with than claim both.