In November 2025, a satellite the size of a mini-fridge trained a language model on the complete works of Shakespeare. Not in a lab. In orbit, 325 kilometers above your head, on the first NVIDIA H100 GPU ever to leave the planet.
The satellite was called Starcloud-1, and it wasn't a stunt. It was a proof of concept for one of the strangest ideas in tech: data centers in space — actual server farms, floating in vacuum, powered by sunlight that never sets.
Your first reaction is probably the correct one: this sounds insane. Rockets are expensive. Space is hostile. We have perfectly good dirt right here.
And yet Google is building AI data centers in space. NVIDIA is building chips for it. SpaceX has filed paperwork for up to a million satellites' worth of it. China already has twelve computing satellites flying. The answer to "why" isn't romance. It's a power bill.
Why AI Is Running Out of Power on Earth
The uncomfortable truth behind every chatbot you've talked to this week: the binding constraint on AI is no longer chips or money — it's electricity. Training clusters are ordered on two-to-three-year horizons; power plants take a decade. That mismatch is the whole story.
The industry's polite term is "capacity constraints." The honest term is a data center power crisis. Utilities in Virginia, Ireland, and Singapore are telling hyperscalers some version of no, you cannot have another gigawatt, we don't have one. Grid queues stretch for years; billion-dollar campuses sit waiting for substations. To understand why anyone would put servers on a rocket, you first need to see how deep this hole is.
How much energy does AI actually use?
So, how much energy does AI use? AI power consumption now rivals whole countries — and it's accelerating.
According to the International Energy Agency's "Energy and AI" report, global data center electricity consumption hit roughly 415 terawatt-hours in 2024 — about 1.5% of everything humanity generates — and the IEA's base case has that more than doubling to around 945 TWh by 2030. That's more than the entire annual electricity consumption of Japan. For data centers. Mostly because of AI.
The American numbers are starker. A Lawrence Berkeley National Laboratory report for the Department of Energy found US data centers consumed 4.4% of national electricity in 2023 — and projected between 6.7% and 12% of all US electricity by 2028 as AI electricity usage scales. Twelve percent. One in every eight watts, feeding AI data center power demand and the cloud.
Former Google CEO Eric Schmidt told Congress that data centers may need an extra 29 gigawatts by 2027 and 67 more by 2030: "These things are industrial at a scale I have never seen in my life." This is a man who ran Google. His scale calibration is not the problem.
And efficiency won't save us. Chips get dramatically more efficient every generation, yet total AI energy consumption still climbs, because cheaper compute means we use absurdly more of it. Economists call this the Jevons paradox. Your utility calls it a headache.
The water problem: cooling at hyperscale
Electricity is only half the bill. The other half is wet.
Every hyperscale data center is, thermodynamically speaking, a giant kettle: chips turn electricity into heat, and the heat usually leaves as evaporated water. US data center water usage went from 21 billion liters in 2014 to 66 billion liters in 2023, the overwhelming majority at hyperscale facilities. A single large site can drink up to 5 million gallons per day — a town of 16,000 households. A UN University report projected AI's water footprint could hit 9.3 trillion liters by 2030 in a high-adoption scenario — roughly the basic annual domestic water needs of all of Sub-Saharan Africa.
Now picture pitching a new data center to a drought-stricken county. Data center cooling isn't an engineering line item anymore; it's a political problem. Communities are saying no. Loudly.
The 2030 capacity shortfall
Then there's the money, which is somehow the least crazy part.
McKinsey calculates that companies will need to invest about $5.2 trillion in AI data centers by 2030, based on roughly 156 gigawatts of AI-related capacity demand. JPMorgan independently landed above $5 trillion and noted the snag: gas turbine lead times have ballooned to three or four years; nuclear plants take a decade. Capital is available. Watts are not.
Even if every project on the books lands on schedule, McKinsey warns the US alone could be short more than 15 GW by 2030. In spreadsheet-speak: the demand curve and the supply curve no longer intersect on this planet — which is exactly where engineers start looking up.
What Are Orbital Data Centers?
Time for the definition, now that you've earned it. Orbital data centers (also called space based data centers, or simply a data center in orbit) are satellites — or constellations of them — doing the work of a terrestrial server farm: storage, processing, AI inference, maybe one day training. Instead of a strained grid, solar panels in near-constant sunlight. Instead of evaporating a river, waste heat radiated into the void.
