A Million Compute Satellites Is a Conjunction Problem
Environmental groups have asked the FCC to pause orbital data center applications. Whatever the agency decides, there is one effect of a million-satellite constellation that is simply arithmetic.
On July 8 a coalition represented by Earthjustice, acting for DarkSky International, Environment America and PEER, petitioned the FCC to stop processing orbital data center applications until the agency has done a programmatic environmental review. The filings in question add up to more than a million satellites. SpaceX alone asked for a million. Starcloud filed for 88,000. Blue Origin’s proposal is a little over 50,000. There are smaller ones behind those.
The petition lists rocket emissions, re-entry pollutants, ozone, orbital debris and the effect on astronomy. We are not going to litigate most of that here. We are not atmospheric chemists and this is not a policy blog.
But one item on that list is something we compute every day, and the scale being proposed changes it in a way worth spelling out.
The pairs
Conjunction screening is the job of working out which objects in orbit might come close enough to each other to worry about. The naive version checks every pair. With n objects that is n(n-1)/2 pairs, and we wrote last December about what that looks like at today’s scale: around fifty million pairs for ten thousand active satellites, over a billion once you include the trackable debris.
Now put a million compute satellites into the same catalogue.
Pairs to screen, n(n-1)/2, as the catalogue grows. Ten thousand satellites to a million is a factor of a hundred in objects and ten thousand in pairs.
Half a trillion pairs per screening window. Spatial filtering, which throws out pairs whose orbits can never touch, buys you two or three orders of magnitude. That is a lot, and it is not enough. And the thing we learned in December still applies: within a dense shell, where every satellite is in a near-identical orbit, filtering buys you almost nothing, because every satellite in the shell can in principle approach every other.
They all want the same orbit
Here is the part that makes it worse than the raw count suggests.
A compute satellite wants sunlight. All of it, all the time, because power is the business. The orbit that gives you that is a dawn-dusk sun-synchronous orbit, riding the terminator so the arrays never see the Earth’s shadow. Google has said that is where its constellation would go, at around 650 kilometres. It is the obvious choice and everyone with a power budget will make it.
Sun-synchronous orbit is already the busiest neighbourhood in low Earth orbit, because it is where Earth observation satellites go for consistent lighting. Put even a fraction of a million compute satellites into the same family of planes and you have built the densest shell that has ever existed, in the one place that was already crowded. Within that shell, the pairs problem is at its worst, and the conjunction rate per satellite goes up with the density of its neighbours.
We do not have a strong view on what the FCC should do with the petition. We do have a strong view that anyone planning capacity at this scale needs a screening system that scales with it, and that most of the ones in use today do not.
Why this is a compute problem and not just a safety one
A conjunction warning is not only a risk. It is downtime.
When a satellite has to manoeuvre to avoid something, the attitude changes, the arrays may come off the sun, the payload is usually paused, and the orbit afterwards is slightly different from the one your schedule assumed. For a communications satellite that is an inconvenience. For a compute node that is an availability gap, and in a dense shell those gaps get frequent.
That matters for the capacity question we wrote about in May. If you are selling compute from a fleet, your availability is not just eclipse and ground contact. It is eclipse, ground contact, and every manoeuvre the screening system tells you to make. In a crowded shell that third term is not small, and it is the one you cannot predict from the orbital elements alone. It depends on everyone else.
This is the thing we would most like the people filing for a million satellites to say out loud. Not whether the chips survive. What fraction of the orbit they expect to spend dodging each other, and what that does to the capacity they are selling.
Our own conjunctions endpoint screens a satellite against the live catalogue, using the time-partitioned approach we described in December. It was built for tens of thousands of objects. We are already thinking about what a million looks like, because it seems unwise to assume the filings will all be denied.