Space Data Centers: Why Google and SpaceX Want Servers in Orbit

Space data centers are clusters of computing hardware launched into orbit, powered by solar panels that see the Sun almost constantly and cooled by radiating heat into the vacuum of space. The idea has gone from paper to hardware in under two years: Google put a prototype satellite carrying its own TPU chips into orbit on October 1, 2026, Starcloud has been running an Nvidia H100 GPU in space since November 2025, and SpaceX has asked U.S. regulators for permission to fly up to one million solar-powered AI satellites. The case rests on three things Earth cannot easily offer: near-continuous sunlight, cooling without water, and no land, grid queue or permits.

Whether orbital data centers can ever beat a warehouse of servers on cost is a live engineering question. This guide covers who has done what, the physics, the launch price that decides the economics, a side-by-side table, a timeline and the sceptic case.

Why put data centers in space? The three arguments

Every serious proposal leans on the same three advantages.

• Power. Above the atmosphere, sunlight arrives at about 1,361 watts per square metre, the total solar irradiance NASA measures from satellites, with no clouds and, in the right orbit, effectively no night. Google's Project Suncatcher team calculates that a panel in a dawn-dusk sun-synchronous orbit can be up to eight times more productive than on the ground.
• Cooling. Terrestrial AI data centers rely on chillers and evaporative towers that can use billions of litres of water a year. In orbit heat leaves only as infrared radiation from large panels: partly a drawback, but zero water and no refrigeration.
• Land, grid and permits. On Earth the bottleneck for new AI capacity is increasingly grid connections and local opposition rather than chips. SpaceX told the Federal Communications Commission its satellites "will achieve transformative cost and energy efficiency" with little operating or maintenance cost.

Who is doing what: the space data center projects, with dates

Google: Project Suncatcher

Google announced Project Suncatcher on November 4, 2025, with a research paper (since published in the journal Joule) describing compact constellations of solar-powered satellites carrying its Tensor Processing Units, linked by lasers and flying in close formation in a dawn-dusk sun-synchronous orbit at about 650 km. Two findings mattered most. Google's Trillium (v6e) TPUs survived a 67 MeV proton beam better than expected: the high-bandwidth memory only began to misbehave after a cumulative dose of 2 krad(Si), almost three times the 750 rad(Si) a shielded five-year mission would receive, with no hard failures up to 15 krad(Si). And a bench test of the laser links reached 1.6 Tbps over a single transceiver pair.

On October 1, 2026, the first Suncatcher prototype, built with the satellite company Planet, launched on SpaceX's Transporter-18 rideshare mission. Google confirmed contact and said the satellite carries four TPUs and will run its open Gemma model to see how the chips cope with launch stress, radiation and thermal extremes. A two-satellite mission with Planet is planned for 2027 to test optical links for distributed machine learning.

Starcloud: the first H100 in orbit

Starcloud, a Redmond, Washington startup founded in 2024, launched Starcloud-1 in early November 2025 aboard a SpaceX rocket. It carries one Nvidia H100, about 100 times more GPU compute than had ever flown. Within weeks the company said it had trained a small language model in orbit and run Google's Gemma model on the H100. Starcloud raised $170 million in March 2026 and $250 million more on August 21, 2026 at a $2.3 billion valuation, with Nvidia and Cisco among the new investors, and is working with Nvidia on the Space-1 Vera Rubin Module, a space-rated computing platform announced at GTC on March 16, 2026. Its stated goal is 88,000 satellites and 20 gigawatts of orbital compute; its white paper sketches a 5 GW data center with panels roughly 4 km on a side. CNBC reported in December 2025 that a second satellite with Nvidia Blackwell hardware was planned for October 2026; as of October 19, 2026, Starcloud has not announced that launch, and its August release speaks of production lines for a next-generation Starcloud-3.

SpaceX: one million satellites and the AI1 design

SpaceX is the largest bet by scale. On January 30, 2026, it filed with the FCC for a constellation of up to one million solar-powered satellites to operate as AI data centers. Reuters noted that operators often request more than they intend to fly: SpaceX sought approval for 42,000 Starlink satellites and had roughly 9,500 in orbit at the time, out of about 15,000 satellites in space overall. SpaceX and Elon Musk's xAI combined in February 2026, and ahead of its June 12, 2026 stock market debut, SpaceX showed a first-generation AI satellite, AI1, built around large solar arrays, a compute module, a deployable radiator and laser links. Musk said the first version would likely use Nvidia chips with the power of one GB300 rack: "A lot of this is technology we've already made for the Starlink V3 satellites." The IPO prospectus said deployments could begin "as early as 2028"; executives told investors they aim to fly demonstrators by late 2027, with a Bastrop, Texas factory reaching meaningful production by the end of 2027. The enabler is Starship, which reached orbit for the first time on September 28, 2026 and deployed 26 Starlink V3 satellites.