A space data center is not a space station with a server closet. The serious designs are racks of accelerators bolted to a spacecraft, woven together by lasers, with no human for hundreds of kilometers. Less Star Trek, more "warehouse moving at 7.8 km per second."
How orbital data centers work
The recipe, in plain language:
Pick the right orbit. The leading designs — Google's Project Suncatcher among them — use a dawn–dusk sun-synchronous orbit: a low-Earth orbit riding the line between day and night, keeping the satellite in almost perpetual sunshine. No night, no batteries the size of a school bus.
Generate power. Big, lightweight solar arrays feed the chips directly. No grid, no permits, no angry county commission.
Compute. Onboard GPUs or TPUs run workloads like a ground cloud — either radiation hardened chips built for space, or commercial chips wrapped in shielding. Google tested its Trillium TPU in a 67 MeV proton beam and found no hard failures up to the maximum tested dose — "likely acceptable" for inference.
Talk via laser. Satellites connect through free-space optics — laser satellite communication at terabit-class speeds — and reach Earth by optical or radio downlink. Google has demonstrated 1.6 Tbps on a single transceiver pair in the lab.
Send back answers, not raw data. You uplink the question, the orbital cluster does the heavy lifting, and only the distilled result comes home.
Why space solves the power and cooling problem
Two Earth problems simply don't exist up there.
Power: in the right orbit, a solar panel can be up to 8 times more productive than on Earth, per Google — no atmosphere, no clouds, no night. The Sun continuously emits over 100 trillion times humanity's entire electricity production. From orbit, you're standing next to the firehose.
Water: there isn't any, and you don't need it. Waste heat leaves by radiation, not evaporation — a headline advantage in a decade of record droughts, per the Thales Alenia Space ASCEND study.
The catch — and you knew there'd be a catch — is that radiating heat in a vacuum is much harder than it sounds. Hold that thought; we'll get to the radiators.
Who Is Building Data Centers in Space?
A few years ago this was a whitepaper genre. Now it's a competitive field with launched hardware, named partnerships, and real money — all chasing energy that doesn't queue, cooling that doesn't drink, and land that nobody can protest.
Google Project Suncatcher
Announced in November 2025, Google Project Suncatcher is the most academically serious entry. Google's research paper sketches clusters of 81 satellites flying in tight formation — within about a kilometer's radius — in dawn–dusk sun-synchronous orbit, carrying Google's TPUs and stitched together by optical links.
The near-term plan is refreshingly modest: a learning mission with Planet Labs to put two prototype TPU satellites up by early 2027 and see what breaks. CEO Sundar Pichai has said that within "a decade or so," orbital data centers will be seen as "a more normal way to build data centers."
Starcloud and NVIDIA
Starcloud is the startup that actually went first. The Redmond-based company launched Starcloud-1 in November 2025 — that 60-kg fridge with the first data-center-class H100 in space, roughly 100x more GPU compute than anything previously in orbit, per NVIDIA. The roadmap escalates fast: Starcloud-2 with a Blackwell GPU in late 2026, then a 200 kW node, then multi-megawatt — and a long-term pitch for a 5 GW orbital facility with a solar array measured in square kilometers. The company hit a ~$1.1 billion valuation in March 2026.
The Starcloud NVIDIA relationship matters because NVIDIA stopped being a bystander. At GTC in March 2026, it announced the Space-1 NVIDIA Vera Rubin module — a radiation-tolerant building block for orbital AI with up to 25x the H100's inference performance — plus IGX Thor for orbital edge computing. Jensen Huang's framing: "space computing, the final frontier, has arrived." When the world's most valuable chipmaker designs an NVIDIA space data center product line, the idea has formally left the fever-dream phase.
The ASCEND Project and Thales Alenia Space
Europe, true to form, commissioned a study — but a good one. ASCEND (Advanced Space Cloud for European Net zero emission and Data sovereignty) is an EU Horizon Europe feasibility study led by Thales Alenia Space with a consortium including Airbus, ArianeGroup, and HPE. In 2024 it reported promising results: the ASCEND data center concept is technically feasible and could be economically viable, targeting 1 GW of European compute in orbit before 2050.
