Space-based data centers

Scroll sideways for more →

Space-based data centers are data centers launched into Earth orbit (or, in some proposals, placed on the Moon), where satellites carrying AI accelerators draw power from solar arrays and reject heat by radiating it into space. The idea moved from studies to flight hardware between 2024 and 2026: the startup Starcloud put the first data-center-class GPU, an NVIDIA H100, into orbit in November 2025,^[1][2] China launched the first 12 satellites of a planned 2,800-satellite computing constellation in May 2025,^[25] and by mid-2026 SpaceX, Blue Origin, and Starcloud had filed U.S. Federal Communications Commission applications covering, respectively, up to one million, 51,600, and 88,000 orbital data center satellites.^[11][17][7] Proponents including Elon Musk, Jeff Bezos, and NVIDIA CEO Jensen Huang argue that near-constant orbital sunlight can escape the electricity, land, water, and permitting constraints closing in on terrestrial AI infrastructure; critics, including OpenAI CEO Sam Altman and several independent engineering analyses, counter that heat rejection, launch cost, radiation, and the impossibility of hardware maintenance make orbital compute uneconomical for at least the rest of the decade.^[34][30][32]

Rationale

The case for computing in orbit rests on power. In a dawn-dusk sun-synchronous low Earth orbit, a satellite rides the day-night terminator and sees the Sun almost continuously; Google's Project Suncatcher research estimates that a solar panel in such an orbit can be up to 8 times more productive than the same panel on Earth, with no batteries needed to cover nights or bad weather.^[9] Bezos made the same argument at Italian Tech Week in October 2025: "We're going to start building these giant gigawatt data centers in space. So, these giant training clusters, those will be better built in space, because we have solar power there, 24/7."^[16]

This matters because terrestrial AI buildouts increasingly collide with grid limits. The International Energy Agency estimates global data-center electricity consumption reached about 485 TWh in 2025 and, in its central projection, roughly doubles to 950 TWh by 2030,^[36] and interconnection queues, substation equipment, land, cooling water, and local opposition have all become gating items for new gigawatt-scale campuses. Starcloud CEO Philip Johnston framed the pitch in March 2026: "The AI revolution is colliding with the physical limits of our terrestrial energy grid... By moving AI compute to space, we unlock access to unlimited solar power and completely remove the energy bottleneck."^[5] Former Google CEO Eric Schmidt gave the same rationale for taking control of the rocket company Relativity Space in 2025, citing projections that U.S. AI data centers could need tens of additional gigawatts by 2030; asked directly whether the acquisition was about launching data centers, he answered "Yes."^[29]

The second pillar is the launch-cost curve. Falcon 9 rideshare pricing sits between $1,400 and $1,800 per kilogram; nearly every published feasibility case depends on fully reusable heavy lifters, chiefly SpaceX's Starship, driving that figure toward a few hundred dollars per kilogram or below.^[32] Google's Suncatcher paper puts the threshold for rough parity with terrestrial data-center energy costs at less than $200/kg, a level it argues launch-price trends could reach by the mid-2030s.^[9] Vendors also point to cooling: satellites reject heat to the vacuum by radiation alone, consuming no water and, per SpaceX, cutting cooling power overhead "by an order of magnitude" versus chillers and cooling towers.^[12] Whether that trade is actually favorable is the central point of dispute with skeptics (see below).

