Chips in Orbit. Now What?
In February we said everyone was racing to launch orbital data centers and nobody was talking about the software. Eight months later, here is the scorecard.
On October 1 a Falcon 9 left Vandenberg on the Transporter-18 rideshare with 130 payloads on board. Three of them matter for this post.
Google’s first Project Suncatcher prototype, built with Planet, carrying four TPUs and drawing about a kilowatt. Google said on Monday that it has made contact and the satellite is working as expected, and that the coming weeks are about seeing how the chips handle launch stress, radiation and thermal extremes.
The first Orbitworks satellite, the opening member of the fifty-satellite constellation announced in Paris last month, with two NVIDIA Jetson AGX Orin modules and around twenty models on board, including a couple of Mistral language models for summarising what the sensor saw. Orbitworks also said on Monday that customers will pick applications from an app store and run them on shared satellites.
And TakeMe2Space’s MOI-1A, a sub-50 kilogram pathfinder with Orin NX processors, a second attempt after an earlier mission failed.
In the same fortnight SpaceX described a Starlink generation with on-board compute, pending the regulatory approvals that would let it fly. Eight months ago we wrote that the hardware was going up and nobody was talking about the software. Time to check the score.
The scorecard
| What we said in February | Where it stands in October |
|---|---|
| Starcloud had an H100 in orbit | Still flying. The follow-on with Blackwell is slated for this quarter and has not launched as we write. |
| Google’s Suncatcher was targeting 2027 | A four-TPU prototype has been in orbit since October 1, with contact confirmed. |
| Axiom had launched two nodes | Operational. Named by NVIDIA as a partner in March. |
| SpaceX had filed for a million satellites | Still pending at the FCC, now with an environmental petition against it and a Starlink-with-compute generation announced alongside. |
| China had a 200,000-satellite plan | Supercomputing-1 launched in September with an AI computer next to a camera. |
| Nobody was talking about the software | NVIDIA’s own space computing announcement lists three hardware platforms and no orchestration or scheduling layer. Orbitworks has an app store. The scheduler behind it is not described. |
Every row in the left column has moved. The hardware layer is not a prediction any more. It is crowded, it is funded, and some of it is talking to the ground.
The bottom row is the one we care about, and it has barely moved.
Four chips in a box is a test. A product is a schedule.
Here is what Suncatcher is, in Google’s own framing: a test of whether TPUs survive launch and then compute reliably in vacuum, cold and radiation. That is exactly the right first question and we are glad someone with Google’s resources is asking it carefully.
But notice what happens the day the answer is yes. The next question is not about the chip. It is: which job, in which window, on which satellite, with what power, and what comes down on which pass. The moment a satellite with compute on it has two customers, or one customer with two jobs, that question has to be answered by something, and the something is software.
Orbitworks has already arrived at that moment on paper. An app store with applications from several providers running on shared spacecraft is a multi-tenant scheduling problem by definition. We wrote about what that looks like last month. What we have not seen from anyone is a description of how those satellites decide whose job runs when, and whether the answer to a job that does not fit is a refusal up front or a missed deadline later.
NVIDIA’s March announcement is the clearest illustration. It is a thorough piece of work. Space-1 for the heavy end, IGX Thor in the middle, Jetson Orin for on-board processing, named partners for each. It describes CUDA as the way workloads run in orbit. It says nothing about how work gets placed, timed, or coordinated across nodes, because that is not NVIDIA’s layer. It is nobody’s layer, as far as the announcements go.
Where we are
We have been building that layer for a year, and we would rather be precise than grand about it.
What runs today: a planner that takes a satellite and a workload and returns a feasible plan on a timeline, or an explicit refusal with the binding constraint named. An agent that runs on the satellite and executes that plan. A CLI that lets you bring your own program. A paper that settles, at least for us, how to survive the gaps.
What does not run yet: fleet-scale orchestration at the size of the constellations being announced, and hosted multi-tenant scheduling with the operator, the customer and the spacecraft all reading the same plan. We have pieces. We do not have the whole thing, and anyone who tells you they do should be asked to show the schedule.
Now what
The February post ended with a question: once you get thousands of compute nodes into orbit, how do you actually run them? Eight months on, the nodes are arriving. Four TPUs here, two Orins there, a Blackwell soon. The question has not been answered. It has just become urgent.
That is fine by us. It is the question we have been working on all along.