The fine print: the carbon math only works with a future launcher about ten times less emissive than today's rockets, and no hardware has flown. But as government-backed roadmaps go, it's the most concrete one outside the US and China — and "data sovereignty" in the name tells you Europe sees orbit as strategic territory, not a gimmick.
Axiom Space and the orbital data center (AxDCU-1)
While others publish papers, Axiom ships boxes. The Houston company flew its shoebox-sized AxDCU-1 prototype to the International Space Station in 2025, then launched its first two free-flying Orbital Data Center nodes in January 2026 aboard Kepler Communications' optical relay satellites, linked by 2.5 Gbps laser connections.
The Axiom Space data center pitch is unglamorous and probably right: process satellite imagery in orbit instead of downlinking petabytes of raw pixels, serve defense customers who like their compute unreachable, and scale "from kilowatts to megawatts." The tortoise strategy — small, useful, already flying.
Other Notable Players in The Field
China's Xingshidai constellation. In May 2025, China launched the first 12 satellites of its "Three-Body Computing Constellation" — also covered under the Xingshidai banner — led by startup ADA Space with Zhejiang Lab. The dozen satellites deliver about 5 peta-operations per second over 100 Gbps laser interlinks; the full plan calls for 2,800 satellites, and Alibaba's Qwen model was reportedly running on orbit by early 2026. The China space data center program is not waiting for anyone's feasibility study.
SpaceX and Elon Musk. Musk has called space "a no-brainer for building solar-powered AI data centers," and SpaceX has filed with the FCC for up to one million data-center satellites — the "SpaceX million data centers" plan, in headline shorthand. But SpaceX's own pre-IPO filing speaks in a quieter voice: orbital AI compute involves "unproven technologies" and "may not achieve commercial viability." When the loudest evangelist's lawyers write that sentence, keep both halves in mind.
Meta and Overview Energy. Meta took a different road: keep the servers on the ground, get the power from space. In April 2026 it signed a first-of-its-kind deal with startup Overview Energy for up to 1 gigawatt of space based solar power — geosynchronous satellites collecting sunlight and beaming it down as low-intensity near-infrared light onto existing solar farms, extending their generating hours into the night. Demo in 2028, commercial delivery around 2030. It's the first time space solar energy beaming has been bought at nuclear-plant scale to feed AI — and arguably the most bankable bet here, because a solar power satellite doesn't care about latency.
Add NTT and SKY Perfect JSAT in Japan, Blue Origin, and ESA-commissioned studies with IBM, and the picture is clear: every major tech power is hedging the same bet.
The Alternatives: Underwater and Arctic Data Centers
Before we strap servers to rockets, fairness demands a look at Earth's two proven "free cooling" frontiers — the ocean and the Arctic. Data center alternatives don't have to clear the atmosphere to be useful.
Microsoft Natick and underwater servers
Microsoft Natick is the patron saint of weird data center experiments. In 2018, Project Natick sank a sealed cylinder with 855 servers off Orkney, Scotland — an underwater data center cooled by the sea itself. Retrieved two years later, only 6 of 855 servers had failed, roughly 8x more reliable than the identical land-based control group. The secret: a nitrogen atmosphere and zero humans bumping into things.
And then Microsoft killed it. The fatal flaw wasn't reliability — it was access. A sealed undersea data center can't be upgraded without hauling it up, and AI hardware goes stale in 1–3 years. China's Highlander has since commercialized the concept off Hainan and Shanghai (a 24 MW underwater facility powered ~97% by offshore wind went live in 2025–2026). The idea has legs — just not hyperscale ones, so far. Remember Natick's lesson; it haunts the orbital plans too.
Peter Thiel's Panthalassa project
The most cinematic entry: in May 2026, Peter Thiel led a $140 million Series B into Panthalassa, an Oregon company building wave-powered, floating AI data centers — the marquee Peter Thiel data center bet, valuing the startup near $1 billion.
The Panthalassa project nodes are 85-meter, mostly submerged steel structures that bob in the open Pacific, converting wave motion into electricity, cooling chips with seawater, and talking to shore via Starlink. Pilots deploy in 2026; commercial systems target 2027. Skeptics point at corrosion, storms, and satellite-link bandwidth — but as a dress rehearsal for off-grid autonomous compute, the ocean is a far gentler teacher than orbit.