Technical challenges

Heat rejection

In vacuum there is no convection, so all waste heat must leave by radiation, governed by the Stefan-Boltzmann law. ABI Research analyst Andrew Cavalier calculated in IEEE Spectrum that a single 700 W H100 needs about 1.4 square meters of radiator surface to hold a 60 degree Celsius operating temperature (about 1 square meter if the chips are allowed to run at 85 degrees), so a 40 kW rack of 32 GPUs needs an 80-square-meter radiator, and a 100 MW facility needs at least 2,500 of them.^[30] Radiator coatings also degrade under ultraviolet exposure and atomic oxygen; Cavalier estimates the required area grows about 40 percent over a five-year mission.^[30] For scale, the International Space Station's external thermal control system rejects only about 70 kW, using radiator assemblies covering hundreds of square meters that SemiAnalysis prices at $340 million to $500 million.^[32] Because a square meter of solar array in orbit yields around 400 W while a square meter of radiator sheds around 450 W, Cavalier notes that "every square meter of power generation now demands approximately another square meter of cooling."^[30] Whether radiators sink the economics is contested: Forethought's May 2026 analysis finds radiator hardware is only 2 to 5 percent of total system cost in its median scenarios (assuming modern deployable radiators at 163 to 346 W/kg rather than ISS-era designs), but concedes that under worst-case radiator performance and $250/kg launch, cooling would add more than $8 billion per gigawatt, versus roughly $3 billion for terrestrial cooling.^[33]

Radiation

Commercial accelerators built on leading-edge process nodes were not designed for the space radiation environment, where energetic particles cause single-event upsets, latch-up, and cumulative total-ionizing-dose damage. Traditional radiation-hardened processors give up orders of magnitude in performance; a widely shared critique by a former NASA space-electronics engineer notes that hardened parts perform like a 20-year-old PowerPC.^[31] The counter-evidence so far is encouraging for the boosters: Google put its Trillium (TPU v6e) chips in a proton beam and found no hard failures up to 15 krad(Si), with the high-bandwidth memory subsystem, the most sensitive component, showing irregularities only beyond a cumulative 2 krad(Si), nearly three times the 750 rad(Si) expected (behind shielding) over a five-year mission.^[9] Starcloud-1's H100 has operated in orbit since November 2025,^[3] though a 325 km orbit below the inner Van Allen belt is a far gentler environment than a multi-year mission at 600 km or above, and error rates at fleet scale remain undemonstrated. SemiAnalysis assumes orbital operators must carry about 20 percent spare GPU capacity, versus about 5 percent on the ground.^[32]

Bandwidth and latency

Training clusters need enormous internal bandwidth, and users need a fat, low-latency pipe to Earth. Inside a cluster, Google's answer is free-space optics: its bench demonstration reached 800 Gbps each way (1.6 Tbps aggregate) over a single transceiver pair, and its design flies satellites in tight formation, with next-nearest-neighbor satellites 100 to 200 meters apart within a cluster of about 1 km radius, to keep received optical power high; the published design models clusters of 81 satellites.^[9][10] SpaceX's argument rests on its existing laser mesh: Musk wrote in October 2025 that "simply scaling up Starlink V3 satellites, which have high speed laser links would work," with V3 designed for 1 Tbps-class throughput.^[15] Skeptics point at the ground segment: a LEO satellite passes over a given ground station for only minutes at a time, and SemiAnalysis estimates traffic routed through intersatellite-link meshes accumulates 30 to 80 milliseconds of one-way delay, eroding any claimed latency advantage for interactive workloads.^[32]

Servicing, lifetime, and debris

There is no technician in orbit. A failed GPU, power supply, or pump stays failed, so operators must overprovision hardware and launch replacements; Forethought estimates that losing about 9 percent of compute per year with no repair forces roughly a 38 percent overbuy of launch and non-chip hardware across a data center's life.^[33] Hardware also depreciates against a moving frontier: SemiAnalysis assumes a five-year useful life in orbit versus 15 years for a terrestrial facility shell, and orbital hardware cannot be upgraded rack by rack.^[32] Finally, the filings themselves have raised debris alarms. NASA formally objected to Blue Origin's 51,600-satellite Project Sunrise application in May 2026, telling the FCC the constellation "presents significant safety and sustainability concerns" and requesting sustained technical engagement; DarkSky International objected over night-sky brightness, and the Center for Space Environmentalism called for a full environmental impact statement before approval.^[18] Similar proceedings await SpaceX's million-satellite application, which asked the FCC to waive its standard six-year deployment milestones.^[11]

