Table of Contents
Elon Musk just declared that the computing power behind future super-intelligence will be “overwhelmingly” in space — a share he says will “round up to 100%.” Days earlier, a Google satellite beamed down the first AI-generated poem ever computed in orbit. Put those two headlines together and data centers in space stop sounding like science fiction. They start sounding like a construction schedule.
This is no longer one company’s moonshot. Five companies have filed paperwork for a combined 1.26 million orbital compute satellites. Four separate programs already have AI chips running in orbit. And the reason is brutally practical: Earth’s power grids and water supplies are buckling under AI demand, while orbit offers near-constant sunlight and no neighbors to complain.
Here is the full picture of the race to move the cloud off the planet — who is flying, who is filing, and what physics still stands in the way.
Why Data Centers in Space Suddenly Make Sense
The case for data centers in space starts with a wall Earth is hitting. Goldman Sachs Research estimates global data center power demand will grow 175% by 2030 compared with 2023 — the equivalent of adding a top-10 economy’s power consumption to the grid in under a decade. In the United States alone, data centers could account for 11% of total electricity demand by 2030.
Water is the quieter crisis. AI data centers are projected to drive an 11-fold increase in annual water consumption for cooling and power generation by 2028, reaching roughly 1,068 billion litres. Data center development has already become a political flashpoint: community opposition and regulatory pressure delayed or blocked nearly $200 billion worth of projects in early 2026.
Meanwhile the money keeps pouring in. Microsoft, Alphabet, Amazon, Meta and Oracle spent over $556 billion on AI infrastructure in the first nine months of 2026 — pushing the group into a free cash flow deficit for the first time in data going back to 2011. The industry is desperate for anywhere cheap, abundant power exists.
Orbit sidesteps both problems at once. Satellites in low Earth orbit get near-constant sunlight — Google estimates up to eight times more solar energy than equivalent ground panels. And in the vacuum of space, waste heat radiates away without a single cooling tower or litre of water. That combination is why the idea went from joke to boardroom strategy in under two years.
Data Centers in Space: 5 Projects Racing to Orbit
The field splits into two camps: hardware already flying, and paper filings for hardware that might. DatacentersX’s tracker counts four programs with AI compute already in orbit. Here are the five names that matter.
1. SpaceX Starmind — the million-satellite filing
The biggest bet on data centers in space belongs to SpaceX. Its “Starmind” orbital data center system was filed with the FCC on January 30, 2026, requesting permission for up to one million satellites flying between 500 and 2,000 km. Each planned AI1 satellite would carry more than 250 kilowatts of solar power — roughly the compute of one Nvidia GB300 rack in orbit, built on Nvidia Rubin GPUs and Vera CPUs, with laser links joining clusters of about ten satellites.
The schedule is aggressive: prototypes in early 2027, orbital compute demonstrations in late 2027, and commercial service as early as 2028, according to SpaceX’s prospectus. The hardware would roll off the Gigasat factory in Bastrop, Texas — an 11-million-square-foot campus whose solar cell plant is designed for 10 GW a year. SpaceX’s target is 1 GW of orbital compute a year by the end of 2027, launched 30 to 50 satellites at a time aboard Starship.
Musk’s own forecast sets the tone. Replying to Sequoia partner Shaun Maguire’s financial model on October 30, he wrote that “terrestrial compute will be peanuts compared to orbital compute,” adding that monthly orbital deployment would “probably” exceed terrestrial additions around the end of 2028 — and “far exceed” them in 2029, as TeslaNorth reported.
2. Starcloud — the first GPU already flying
While SpaceX files paperwork, Starcloud has hardware in orbit. Its Starcloud-1 satellite has carried an Nvidia H100 — the first data center GPU in space — since November 2025, and has already trained a small AI model and run Google’s Gemma model onboard.
The startup has raised $450 million, most recently at a $2.3 billion valuation in August 2026, with Nvidia and Cisco among its investors. Its next satellite, Starcloud-2 — carrying what would be the largest commercial deployable radiator ever flown — has slipped to 2027. Starcloud has also filed for up to 88,000 satellites of its own.
3. Google Project Suncatcher — the poem from orbit
Google’s entry is the most science-forward. Its Suncatcher prototype, built with satellite company Planet, launched October 1, 2026 aboard a SpaceX Falcon 9, carrying four of Google’s Tensor Processing Units. With only about 1 kW of power, it can run just one TPU at a time for 15 to 20 minutes — but on October 30, engineers uplinked a prompt to the onboard Gemma 4 model and received a haiku back as the satellite passed over Svalbard, Norway, according to The AI Insider. We covered that milestone in detail here: Google Project Suncatcher: First AI Prompt Runs from Orbit.
Google is explicit that this is a feasibility test, not a working data center. Next up: radiation and thermal testing over a roughly one-year mission, then two interconnected satellites in 2027 with inter-satellite optical links. Google’s researchers put the economical horizon in the 2030s — estimating it would take about 1,800 Starship flights to drive the learning-curve cost reductions the economics need.
4. China’s Three-Body constellation — 12 satellites up
China got there first. ADA Space and Zhejiang Lab launched 12 satellites of the Three-Body Computing Constellation on a Long March 2D in May 2025, each rated at 744 TOPS with laser crosslinks between them. In January 2026, Alibaba’s Qwen3 model ran inference in orbit. The stated plan: 1,000 satellites by 2030 and 2,800 by 2035.