Arctic and Nordic data centers
And then there's the option that already works and bores everyone: build where it's cold. Meta's Luleå facility in northern Sweden — opened 2013, ~100 km from the Arctic Circle — cools itself with outdoor air year-round and runs entirely on local hydropower, hitting a power usage effectiveness around 1.07. Nearly perfect.
The broader Nordic data center belt — Norway's fjord-side facilities, Iceland's geothermal sites, the giant new AI campuses near Narvik serving Microsoft and OpenAI — is the most mature alternative here. An Arctic data center needs no rockets, no submarines, no new physics. Its limits are mundane: only so many cold rivers, so much hydro, so much fiber. Which is exactly why the conversation keeps drifting upward anyway.
The Engineering Challenges
Now for the cold shower. Putting a hyperscale data center in orbit means solving three problems Earth solves for free. Space computing is hard precisely where ground computing is trivial.
Powering a data center with orbital solar
Orbital solar power is the good news. Sunlight in space is stronger (no atmosphere), more reliable (no weather), and in a dawn-dusk sun synchronous orbit, nearly constant — the up-to-8x edge Google cites, and the core of every space solar power pitch.
The bad news is mass. A feasibility study puts a 1 MW orbital cluster at around 5,600 m² of solar array plus thousands more of radiator — 34–59 kg of hardware per kilowatt of IT power. A modest 10 MW cluster is hundreds of tons in orbit. Every panel rides a rocket, and solar cells degrade under radiation, so you oversize from day one. Free energy, expensive real estate.
Cooling in a vacuum
The dirty secret of "space is cold": vacuum is the best thermal insulator there is. Your thermos exploits this. No air for convection, no water for evaporation — heat leaves only by infrared radiation, governed by the Stefan–Boltzmann law.
The numbers are brutal. Rejecting 1 MW of waste heat requires roughly 1,200–1,600 square meters of radiator surface — a structure bigger than the compute it serves, shedding heat over a thousand times slower than water cooling rips it off an AI chip on Earth. Scaling ISS-style thermal hardware to megawatt class implies ~100 tons of radiators — potentially 10x the mass of the servers themselves. The entire ISS, for reference, rejects just 70 kW.
Run radiators hotter and they shrink, but then your chips cook closer to their limits. Thermal management, not power, may be the real binding constraint of the orbital data center — Google's own paper flags it as unsolved.
Latency and laser connectivity
The third wall is physics' speed limit. Laser interlinks between satellites are spectacular — terabit-class, near-zero added latency over thousands of kilometers. The trip to the ground is not: realistic round-trip latency runs from tens to a couple hundred milliseconds, and cloud cover can flat-out block an optical downlink.
That draws a hard line through the AI workload map. Inference, batch processing, and Earth-observation analytics: fine. Frontier-model training — which demands microsecond-tight coupling between thousands of chips — stays on the ground. Inconveniently, training is the workload driving the power crisis. The orbital cloud, at least at first, will be an inference cloud.
The challenges above don't exist in isolation. Every infrastructure decision — space, ground, or Arctic — involves the same set of tradeoffs, just distributed differently. Some problems that orbit eliminates entirely, Earth has been quietly solving for decades. Others that Earth treats as line items, space turns into existential engineering questions. The table below maps all three across the dimensions that actually determine whether a data center gets built, funded, and kept running.
| Orbital | Ground | Arctic | |
|---|---|---|---|
| Power Source | Unlimited Potential | Constrained | Limited but Clean |
| Water Usage | None | Critical Problem | Near Zero |
| Cooling Method | Unsolved at Scale | Solved & Expensive | Solved & Nearly Free |
| Build Cost | 3–7x Ground Cost Today | High but Known | Cost-competitive |
| Latency | Physics Limit | Excellent | Acceptable |
| Workloads | Inference Only | Universal | Universal |
| Status 2026 | Early Stage | Capacity Crisis | Operational |
The Real Bottleneck: Launch Costs
Strip away the radiators and the lasers, and the whole argument collapses to one number: launch cost per kg.
Today, a reusable Falcon 9 puts mass in low-Earth orbit at roughly $1,500–2,900 per kilogram. A 1 MW orbital cluster weighs around 40,000 kg at realistic mass budgets — over $100 million in launch costs alone, before you've paid for a single GPU, ground station, or replacement satellite. At current prices, independent analyses agree, orbital compute runs several times the cost of equivalent terrestrial capacity. The economics don't close. Period.