Programs and players

Scroll sideways for more →

Starcloud

Starcloud, a Y Combinator-backed startup founded in 2024 in Redmond, Washington, flew the first modern AI accelerator to orbit. Starcloud-1, a 60 kg satellite built on an Astro Digital bus, launched on a SpaceX Falcon 9 on November 2, 2025 into a 325 km orbit carrying a single NVIDIA H100, which NVIDIA described as delivering 100 times more GPU compute than anything previously flown.^[1][2][3] In December 2025 the company ran Gemma and a version of Gemini in orbit and trained Andrej Karpathy's nanoGPT model, the first LLM training run in space.^[4] On March 30, 2026 Starcloud announced a $170 million Series A at a $1.1 billion valuation, led by Benchmark and EQT Ventures, bringing total funding to about $200 million; its FCC application for up to 88,000 satellites in 600-850 km sun-synchronous orbits was accepted for filing on March 13, 2026.^[5][6][7] Starcloud-2, slated as of May 2026 for a Falcon 9 launch in January 2027 (an October 2026 target had been reported in December 2025),^[39] is a far larger spacecraft with roughly 100 times Starcloud-1's power generation, carrying an NVIDIA Blackwell GPU, AWS server hardware, bitcoin-mining ASICs, and what the company calls the largest deployable radiator flown on a commercial satellite; Crusoe is an announced early customer.^[6][7] Johnston has been explicit that the roadmap depends on cheap heavy lift: "we do need Starship, or a much lower launch cost vehicle."^[6] Starcloud's long-term concept, described on NVIDIA's blog, is a 5 GW orbital data center with solar and cooling panels roughly 4 kilometers on a side.^[1]

SpaceX Starmind

SpaceX runs the most aggressive program by filed scale. It applied to the FCC on January 30, 2026 for an "Orbital Data Center System" of up to one million solar-powered satellites at 500-2,000 km, days before announcing its acquisition of xAI with orbital data centers as a stated rationale.^[11] Musk unveiled the AI1 satellite design in June 2026 (a roughly 70-meter-wingspan spacecraft whose official page now lists up to 250 kW peak power, deployable liquid radiators, and laser links into the Starlink constellation), named the constellation Starmind, and on August 4, 2026 announced a partnership under which Starmind satellites will carry NVIDIA Rubin GPUs and Vera CPUs from the Vera Rubin platform.^[12][13] On the same day's earnings call Musk said SpaceX is "exclusive to Nvidia" and expects to begin launching Starmind satellites in 2027, with mass production planned at an 11-million-square-foot "Gigasat Factory" in Bastrop, Texas,^[42] and chips eventually supplied by the TeraFab joint project with Tesla.^[14][43] All Starmind capacity and schedule figures are company plans; nothing has launched. The full program history is covered at SpaceX Starmind.

Google Project Suncatcher

Google announced Project Suncatcher, "a research moonshot," on November 4, 2025, alongside a paper titled "Towards a future space-based, highly scalable AI infrastructure system design."^[8][9] The design flies TPU-equipped satellites in tight formations in a 650 km dawn-dusk sun-synchronous orbit, connected by free-space optical links, and treats the whole formation as one distributed machine-learning cluster. Unusually for the field, Google published its caveats up front: the paper concludes the concept is "not precluded by fundamental physics or insurmountable economic barriers" but names thermal management, ground bandwidth, and on-orbit reliability as open engineering problems, and conditions economics on launch prices below $200/kg by the mid-2030s.^[9] Planet Labs will build and operate two prototype satellites carrying TPUs, targeted for launch by early 2027.^[10]

Blue Origin

Bezos spent late 2025 talking up gigawatt orbital data centers on a 10-to-20-year horizon.^[16] Blue Origin's concrete moves came in early 2026: in January it announced TeraWave, a 5,408-satellite laser-linked communications network aimed at data-center, enterprise, and government traffic, and on March 19 it filed the Project Sunrise FCC application for up to 51,600 compute satellites in 500-1,800 km sun-synchronous orbits, with TeraWave as the backbone.^[17] NASA's objection to the filing in May 2026 was the first formal pushback by a U.S. government agency against an orbital data center constellation.^[18] Bezos himself has stayed conservative on timing, saying in May 2026 that the two-to-three-year timelines circulating in the industry are "probably a little ambitious" given energy, chip, and launch costs.^[19]