5. Axiom Space — quiet nodes on an optical network
The least flashy entry may be the most operational. Axiom Space’s orbital data center Nodes 1 and 2 launched in January 2026 aboard Kepler’s optical relay satellites, communicating over 2.5 Gbps optical links. Compute specs are undisclosed — but unlike the filings above, this hardware is already working.
The Paper Trail: 1.26 Million Satellites, Zero Approvals
Add up the FCC filings and the number is startling: five companies want permission for 1,259,600 data center satellites — roughly 75 to 80 times the number of working satellites in orbit today. The list: SpaceX (up to 1 million), Orbital Compute (100,000), Starcloud (88,000), Blue Origin’s Project Sunrise (51,600), and Cowboy Space’s Stampede (20,000) — see the full filing-by-filing breakdown.
Read the fine print, though. Every one of these has been “accepted for filing” — FCC-speak for “the paperwork was complete enough to open for public comment.” None has been approved. And all five asked the FCC to waive the milestone rule that normally requires half a constellation within six years; SpaceX, Starcloud and Orbital Compute also asked to skip the surety bond. The licenses are land grabs without deadlines — for now.
SpaceX’s own filing calls the project a first step toward a Kardashev II civilization. Musk’s take on orbital crowding — that space is “really big” so it won’t get crowded — is about to be tested against the math below.
Data Centers in Space Face Brutal Physics
The physics doesn’t care about pitch decks. In a vacuum, heat can only leave by radiation — roughly one square metre of two-sided radiator per kilowatt at 330 K. SpaceX’s AI1 design needs about 110 m² of liquid radiators for 120 kW of average compute. Radiators, not chips, are what size these satellites — and Google’s project lead has said the radiators are among the heaviest parts of its test satellite.
Launch cost is the second wall. At today’s roughly $1,500 per kilogram and about 100 kW of compute per tonne, launch alone costs around $15 per watt. If Starship reaches its $100-per-kg goal, that falls near $1 per watt — which is why every orbital data center business plan is really a Starship schedule in disguise. Nothing up there can be repaired or upgraded, either: a dead satellite is dead money.
Then there’s radiation. Data center GPUs were never radiation-hardened: cosmic rays flip bits silently, corrupting results without tripping alarms. A 2022 solar storm destroyed 38 Starlink satellites outright. Google’s ground tests were encouraging — its Trillium TPU memory tolerated nearly three times the expected five-year dose — but the orbital verdict is still out.
And orbit is getting crowded. Some 16,000 to 17,000 satellites are working up there now; researchers’ CRASH clock — the time to a likely collision if everything stopped dodging — reads 2.2 days, down from 164 days in 2018. ESA’s space environment health index jumped roughly tenfold in a single year, and astronomers put the sustainable limit near 100,000 faint satellites total. A million-satellite constellation doesn’t fit in that math, and one expert model suggests four of the five filed plans risk runaway debris on their own.
Timeline: When Data Centers in Space Become Real
So when does this stop being experiments and filings? The honest answer, from the companies themselves: the late 2020s for demonstrations, the 2030s for economics.
- 2027: SpaceX flies AI1 prototypes; Starcloud-2 launches with its giant radiator; Google launches two interconnected Suncatcher satellites with laser links; Aetherflux targets its first “Galactic Brain” node.
- Late 2027: SpaceX plans its first orbital compute demonstrations.
- 2028: commercial service “as early as 2028” per SpaceX’s prospectus — and Musk’s forecast that monthly orbital deployment passes terrestrial additions around year’s end.
- 2030: China’s constellation aims for 1,000 satellites; data center power demand hits +175% vs 2023.
- 2030s: Google’s target for economical orbital AI infrastructure — if launch costs fall far enough.
Data Centers in Space: Frequently Asked Questions
Are there data centers in space right now?
No full-scale facility exists yet — only single-satellite experiments. Starcloud’s H100 satellite, Google’s four-TPU prototype, China’s 12-satellite constellation and Axiom’s two nodes are all flying, but each is a test article, not a working data center. Everything bigger is still on paper.
Why put data centers in space instead of building on Earth?
The pitch for data centers in space is about Earth’s limits: power demand set to grow 175% by 2030, water use for cooling exploding, and $200 billion in projects already delayed by local opposition. Orbit offers up to eight times more solar energy, no water cooling, and no neighbors.
What is SpaceX Starmind?
SpaceX’s planned orbital compute constellation: AI1 satellites with 250+ kW of solar power each, Nvidia Rubin GPUs, laser-linked clusters, built at the Gigasat factory in Texas. Filed with the FCC in January 2026 for up to one million satellites; prototypes are due in early 2027.
When will orbital data centers be commercially available?
SpaceX targets commercial service as early as 2028, and Musk forecasts monthly orbital deployment overtaking terrestrial additions around the end of 2028. Google is more cautious, targeting economical orbital AI infrastructure in the 2030s.
Will data centers in space replace Earth data centers?
Not for decades, if ever. Launch costs, radiation, debris and the impossibility of repairs keep orbit as a complement, not a replacement. The near-term role: easing Earth’s power crunch and processing space-collected data where it’s gathered, instead of beaming it all down first.
The Bottom Line
A year ago, orbital data centers were a punchline. Today there is a GPU in orbit that has trained a neural network, a Google satellite that wrote a poem from space, a Chinese constellation running a real AI model, and FCC filings for over a million more satellites. The physics problems — heat, radiation, launch cost, debris — are real and unsolved. But for the first time, so is the hardware. Musk’s “100%” may be bluster; the trajectory underneath it is not.
About the Author
