Orbital data centers

12 Sep, 2026

The startup Starcloud probably has the best demo so far. Per IEEE Spectrum, they already have a satellite with an Nvidia H100 GPU, launched in November 2025 via a SpaceX Falcon 9. CNBC reports that, using that GPU, they trained and ran nanoGPT on the works of Shakespeare, and also ran some version of Gemma (Google's open LLM)¹.

Orbital data centers' main advantage is energy generation. A satellite in the right orbit (a sun-synchronous orbit) can stay in the sun and generate energy continuously. A smaller, murkier advantage is less regulation: building a data center on Earth requires getting more permits and working with more governments, and may get harder given data centers' current unpopularity. There are still laws in space, just fewer.

They have many disadvantages, but none appear insurmountable.

The biggest, most discussed disadvantage is that cooling chips is much harder. This is a bit counterintuitive since space is cold, but that's because space is a fantastic insulator. Up there, (thermal) radiation is the only way to get rid of excess heat, and if you directly compare what chips need to what the International Space Station currently gets from its radiators, you'd conclude that you'd need to add prohibitively heavy radiators to any launch. Fortunately, the Stefan–Boltzmann law says energy is radiated as temperature⁴, and GPUs run hotter than the ISS, which already improves the rate "for free". Also, there hasn't been much pressure to design better radiators because there hasn't been much need for them, so there might be easy improvements from research or just recombining existing technology. Mesophase pitch-based carbon fibers are apparently both lightweight and great at conducting heat to a radiator, but I don't really know what I'm talking about here. Or Fu et al. 2019?

The other big concern is launch cost. You kind of just have to get good at launching stuff cheaply and frequently. Maybe by designing more damage-tolerant heat shields so your rockets survive reentry well enough to be reused; maybe by improving your launchpad so it can be reused faster. To some extent, this and the previous concern can compensate for each other: smaller/lighter cooling technology is easier to launch, more efficient launching can let you get away with bulkier/heavier cooling tech.

For better or worse, SpaceX is the clear frontrunner here, probably years ahead. SpaceX uses standard two-stage rockets (named "Super Heavy" and "Starship"), both of which are designed to be reusable; and they've already been able to catch and reuse the first stage, but not the second stage yet. In their most recent launch, Flight test 13 (July 2026), the second stage rocket splashed down in the Indian Ocean in a fairly controlled manner; it's sad that this is the only source I have for this, but Elon tweeted that the landing was precise enough that it would have been caught by the tower arms. Per Space.com, the next flight test (mid-September 2026) was previously planned to be the first launch where the second-stage rocket would be caught in an attempt to reuse it, but that's no longer the case.

Smaller concerns:

• The increased (particle) radiation in space can flip bits or damage chips. But adding error-correction and shielding aren't that expensive and heavy. Google tested their TPUs, things seem fine.
• Space debris could strike data centers and cause issues. Empirically, this hasn't really been an issue for Starlink and so seems manageable. With ambitious plans for how many orbital data centers to launch, it could get worse.
• Maintenance is basically impossible; the bet is that AI chips keep getting better so much faster than it's fine to just write off.
• Bandwidth will be low, but this seems acceptable if the data center is mostly used for running already-trained models (i.e., inference), rather than training models and then uploading/downloading their weights. Inputs and outputs for inference are pretty small, and inference is useful.

I only put in a little extra research beyond reading Forethought's report.