Axiom Space, Aetherflux, and Lonestar

Axiom Space is building small, service-oriented "orbital data center" nodes rather than gigawatt AI farms: processing capacity in orbit that other satellites can use for storage, AI/ML inference, and secure government workloads. After testing its AxDCU-1 unit on the International Space Station in 2025, Axiom's first two free-flying ODC nodes launched on January 11, 2026 aboard Kepler Communications optical-relay satellites, each carrying multi-GPU compute modules and 2.5 Gbps optical terminals; the company talks about scaling from kilowatts to megawatts.^[20][21] This is closer to edge computing for spacecraft than to a terrestrial data center replacement.

Aetherflux, the space solar power startup founded by Robinhood co-founder Baiju Bhatt, entered the orbital data center race in December 2025 with plans for a first "Galactic Brain" compute node in low Earth orbit, targeted to be operational in the first quarter of 2027, alongside its original mission of beaming solar power to the ground (a first power-beaming demonstration satellite is planned for 2026).^[22] NVIDIA named Aetherflux among the users of the space-computing platforms it announced at GTC in March 2026, alongside Axiom Space, Kepler, Planet Labs, Sophia Space, and Starcloud, where Huang declared that "space computing, the final frontier, has arrived."^[35]

Lonestar Data Holdings pursues data storage rather than AI compute, and targets the Moon. Its Freedom payload, a solid-state storage device marketed as the first lunar data center and aimed at resilience and disaster-recovery use cases, launched aboard Intuitive Machines' IM-2 Athena lander on February 26, 2025; Lonestar said the payload completed its commercial operational tests during transit and remained healthy after entering lunar orbit.^[23][44] Athena touched down on March 6, 2025 near the lunar south pole but came to rest on its side inside a crater and depleted its batteries within a day, ending surface operations prematurely.^[24]

China

The most satellites actually computing in orbit today are Chinese. On May 14, 2025 a Long March 2D launched the first 12 satellites of the Three-Body Computing Constellation, a project of Chengdu-based ADA Space (Guoxing Aerospace) and Zhejiang Lab. Each satellite processes up to 744 trillion operations per second, and the batch is linked by 100 Gbps laser links, giving a combined 5 POPS and 30 TB of storage; the full "Star-Compute" plan calls for 2,800 satellites (2,400 for inference, 400 for training) totaling 1,000 POPS.^[25][40][41] ADA Space executive vice president Zhao Hongjie said in April 2026 that the second and third satellite groups were in production and scheduled for orbital deployment in 2026.^[40] Separately, SatNews reported in February 2026 that the state-owned China Aerospace Science and Industry Corporation (CASIC) had completed in-orbit testing of a three-satellite computing testbed it plans to scale to a 32-satellite computing grid by 2028.^[26] The program is frequently framed in Chinese state media as a bid for first-mover advantage in space computing infrastructure.^[25]

Europe and Japan

Europe's activity has been studies rather than filings. The ASCEND feasibility study (Advanced Space Cloud for European Net zero emission and Data sovereignty), coordinated by Thales Alenia Space for the European Commission under Horizon Europe, published results on June 27, 2024. It concluded that space data centers could reduce digital-sector emissions only if a launcher ten times less emissive over its lifecycle is developed, proposed modular infrastructure assembled in orbit by robotic systems, and sketched a target of one gigawatt of European capacity in orbit before 2050.^[27] In Japan, NTT and SKY Perfect JSAT founded Space Compass in 2022 around a "space integrated computing network" concept that would process satellite data in orbit; its near-term business is geostationary optical data relay and high-altitude platform connectivity, and as of August 2026 the venture's public announcements center on relay services rather than an operating orbital compute satellite.^[28]

Economics

Every serious analysis reduces the question to dollars per kilogram and hardware lifetime. The numbers below come from named models and should be read as estimates with strong assumptions.