Axiom Space: the first dedicated nodes

Houston-based Axiom Space builds small orbital data center nodes for edge processing and secure storage. Its first two launched on January 11, 2026 on Kepler Communications' optical relay satellites, with 2.5 Gbps laser links; Axiom says both are operating.

Lonestar: data storage on the Moon

Florida-based Lonestar Data Holdings treats the Moon as a disaster-recovery vault. Its Independence payload completed a data storage test from the lunar surface on Intuitive Machines' IM-1 lander in February 2024; its Freedom payload launched on February 26, 2025 aboard the IM-2 Athena lander; Lonestar says it completed data transfers in transit and lunar orbit. Athena touched down near the lunar south pole on March 6, 2025, but came to rest on its side and surface operations ended early.

Europe: the ASCEND study

ASCEND, led by Thales Alenia Space for the European Commission with partners including Airbus, ArianeGroup and the German space agency DLR, reported on June 27, 2024 that space data centers are technically feasible and could cut the carbon footprint of digital services, provided Europe builds a launcher about ten times less emissive over its life cycle. The target is one gigawatt in orbit before 2050.

Jeff Bezos and Blue Origin

In Turin on October 3, 2025, Amazon founder Jeff Bezos predicted gigawatt-scale data centers in space within 10 to 20 years. In December 2025, Reuters, citing the Wall Street Journal, reported that Blue Origin had worked on the technology for over a year. In May 2026, Bezos told CNBC the outcome was "very realistic" but a two-to-three-year timeline was "a little ambitious".

The physics and engineering: power, heat, radiation, latency, cost

Solar power in orbit versus on the ground

The solar constant above the atmosphere is about 1,361 W/m². A fixed panel on the ground averages far less over a year once night, atmosphere, weather and Sun angle are counted; that gap is where Google's "up to eight times" comes from. In a dawn-dusk sun-synchronous orbit the satellite rides the line between day and night, so its panels stay lit through almost every orbit. Constant input means less battery mass, and every kilogram of battery costs launch money; see our explainer on sodium-ion batteries.

Cooling: radiation is the only way out

On Earth a data center dumps heat into air or water. In vacuum heat can only leave as infrared radiation, governed by the Stefan-Boltzmann law. An ideal black radiator at about 27 °C (300 K) emits roughly 460 watts per square metre per side, so a 1 GW system needs radiators covering on the order of one to two square kilometres. That is why Starcloud's 5 GW concept is drawn about 4 km across, why SpaceX's AI1 includes a deployable liquid radiator, and why Google lists thermal management first among the "significant engineering challenges" that remain. Running chips hotter helps, since radiated power rises with the fourth power of temperature, but silicon has limits.

Radiation and chip reliability

Trapped charged particles and cosmic rays cause a slow total ionising dose that degrades transistors and single-event upsets that flip bits in memory. Suncatcher's proton-beam results suggest commercial accelerators can survive a five-year mission at around 650 km with modest shielding, and Starcloud designs around a five-year chip lifetime. Nobody has yet run a large cluster in orbit long enough to measure real failure rates, which is what the current prototypes are meant to find out.

Latency and laser links

Low Earth orbit at 500 to 650 km adds only a few milliseconds each way, which is why Starlink latency is comparable to ground broadband; see how Starlink internet works. The harder problem is bandwidth between satellites, since training large AI models needs tens of terabits per second between accelerators. Starlink's optical inter-satellite links are the proven precedent; Google's approach is to fly satellites only hundreds of metres apart so dense wavelength-division multiplexing can push each link into the terabit class, as its 1.6 Tbps bench test showed. Downlinking large volumes of data through weather is a separate bottleneck.

Launch cost: the number that decides everything

Google's analysis is the clearest public benchmark. From historical and projected launch prices, its team concluded that with a sustained learning rate, prices may fall below about $200 per kilogram by the mid-2030s, at which point the cost of launching and operating a space-based data center could become roughly comparable, per kilowatt per year, to the reported energy costs of an equivalent terrestrial data center. Today's prices are several times higher, which is why every plan depends on fully reusable heavy launch; SpaceX's filing says the scale only works with vehicles that can "deploy millions of tons of mass per year to orbit when launching at rate".

Maintenance, debris and end of life

Nobody swaps a failed board at 650 km, so satellites must be built for redundancy and remote recovery, and must deorbit cleanly. Adding thousands of large, radiator-winged spacecraft to orbits that already hold roughly 15,000 satellites raises collision and debris questions regulators have barely begun to address, and larger surfaces take more micrometeoroid hits. Everything in orbit is temporary, as the spent rocket stage that hit the Moon showed.

