Data Centers in Space: Hype, Physics and the $100-per-Kilogram Question
In 2026 SpaceX asked to launch a million computing satellites, Google put its first AI chips into orbit and a Seattle startup became a unicorn on the idea. This guide explains why orbit suddenly looks attractive for AI, what has actually flown, what physics makes it hard, and the one number that decides whether it ever beats a data center on the ground.
Why anyone wants to put data centers in orbit
AI's limiting resource is no longer just chips. It is electricity, cooling water, land and the years it takes to connect a new site to the grid. Big AI campuses now draw hundreds of megawatts to gigawatts, and in many places local opposition, grid queues and water rights decide where they can go. In June 2026 CNBC summed up the mood: no one wants AI data centers on Earth, so do they make sense in space?
Orbit offers what the ground is short of. In the right orbit the Sun almost never sets, so solar panels work nearly all the time without batteries for the night. Nobody lives next door. There is no grid queue and no water bill. The catches are that everything must be launched at thousands of dollars per kilogram, heat can only leave by radiating it away, chips are bombarded by radiation, and nothing can be repaired.
What changed in 2025 and 2026 is that this went from conference slides to hardware and regulatory filings. The question this article tries to answer is not "is it possible?" (small versions clearly are) but "at what launch price and for which jobs does it beat the ground?"
The pitch in one table
Here is how proponents compare a data center on the ground with one in a dawn-dusk sun-synchronous orbit, the orbit Google's paper uses, which rides the line between day and night so its panels see the Sun almost continuously.
Earth vs orbit, resource by resource
Green = advantage, red = disadvantage. Qualitative summary of the sources in this article.
| Resource | On Earth | In orbit |
|---|---|---|
| Energy | Grid power with long connection queues; solar only works part of the day | Near-continuous sunlight in dawn-dusk orbit, no atmosphere filtering it |
| Cooling | Air or water cooling; large water use in some designs | No air or water at all: heat must be radiated away by large, heavy panels |
| Land and permits | Local opposition, zoning, environmental reviews | No neighbours, but a new fight over orbital space, debris and astronomy |
| Getting there | Trucks | Rockets, at roughly $1,400 to $3,600 per kg today depending on the source |
| Repairs | Technicians swap failed parts daily | Effectively none; spares must be launched in advance |
| Radiation | Shielded by the atmosphere and magnetic field | Total dose and bit flips; memory is the weak point in Google's tests |
| Network | Fibre at terabits per second, microsecond latency inside a hall | Laser links of roughly 0.1 to 1.6 Tbps; a few to tens of ms to the ground from low orbit |
From study to orbit: 2024 to 2026
Click an event. Confidence badges say how firmly each one is documented.
Timeline
Scroll sideways on a phone.
Who is building what
Filter by status. "Flown" means hardware is in orbit (or reached the Moon); "Filed" means a regulatory application; "Planned" means announced plans only.
The players
Funding and satellite counts are as reported; filings are requests, not approvals.
| Who | What | Status |
|---|
Anatomy of an orbital data center satellite
Whatever the company, the design ends up with the same parts. Click each one to see why it is there and what makes it hard.
One compute satellite
A schematic, not any specific company's design. Proportions are illustrative: in practice the radiators and solar arrays dwarf the computer.
The physics problems
Six engineering constraints decide whether any of this works. Pick one.
Six hard problems
The economics: it all comes down to launch price
Every serious analysis lands on the same variable. The chips cost the same in orbit as on the ground; what differs is everything you must launch to power and cool them, and the price of launching it. JPL physicist Slava Turyshev's April 2026 paper estimates about 29.4 kg per kW just for solar arrays, storage and radiators in a 1 MW system, and 34 to 59 kg per kW once the rest of the spacecraft is included. If a ground data center costs $10,000 to $40,000 per kW, that leaves only $250 to $1,000 per kg to cover both launch and building the spacecraft, which he puts at 3.4 to 13.5 times below the current public Falcon 9 price for a dedicated launch.
What different analysts say the launch price must reach
Each bar is the threshold an analysis gives for orbit to compete (lower is harder). The grey bars show reported current prices for comparison. Log scale.
- BCG (2026): a 20-year cost of about $660 to $750 million per MW in orbit versus $230 to $300 million on Earth, a 2.5 to 3 times premium today, narrowing to about 1.5 times over the next decade but unlikely to close.