Scroll sideways for more →

The disagreements trace to a few parameters. Launch: SemiAnalysis notes Starship's target of about $250/kg against Falcon 9's current $1,400-1,800/kg, and launch is the single most sensitive input in every model.^[32] Lifetime and maintenance: a 5-year orbital life versus 15 years terrestrial, plus spare-capacity overbuy of around 20 percent, drives SemiAnalysis's estimate that monthly non-chip data center capex runs about 18 times higher in space.^[32] Radiators: Forethought's median scenarios make cooling hardware a minor cost line (2 to 5 percent), while worst-case radiator assumptions at $250/kg add over $8 billion per gigawatt.^[33] Vendor claims sit well outside these ranges: Starcloud projects 10x lower energy costs and 10x lifecycle CO2 savings versus terrestrial facilities,^[1] and the ASCEND study projected returns of several billion euros by 2050,^[27] but both are advocacy positions from parties selling the concept.

A separate line of argument, made by SemiAnalysis, is that even a launch-cost miracle would not remove the binding constraint on AI buildout, which it identifies as semiconductor supply: in its projections AI demand consumes about 86 percent of TSMC's N3-class capacity and 70 percent of DRAM wafer capacity by 2027, so orbit adds power capacity but not chips.^[32]

Criticism

The concept has drawn unusually pointed criticism for something attracting billions in filings and funding. Altman said in February 2026: "I honestly think the idea with the current landscape of putting data centers in space is ridiculous," citing launch costs relative to terrestrial power costs and the impossibility of fixing broken GPUs, and adding that orbital data centers are "not something that is going to matter at scale this decade."^[34] Musk's rebuttal to thermal critiques has been that it is "safe to say SpaceX knows how to do heat rejection in space," pointing at the more than 10,000 Starlink satellites it operates,^[38] though a Starlink satellite dissipates orders of magnitude less heat than a single AI rack.

The engineering critiques cluster around thermodynamics. Cavalier's IEEE Spectrum analysis argues the industry's "free cooling" framing inverts reality, because radiative-only heat rejection is the binding constraint: "the constraint is never the silicon. It's the thermodynamics."^[30] A widely circulated November 2025 essay on the Taranis blog, written by a self-described former NASA space-electronics engineer who later spent a decade at Google, works the same numbers from the ISS baseline (a station-sized 2,500-square-meter solar array yields power for only a couple hundred GPUs) and calls the concept "catastrophically bad" as a replacement for terrestrial facilities, while conceding small demonstrations are feasible.^[31] SemiAnalysis adds that several standard talking points are simply wrong as stated: generic LEO satellites see sunlight only about 60 percent of the time (dawn-dusk sun-synchronous orbits fix this but are a narrow, crowded orbital slot), and even those orbits can suffer roughly 35 minutes of eclipse per day.^[32]

Institutional objections center on the space environment. NASA's May 2026 filing against Project Sunrise, and parallel objections from DarkSky International and the Center for Space Environmentalism, mark the start of a regulatory fight over whether tens of thousands to a million additional large satellites can be operated safely, with collision-cascade risk and interference with ground-based astronomy the recurring themes; SpaceX's application meanwhile asks the FCC to waive the standard milestone that would require half the constellation to be in orbit within six years.^[18][11] Even sympathetic observers note the gap between filings and demonstrated hardware: as of August 2026, the entire industry's flown inventory amounts to one H100 smallsat, two multi-GPU relay nodes, a dozen Chinese inference satellites, an ISS testbed, and a lunar storage payload on a tipped-over lander. Bezos, whose company filed one of the largest applications, put the sober version on the record himself: timelines of two to three years are "probably a little ambitious."^[19]

See also

• SpaceX Starmind
• Data center
• AI infrastructure
• AI energy consumption
• NVIDIA Vera Rubin
• Edge computing
• TeraFab