Ground vs. orbital data centers: side by side

Factor	Ground data center	Orbital data center
Power	Grid, on-site gas, solar, wind, nuclear deals	Solar arrays in near-constant sunlight; up to 8x ground yield (Google)
Cooling	Air, chilled water, evaporative towers	Radiation only; square kilometres of radiator per gigawatt
Water	Billions of litres a year at hyperscale sites	None
Cost per kW	Known; electricity and chips dominate	Launch mass dominates; comparable only below about $200/kg (Google)
Latency	Milliseconds to tens of milliseconds	A few extra milliseconds; downlink capacity is the constraint
Repairability	Hot-swap a board in minutes	None; replace whole satellites
Regulatory	Zoning, permits, interconnection, water rights	Spectrum and orbital licences (FCC, ITU), debris rules, data sovereignty

What it means for AI energy demand on Earth

The International Energy Agency estimated that data centres consumed about 415 TWh in 2024, roughly 1.5% of global electricity, with the United States accounting for 45%. In its base case, consumption more than doubles to around 945 TWh by 2030 and reaches roughly 1,200 TWh by 2035; the IEA's 2026 update kept the 2030 figure at about 950 TWh while raising the 2025 estimate to 485 TWh.

Against those numbers, space will not relieve the grid this decade: the hardware in orbit today adds up to a few kilowatts, and SpaceX's own prospectus points to deployments no earlier than 2028. Google's framing is the honest one: the core concepts are "not precluded by fundamental physics or insurmountable economic barriers", but the engineering and launch economics still have to be proven at scale. Orbit is a bet on the 2030s.

Timeline: announced milestones, 2024 to the 2030s

Future entries are announced targets, not commitments; this table is the part of the guide most likely to need refreshing.

Date	Milestone	Who	Status
February 2024	Data storage test from the lunar surface (IM-1)	Lonestar	Done
June 27, 2024	ASCEND study: 1 GW in orbit before 2050 if a 10x cleaner launcher is built	Thales Alenia Space	Published
Feb 26 to Mar 6, 2025	Freedom payload flies to the Moon on IM-2; lander tips on landing	Lonestar	Partial
November 2025	Starcloud-1 launches with the first Nvidia H100 in orbit; Project Suncatcher announced (Nov 4)	Starcloud, Google	Done
January 11, 2026	First two orbital data center nodes launch	Axiom Space	Operating
January 30, 2026	FCC filing for up to one million AI satellites	SpaceX	Pending
June 2026	AI1 satellite shown; IPO filing says deployments "as early as 2028"	SpaceX	Announced
August 21, 2026	$250 million raise at $2.3 billion valuation; 88,000-satellite goal	Starcloud	Done
October 1, 2026	First Suncatcher prototype with four TPUs launches	Google, Planet	In orbit
2027	Two-satellite Suncatcher mission; SpaceX demonstrators; Bastrop factory output	Google, SpaceX	Target
2028	Earliest deployments per SpaceX prospectus	SpaceX	Target

The sceptic case, stated plainly

Chips already run in space; the debate is whether the whole system beats a solar farm, a battery bank and a building. The main objections:

• Radiators are the real payload. Square kilometres of thermal panels per gigawatt add mass, drag and deployment risk. Google's paper says gigawatt-scale constellations may need "a more radical satellite design" integrating power, compute and cooling.
• The economics hinge on a forecast. Parity assumes launch prices below roughly $200/kg; Reuters noted in June 2026 that Starship remained years behind its original targets and had yet to demonstrate rapid reusability.
• Chips age faster than satellites. Ground operators swap in new accelerators every few years; an orbital cluster is frozen at launch.
• No repair, more debris. Morgan Stanley analysts, cited by CNBC, listed radiation, in-orbit maintenance, debris hazards and regulatory issues around data governance and space traffic as the main hurdles.

The builders argue these are bounded engineering problems. Bezos's "a little ambitious" and Musk's statement that SpaceX was making no promises about when or how large the deployment would be say the same thing: the physics works, the schedule is the uncertain part.

What to watch next

Three signals will show whether space data centers are becoming real: how Google's TPUs behave over months in orbit; whether SpaceX flies demonstrators by late 2027 and what Starship launch prices settle at; and whether anyone sells orbital compute to a paying customer at a price that undercuts a ground cloud. For more on frontier computing, read our guide to quantum computing breakthroughs, follow crews returning to deep space in Artemis 2's record lunar flyby, or browse the rest of our tech coverage.