- Forethought: competitive with an aggressive terrestrial build-out near $100 per kg, with savings at $50 per kg; about 38% extra non-compute hardware because failed chips cannot be swapped; probably not a meaningful share of compute before 2030, possibly much of the build-out in the 2030s.
- Bain: at current prices, launch alone would cost more than the computers for a 100 MW system; $50 to $100 per kg is needed.
- Google: if launch falls below about $200 per kg by the mid-2030s, the launched cost of a satellite's power capacity, spread over its life, could roughly match what US data centers pay for electricity.
Break-even calculator
Try the trade-off yourself. The model is deliberately simple: the chips are left out because they cost the same in both places, so it compares the cost of providing one kilowatt of powered, cooled capacity.
Does orbit beat the ground?
Orbit cost per kW = mass per kW ร (launch price + spacecraft build cost per kg). Earth cost per kW = your terrestrial benchmark (building, power and cooling over the same lifetime). This site's simplification of Turyshev's framing; ignores financing, links, operations and replacement.
What runs well in orbit, and what does not
Not all computing is equal. A July 2026 preprint by Kees van Berkel (arXiv 2607.14172) uses bandwidth models to conclude that inference in low orbit may be feasible, but that training frontier-scale models there is unlikely to compete. Training needs thousands of chips exchanging data constantly at terabits per second with microsecond delays; laser links between satellites today run at roughly 100 to 200 Gbps, and Google's best bench test reached 1.6 Tbps for one pair. Turyshev's paper reaches a similar ranking: the credible early markets are processing data that is already in space and computing built into communications networks.
Workload fit
Our summary of the analyses above. โ good fit, ~ possible with caveats, โ poor fit. Hover a cell for the reason.
A scale reality check
The filings are measured in tens of thousands to a million satellites. Musk's projection for SpaceX is that launching a million tonnes a year of compute satellites at about 100 kW per tonne would add roughly 100 GW of AI capacity a year. To see what those numbers mean, compare them with things that exist.
Satellites: what exists versus what has been filed
Log scale. Filed counts are applications, not approvals or plans with dates.
Debris, astronomy and the regulatory fight
The filings have triggered the biggest argument yet over who gets to use low Earth orbit.
- Astronomers: the American Astronomical Society asked the FCC to deny an "Orbital Compute" application, citing overlap with earlier filings and harm to astronomy. (Coverage differs on whether that application is for 100,000 or a million satellites.)
- NASA: on about 5 May 2026 NASA objected to Blue Origin's 51,600-satellite Project Sunrise, citing "significant safety and sustainability concerns".
- Environmental groups: DarkSky International, Environment America and PEER asked the FCC for a programmatic environmental impact statement before licensing any of these constellations, citing rocket emissions, pollutants from re-entering satellites and ozone effects.
- Debris: critics warn that a million objects in low orbit would sharply raise collision risk, and that automated collision avoidance is only as good as tracking data. This is the critics' framing, not a settled finding.
- Rivals: Blue Origin called SpaceX's plan "profoundly disproportionate"; SpaceX called Blue Origin's arguments "naive" and "simplistic". One Italian report expects FCC decisions in 2026 to 2027; none had been made in the coverage we found.
Europe's ASCEND study (Thales Alenia Space, 2024) adds an environmental condition of its own: space data centers only beat ground ones on emissions if launchers become about ten times less emitting over their whole lifecycle.
Lessons from data centers in odd places
๐ Microsoft's Project Natick (under the sea)
A sealed, nitrogen-filled server pod on the seabed off Orkney had one eighth the failure rate of a land control group: 6 of 855 servers failed against 8 of 135 on land. Reporters credited the dry, inert atmosphere and the absence of people jostling equipment. Microsoft still ended the programme ("I'm not building subsea data centers anywhere in the world", one executive said). The lesson for orbit: sealed hardware left alone can be reliable, but reliability alone did not make the economics work.
๐ Lonestar's Freedom (on the Moon)
An 8 TB solid-state data store rode Intuitive Machines' Athena lander, launched 26 February 2025, which landed on its side in a crater on 6 March and died. Lonestar says the data center was intact, ran during the trip and briefly on the surface, and exchanged data and telemetry with the lander. We found no confirmation that the 8 TB archive was read back. The lesson: the computer is rarely the part that fails; the ride is.