References

1. ^NVIDIA Blog. "How Starcloud Is Bringing Data Centers to Outer Space." October 15, 2025. blogs.nvidia.com/...starcloud
2. ^Gunter's Space Page. "Starcloud 1 (Lumen 1)." Accessed August 12, 2026. space.skyrocket.de/...starcloud-1
3. ^Data Center Frontier. "Starcloud Launches Orbital AI Data Center With NVIDIA H100 GPU." November 2025. datacenterfrontier.com/...ter-with-nvidia-h100-gpu
4. ^CNBC. "Nvidia-backed Starcloud trains first AI model in space." December 10, 2025. cnbc.com/...ai-model-in-space-orbital-data-centers
5. ^Via Satellite. "Starcloud Raises $170M to Fund Orbital Data Center Plans." March 30, 2026. satellitetoday.com/...nd-orbital-data-center-plans
6. ^Payload. "Starcloud Raises $170M Series A at $1.1B Valuation." March 30, 2026. payloadspace.com/...70m-series-a-at-1-1b-valuation
7. ^SpaceNews. "Starcloud files plans for 88,000-satellite constellation." March 2026. spacenews.com/...for-88000-satellite-constellation
8. ^Google. "Meet Project Suncatcher, a research moonshot to scale machine learning compute in space." The Keyword, November 4, 2025. blog.google/...google-project-suncatcher
9. ^Google Research. "Exploring a space-based, scalable AI infrastructure system design." November 4, 2025. research.google/...ai-infrastructure-system-design
10. ^Planet Labs PBC. "Planet to Build and Operate Advanced Space Platform for Google's Project Suncatcher Moonshot." November 2025. planet.com/...form-for-project-suncatcher-moonshot
11. ^SpaceNews. "SpaceX files plans for million-satellite orbital data center constellation." January 2026. spacenews.com/...orbital-data-center-constellation
12. ^SpaceX. "Starmind." Official SpaceX website, accessed August 2026. spacex.com/...starmind
13. ^SpaceX (@SpaceX) on X. "SpaceX is partnering with @Nvidia to design the Starmind AI1 satellite compute payload." August 4, 2026. x.com/...2084723854534951218
14. ^Daniel Howley. "Nvidia stock rises after Musk says SpaceX will exclusively use company's chips." Yahoo Finance, August 5, 2026. uk.finance.yahoo.com/...e-companys-chips-211803293
15. ^Teslarati. "Starlink V3 satellites could enable SpaceX's orbital computing plans: Musk." October 31, 2025. teslarati.com/...acex-orbital-computing-plans-musk
16. ^Alan Boyle. "Jeff Bezos says orbital data centers will be the 'next step' for space ventures that make Earth better." GeekWire, October 6, 2025. geekwire.com/...zos-orbital-data-centers-next-step
17. ^GeekWire. "51,600 more satellites? Blue Origin adds another twist to the data center space race with Project Sunrise." March 2026. geekwire.com/...-center-space-race-project-sunrise
18. ^SatNews. "NASA Objects to Blue Origin's 'Project Sunrise'." May 5, 2026. satnews.com/...cts-to-blue-origins-project-sunrise
19. ^CNBC. "Bezos says 2-3 year timeline for space data centers is a 'little ambitious'." May 20, 2026. cnbc.com/...bezos-blue-origin-space-data-centers
20. ^Axiom Space. "Axiom Space to Launch Orbital Data Center Nodes to Support National Security, Commercial, International Customers." April 7, 2025. axiomspace.com/...mmercial-international-customers
21. ^Introl. "First Orbital Data Center Nodes Reach Space." January 2026. introl.com/...omputing-infrastructure-january-2026
22. ^SpaceNews. "Space-based solar power startup Aetherflux enters orbital data center race." December 2025. spacenews.com/...x-enters-orbital-data-center-race (see also Data Center Dynamics, "Aetherflux orbital data center to be operational by Q1 2027," December 2025, datacenterdynamics.com/...e-operational-by-q1-2027)
