Houston, We Have a Data Center: AI Infrastructure Goes Orbital
For a while now, this series has stayed grounded, quite literally. We've talked CXL, SOCAMM, High Bandwidth Flash, ZAM, ESUN, all clever ways of tackling the AI world's appetite for memory bandwidth and interconnect, but all firmly bolted to a data center floor somewhere on planet Earth. That floor, it turns out, is running out of room.
Over the past few months, the AI infrastructure conversation has taken an unmistakable turn: upward. Elon Musk folded xAI into SpaceX, in a deal reportedly valuing the combined entity at $1.25 trillion, framing the move explicitly around the idea that Earth-based power and cooling can no longer keep pace with AI's compute appetite. Around the same time, SpaceX filed with the Federal Communications Commission (FCC) for approval to deploy a constellation of up to one million satellites designed to function as orbital data centers. Blue Origin, Google, and quite a few well-funded startups have all entered the fray with their own orbital compute ambitions.
When I first heard “data centers in space,” I dismissed the idea as something akin to flying cars. But the more I dug into this, the more it became clear that serious engineering teams, serious capital, and serious regulatory filings sit behind it. So let's unpack the emerging space race, pun very much intended, for AI infrastructure.
What's Driving This
It's the same old AI infrastructure trilemma showing up: power, cooling, and land.
Global electricity demand for AI is running well ahead of grid capacity in most markets outside China. In the US, data centers consumed about 4.4% of total electricity in 2023 per the Department of Energy; globally, the International Energy Agency (IEA) estimates data centers accounted for roughly 1.5% of electricity use in 2024, a share that's rising fast. PJM, the largest US grid operator, expects data-center-driven demand to grow by roughly 30 GW between 2025 and 2030.
Interconnection queues, meaning the wait to actually connect a new large load to the grid, can run five to fifteen years by IEA estimates, while a data center campus can be built in a fraction of that time.
Land, cooling water, and community pushback, in Northern Virginia, Ireland, and parts of the Netherlands among other places, are adding further friction to new terrestrial builds.
Space sidesteps all three at once, at least on paper. A satellite in a dawn-dusk sun-synchronous orbit sees the sun nearly continuously: no clouds, no night, no seasonal variation. That alone can make solar panels as productive as their terrestrial equivalents. There's no land to lease, no water to draw, no utility queue to wait in, and radiative cooling into the vacuum of space, its own engineering headache as we'll see, doesn't compete with anyone for a permit.
What’s Being Proposed
This isn't one project. It's now a crowded field, and most players are converging on a similar architecture: satellites carrying GPUs or custom AI accelerators, tied to each other and to the ground by optical inter-satellite links (ISLs), flying in sun-synchronous Low Earth Orbit (LEO).
SpaceX's Orbital Data Center System: filed with the FCC for up to one million satellites, in orbital shells roughly 50 km thick between 500 km and 2,000 km altitude, using optical ISLs and connecting to the ground via the Starlink laser mesh. Musk's own figures point to roughly 100 GW of AI compute capacity added per year if the plan comes to pass.
Blue Origin's Project Sunrise: up to 51,600 satellites in sun-synchronous orbits between 500 km and 1,800 km, riding on Blue Origin's planned TeraWave optical backbone (itself a 5,408-satellite network) and launched via New Glenn. Blue Origin is seeking Ka-band spectrum (18.8–19.3 GHz and 28.6–29.1 GHz) for telemetry and control.
Google's Project Suncatcher: a research-first approach, modeling 81-satellite clusters flying in tight, roughly 1 km-radius formations connected by free-space optical links already demonstrated at 1.6 Tbps on the bench. Google has radiation-tested its Trillium (v6e) TPU on a 67 MeV proton beam; the chip held up well, though the HBM subsystem proved the most sensitive component. Two prototype satellites, built with Planet, are due to launch by early 2027.
Cowboy Space's Stampede: a newer entrant (formerly Aetherflux, and founded by a Robinhood co-founder) proposing up to 20,000 satellites. Its pitch is refreshingly specific: this is less about space being inherently cheaper and more about skipping years-long grid interconnection queues altogether.
