The Hardware Layer Is Funded. The Systems Layer Is a Research Agenda.
On the same day last week, Starcloud raised a quarter of a billion dollars and a university group published telemetry from its own satellites. Read together, they say the same thing.
Last Friday, August 21, Starcloud announced a 250 million dollar extension to its Series A at a 2.3 billion dollar valuation, with NVIDIA and Cisco among the new investors. That takes the company to 450 million raised. The money goes to manufacturing, to engineering work with NVIDIA on the Space-1 Vera Rubin module, and to buying launch allocation, which several of the write-ups note is getting scarce. The stated ambition is 88,000 satellites and 20 gigawatts.
The same day, a group at Beijing University of Posts and Telecommunications posted a twenty-page paper on arXiv called “AI Infrastructure in Space: How Far Can We Go?” It is not a funding announcement. It is a systems paper, and unusually for this field it is grounded in telemetry from satellites the authors actually operate.
We read them back to back and they say the same thing from opposite ends.
What the paper found
Three findings, each from a different satellite.
From BUPT-1, the plain measurement: the usable compute on board is bounded by the thermal and energy envelopes, not by the processor. You cannot run the chip at the rate its datasheet says, because the spacecraft cannot carry the heat away or supply the power for long enough. The capacity you can sell is a function of orbital state.
From BUPT-2, a result on lifecycle. They built a system called SateLight for pushing application updates to a satellite and measured the transmission latency dropping by more than half on average, and by over ninety percent in the best case. The point being that deploying and updating software on a satellite is a real engineering problem with real numbers attached, not something you do once at integration.
And from a stateful vision-language serving workload, the one we underlined: thermal interruptions make execution-state recovery a first-class systems problem. The service gets paused by the spacecraft when the thermal margin runs out, and what happens to its state when that occurs is not an afterthought. It is the design.
Their conclusion is a research agenda. They define AI infrastructure in space as the systems layer that manages capabilities across spacecraft, orbital links, ground stations and cloud backends, while treating orbital and physical state as part of the resource model.
We would sign that definition. It is almost word for word the claim in the essay we published in April: that in orbit the resource constraints are a function of time, that the scheduler’s job is to know that function, and that this is a different kind of system from the ones the cloud taught us to build.
The two announcements, as a stack. The top layer has 450 million dollars behind it from one company alone. The middle layer has a paper. We are building the middle layer, and we are honest that the bottom row is where we are, not where the paper ends.
The thing nobody is funding
Here is what strikes us about seeing these two on the same day.
A company can raise 450 million dollars to put chips in orbit, and the hardest unsolved problem in making those chips useful is being written up as an open research agenda by a university group with two small satellites. The paper’s own case studies show it: capacity is a function of orbital state, deployment is a measured latency, and recovery across a thermal pause is a design problem. Every one of those is a software question. None of them is on the funding announcement.
That is not a criticism of Starcloud. Their job is the hardware and they are clearly good at it. It is an observation about where the money is and where the problem is, and the two are not in the same place.
Where we are, honestly
The planner exists and runs. Give it a satellite and a workload and it propagates the orbit, finds the eclipse and the passes, places the steps, and either returns a plan on a timeline or tells you which constraint it could not satisfy. The agent exists and runs on the satellite side. The CLI lets you bring your own program. That is the bottom row of the diagram, and it is real.
The middle row is bigger than that. Fleet-scale orchestration, hosted compute with several customers on one satellite, the lifecycle work the BUPT group measured. We have pieces of it, and we are not going to claim more than we have.
What we will claim is that the paper describes the right problem. Treat orbital state as part of the resource model. Make recovery across a pause a first-class concern rather than a crash. Our own result from July says recovery is free if the checkpoint lands before the pause, which means the valuable thing is knowing the pause is coming, which in orbit is a propagation problem the planner already solves.
The hardware is funded. The systems layer is a research agenda. We think it should be a product, and we are building it in that order.