23. ^PR Newswire. "Lonestar's Data Center is ready for the Moon: Intuitive Machines and Lonestar Complete Final Installation and Testing of Lonestar's Data Center Payload." February 2025. prnewswire.com/...rs-data-center-payload-302372021
24. ^Spaceflight Now. "Intuitive Machines' IM-2 Moon mission ends with lander on its side." March 7, 2025. spaceflightnow.com/...nder-on-its-side-on-the-moon
25. ^Andrew Jones. "China launches first of 2,800 satellites for AI space computing constellation." SpaceNews, May 2025. spacenews.com/...-ai-space-computing-constellation
26. ^SatNews. "China Completes In-Orbit Testing of 'Three-Body' AI Computing Constellation." February 16, 2026. satnews.com/...ree-body-ai-computing-constellation
27. ^Thales Alenia Space. "Thales Alenia Space reveals results of ASCEND feasibility study on space data centers." June 27, 2024. thalesaleniaspace.com/...tudy-space-data-centers-0
28. ^NTT. "NTT and SKY Perfect JSAT Agree to Establish Space Compass Corporation." April 26, 2022. group.ntt/...220426a
29. ^TechSpot. "Former Google CEO Eric Schmidt wants to put data centers in space." May 2025. techspot.com/...le-ceo-eric-schmidt-wants-put-data
30. ^Andrew Cavalier. "Why Thermodynamics Rules Future Orbital Data Centers." IEEE Spectrum, June 11, 2026. spectrum.ieee.org/orbital-data-centers-heat
31. ^Taranis. "Datacenters in space are a terrible, horrible, no good idea." taranis.ie, November 26, 2025. taranis.ie/...are-a-terrible-horrible-no-good-idea
32. ^Daniel Nishball, Pranav Myana, Ellie Holbrook, et al. "To Boldly Go: The Case for Space Datacenters." SemiAnalysis, June 3, 2026. newsletter.semianalysis.com/...r-space-datacenters
33. ^Avi Parrack and Fin Moorhouse. "Will We Really Put Data Centers in Space?" Forethought, May 22, 2026. forethought.org/...eally-put-data-centers-in-space
34. ^Tom's Hardware. "Sam Altman fires back at Elon Musk's proposal for space-based data centers, says orbiting data centers 'ridiculous' for now." February 2026. tomshardware.com/...s-and-cost-as-primary-limiters
35. ^NVIDIA Newsroom. "NVIDIA Launches Space Computing, Rocketing AI Into Orbit." March 16, 2026. nvidianews.nvidia.com/...space-computing
36. ^International Energy Agency. "Key Questions on Energy and AI." 2026. iea.org/...key-questions-on-energy-and-ai
37. ^Investing.com. "Deutsche Bank sees SpaceX orbital data centers matching ground costs." 2026. investing.com/...atching-ground-costs-93CH-4790612
38. ^Luke James. "Elon Musk's first-gen orbital data center craft spans wider than a Boeing 747 and runs an interchangeable chip payload." Tom's Hardware, June 9, 2026. tomshardware.com/...ails-its-ai1-compute-satellite
39. ^Jason Rainbow. "Starcloud orders Starlink lasers for orbital data center network." SpaceNews, May 26, 2026. spacenews.com/...s-for-orbital-data-center-network
40. ^State Council Information Office of China. "From ground to orbit: China eyes computing in space." April 27, 2026. english.scio.gov.cn/...content_118464689
41. ^Ben Turner. "China is building a constellation of AI supercomputers in space, and just launched the first pieces." Live Science, May 23, 2025. livescience.com/...-just-launched-the-first-pieces
42. ^Tom's Hardware. "SpaceX unveils 11-million-square-foot Gigasat factory, a new manufacturing facility for space-based data centers." June 9, 2026. tomshardware.com/...-late-2027-from-its-satellites
43. ^Luke James. "Elon Musk unveils $20 billion 'TeraFab' chip project to make chips, memory, and package processors all under one roof." Tom's Hardware, March 22, 2026. tomshardware.com/...0-billion-terafab-chip-project
44. ^Lonestar Data Holdings. "Lunar Data Center Achieves First Success En Route To The Moon." PR Newswire, March 5, 2025. prnewswire.com/...s-en-route-to-the-moon-302392544