Everything therefore hinges on Starship. SpaceX targets a Starship launch cost of $100–200/kg with full, rapid reuse — internal projections run toward $10–20/kg at airline-like cadence — while independent observers put mature SpaceX Starship cost per kg at $100–500. Nobody actually knows, because the required cadence hasn't been demonstrated yet.
What we do know is the threshold. Google's analysis finds that at around $200/kg, a space-based data center becomes roughly cost-comparable with a terrestrial one's energy costs — plausibly reached by the mid-2030s.
The clean way to think about it: cooling and power are hard engineering; launch economics are existential. Starship-class pricing is necessary for orbital AI at scale — and even then, not sufficient. A 100 MW constellation could still demand 50+ dedicated launches, plus spacecraft mass production and in-orbit assembly. The rocket equation forgives nothing.
The Legal Gray Zone of Computing in Orbit
Suppose the engineering works. Whose laws apply to a server farm that belongs to no country and orbits all of them every 90 minutes?
The foundation is the 1967 Outer Space Treaty: space belongs to no nation, but the state where a spacecraft is registered "retains jurisdiction and control" over it. In practice, a data center in orbit is treated like a ship at sea — governed by the law of its flag state.
You can see where this is going. Ships taught us exactly what happens next: flags of convenience. Nothing stops an operator from registering orbital compute wherever the data, AI, and tax rules are friendliest — while serving users everywhere. Meanwhile, regulations like GDPR follow the data, not the hardware, so a US-flagged satellite processing European personal data is theoretically still on the hook. Multiple states can plausibly claim authority over the same compute job, and the treaties — written to assign blame for falling debris, not to referee AI governance — offer no tiebreaker.
Who audits an autonomous AI facility with no human aboard? When The Register asked Axiom which country's laws govern its on-orbit data processing, the company didn't respond. That silence is the current state of the law. Orbit is not a legal escape hatch — it's a legal traffic jam where the pile-up hasn't happened yet.
Will AI Data Centers in Space Actually Happen?
After all that — the terawatt-hours, the radiators, the rocket math, the lawyerless courtroom 400 km up — here's where I land.
Yes, they'll happen. No, not the way the keynotes promise.
The momentum is real. Within a single recent stretch, Starcloud flew an H100, Axiom launched free-flying nodes, China orbited a dozen Xingshidai satellites, NVIDIA announced space silicon, Google booked a 2027 demo with Planet, Meta bought a gigawatt of space solar, and SpaceX filed for a million-satellite constellation. Not vaporware behavior.
But the skeptics hold the better near-term hand. IEEE Spectrum's analysis pegged a 1 GW orbital system at roughly 3x the cost of its terrestrial twin — down from earlier 7–10x estimates, but still a chasm. Radiative cooling stays stubbornly massive. Latency walls off training, today's dominant workload. Hardware that refreshes every 1–3 years clashes with satellites you can't upgrade — Natick's flaw, at escape velocity. Astronomers warn about the space junk problem: satellite streaks already contaminate a growing share of asteroid-hunting telescope images, and mass reentries deposit ozone-eating aluminum oxide in the upper atmosphere — space debris from thousands of short-lived compute satellites is a bill someone eventually pays. And SpaceX's own IPO filing concedes the whole category "may not achieve commercial viability."
So here's my honest scorecard for the future of data centers:
Late 2020s: demonstrators and niches — kilowatt-scale space AI computing, Earth-observation processing, defense workloads that pay extra for unreachable hardware.
Early-to-mid 2030s: if Starship hits a few hundred dollars per kilogram, tens to hundreds of megawatts in orbit for inference and batch work — the first space based data centers that earn their keep.
Replacing terrestrial hyperscale: not this generation. Even bullish forecasts speak of orbital costs converging with ground costs by 2035 for certain workloads — not winning outright.
The deciding data point arrives soon, and it isn't a press release: it's Google's two TPU satellites in early 2027. If those chips survive, link up, and compute on budget, AI infrastructure 2030 planning everywhere gets a new line item. If they don't, the sector gains a convenient excuse to wait for cheaper rockets.
Either way, the power crisis driving all this isn't going anywhere — and that, not the romance of orbit, is why the idea refuses to die. Earth ran out of easy watts. Space has nothing but.
Keep an eye on early 2027. The sky is about to run a benchmark.