Starcloud: already flying an Nvidia H100-class GPU in orbit, reportedly the first to run an LLM, and a version of Gemini, in space, with a stated ambition of a multi-gigawatt orbital data center over time.
China's orbital compute push: reportedly targeting on the order of 1,000 satellites and roughly 1 GW of orbital compute capacity by the end of the decade, in sun-synchronous orbit around 700–800 km.
The USP: Why Orbit, Really?
The case for orbital AI compute rests on a handful of concrete claims:
Near-continuous solar power. No batteries needed for baseload; panels in the right orbit can run up to roughly 8x more productive than ground equivalents.
No land, water, or grid queue. The economics here aren't really about cheaper power; they're about skipping a multi-year wait to connect a large new load to an already strained grid.
Falling launch costs. Google's own modeling suggests that once launch costs to LEO fall to around $200/kg (current Falcon-class launches run roughly $1,500–$2,900/kg or more), the amortized cost of orbital compute becomes roughly comparable to terrestrial data center energy costs, which run an estimated $570–$3,000/kW per year, on a per-kW basis.
Optical mesh networking. Free-space laser links, some already demonstrated at 1.6 Tbps in lab tests, stand in for the racks-and-switches interconnect fabric this series usually covers (CXL, UALink, ESUN) with a physically separate satellite-to-satellite fabric instead.
A Note on Regulation and Spectrum
As with any new interconnect or memory standard we've covered here, none of this works without the regulatory plumbing. Every major orbital data center filing needs FCC sign-off, and that process has turned unusually adversarial. Amazon formally petitioned the FCC to deny SpaceX's million-satellite filing, arguing it would hand SpaceX outsized control of prime orbital shells. Blue Origin, also under Jeff Bezos alongside Amazon, then filed a broadly similar proposal of its own, days later for Project Sunrise. SpaceX, in turn, has asked the FCC to hold Amazon and Blue Origin to the same standard being applied to its own application. None of the three has yet published detailed satellite specifications or a firm deployment schedule, and orbital congestion, debris mitigation, and spectrum coordination remain open questions for regulators to work through.
Ecosystem and Competitive Landscape
Underneath the filings, there's real hardware progress. Starcloud's H100 has already flown and computed in orbit. Rendezvous Robotics is building self-assembling, tile-based satellite modules that could reduce reliance on large monolithic launches, and has already partnered with Starcloud. Google's radiation and optical testing is public and peer-reviewable, a refreshing contrast to some of the more marketing-forward filings in this space. And this isn't a US-only story: China's own orbital AI compute ambitions are advancing in parallel, with government-backed constellations targeting gigawatt-scale capacity by the decade's end.
Honest Caveats
Not everyone is convinced, and the scepticism deserves airtime.
OpenAI's Sam Altman has publicly called the current crop of orbital data center proposals impractical for this decade, pointing to the mismatch between launch costs and terrestrial power costs, and to the basic problem that GPUs fail regularly on Earth and simply cannot be swapped out in orbit. Industry analysts have raised a related concern: both geosynchronous orbit and popular LEO shells are already crowded, and solar power plus radiative cooling alone don't resolve that congestion problem.
The physics deserves a mention too. In the vacuum of space, only radiative cooling is available: no convection, no air handlers. A high-power orbital cluster of GPUs therefore needs a proportionally large radiator area, which adds mass, and mass is the one thing launch costs punish hardest. Radiation is the other open question. Google's own testing shows HBM subsystems are the most sensitive component to total ionizing dose, and the effect of single event effects on long, multi-week training runs remains an active research question rather than a solved one.
Outlook
None of this makes orbital AI compute a near-term substitute for the terrestrial buildout this series of articles has spent covering. But it does mark a genuine shift in how a subset of very well-capitalized players are framing their own bottleneck. For a while, the conversation was chips, then memory, then interconnect, then power. Now it also includes orbital shells, launch cadence, and optical inter-satellite bandwidth as legitimate infrastructure planning inputs, not science fiction.
My own bet is on the terrestrial grid getting fixed first. But I've been wrong before, and this is a space (sorry) worth watching closely.
What's your take: is orbital AI compute shaping up to be a serious infrastructure tier, or an elaborate way around a regulatory problem? Let me know in the comments; I'll keep tracking this as it develops.