Three scenarios to 2035
๐ป Bear: a niche
Starship pricing stays well above $500/kg, radiators and replacements eat the energy savings, and regulators slow the mega-constellations. Orbital compute stays a niche for processing satellite imagery and defence edge computing. Watch: published Starship prices, FCC decisions, first in-orbit failure rates.โ๏ธ Base: inference and space-native data
Launch falls toward $200 to $300/kg by the early 2030s. Orbit carries a growing share of inference and of processing for data already in space, at a premium buyers accept for speed of deployment and no grid queue. Training stays on Earth. This matches BCG's "narrows but does not close" and Turyshev's early-market ranking.๐ Bull: much of the 2030s build-out
Fully reusable launch reaches about $100/kg or below, spacecraft mass per kW falls, and laser links scale up. Forethought's case: space could host much of the 2030s compute build-out. Requires thousands of heavy launches a year and regulatory approval of very large constellations.Reading the numbers: where sources disagree
| Claim | What the sources say |
|---|---|
| Today's launch price | Estimates for Falcon 9 range from about $600/kg (Bain) and $1,400โ1,800/kg to $3,000โ3,600/kg, depending on rideshare vs dedicated launch and the source. One analysis found the only public Starship price works out to about $600/kg. |
| "$200/kg by the mid-2030s" | A learning-curve projection in Google's paper, not an announced price. Others put the needed figure at $100/kg (BCG, Forethought) or $50โ100/kg (Bain). |
| Starcloud trained an AI model in space | A small nanoGPT model on Shakespeare plus running Gemma, according to the company. A real first, but a demonstration, not a data center. |
| Google's satellite is in orbit | Several outlets report the 1 Oct 2026 launch on Transporter-18 and that Google made contact; no performance results had been published by 23 Oct. |
| Axiom launched data center nodes in January 2026 | Reported by one blog (11 Jan 2026); Axiom's own April 2025 release announced the plan. Not confirmed from Axiom or Kepler. |
| Nvidia's Space-1 Vera Rubin module | Announced at GTC in March 2026 with "up to 25 times H100" inference compute; TechCrunch reported none had been produced or shared at the time; no ship date. |
| "Gigawatt data centers in space within 10 to 20 years" | Jeff Bezos' forecast at Italian Tech Week, an opinion. |
| Radiator area of ~1,200 mยฒ per MW | Appears in one source; not verified. IEEE Spectrum's point that degrading coatings require at least 40% extra radiator mass is better sourced. |
| China's 1,000 POPS constellation | The goal for 2,800 satellites; the first 12 (May 2025) total about 5 POPS. We found no 2026 launches. |
๐ References and papers
- Research papers and analyses
- Towards a future space-based, highly scalable AI infrastructure system design (Google, arXiv 2511.19468)
- Orbital Data Centers: Spacecraft Constraints and Economic Viability (Turyshev, arXiv 2604.27197)
- van Berkel on LEO inference vs training (arXiv 2607.14172)
- Space-Based Data Center Architectures for 6G and Beyond (arXiv 2603.18601)
- BCG: Space-based data centers cost outlook (2026)
- Forethought: Will we really put data centers in space?
- Bain: Orbital data centers, beyond the grid
- IEEE Spectrum: Why thermodynamics rules future orbital data centers
- Analysys Mason: Orbital data centres can complement terrestrial ones
- Company announcements and filings coverage
- Google: Project Suncatcher ยท DCD on the first test launch ยท Fortune
- DCD: SpaceX files for a million-satellite orbital data center constellation ยท Via Satellite on SpaceX and xAI
- TechCrunch: Starcloud raises $170M ยท Wikipedia: Starcloud
- Engadget: Blue Origin's Project Sunrise ยท DCD: Bezos on gigawatt data centers in space
- Via Satellite: Nvidia's space chips
- Axiom Space: orbital data center nodes
- Live Science: China's Three-Body Computing Constellation
- Thales Alenia Space: ASCEND feasibility study results
- DCD: Lonestar's lunar data center after Athena tipped over
- Context and criticism
- CNBC: Do space-based AI data centers make economic sense?
- Scientific American on Project Suncatcher
- Bloomberg: Are AI data centers in space actually feasible?
- SatNews: NASA objects to Project Sunrise
- WebProNews: environmental groups ask the FCC to halt
- Windows Central: Microsoft ends Project Natick
- On this site
- AI chips and compute infrastructure
- The compute frontier, explained