Orbital Data Center Market Landscape 2026

As an Amazon Associate we earn from qualifying purchases.

Key Takeaways

• Twenty-one orbital data center programs have flown hardware in space across 53 tracked global initiatives.
• Kepler Communications, BUPT Tiansuan, ADA Space, and Zhejiang Lab operate multi-satellite orbital networks.
• Five United States operators have filed federal applications for up to 1,259,600 orbital compute satellites.

How the Global Orbital Data Center Landscape Splits Across Seven Readiness Stages

As of November 1, 2026, 21 of the 53 tracked programs building orbital data centers worldwide have flown their own computing or data-storage hardware in space, accompanied by 10 additional software or component-supplier missions and six corporate statements of intent. An orbital data center, also referred to as a space data center or space-based computing node, is a spacecraft or networked constellation of satellites that executes data-center workloads in orbit using solar power and rejects waste heat into deep space through radiative panels. What began between 2017 and 2024 as isolated processor experiments aboard the International Space Station (ISS) and small CubeSats shifted during 2025 and 2026 into dedicated free-flying GPU satellites, multi-satellite optical computing networks, hundreds of millions of dollars in venture financing, and regulatory filings proposing more than one million spacecraft.

The rapid acceleration of orbital data center programs stems directly from physical and regulatory bottlenecks constraining terrestrial artificial intelligence (AI) infrastructure. According to Goldman Sachs power projections, United States data center electricity demand is projected to more than double from 31 gigawatts in 2025 to 66 gigawatts by 2027, whereas an April 2026 RAND analysis estimates only 82 gigawatts of net new utility generation capacity will come online by 2030. In primary terrestrial markets such as Northern Virginia, Silicon Valley, and Frankfurt, utility grid interconnection queues stretch from seven to 12 years, and the first quarter of 2026 recorded roughly 75 terrestrial projects worth $130 billion delayed or blocked by power shortages and local opposition. As examined in coverage of why data centers inspire bipartisan pushback and empirical studies showing data centers yield modest job growth relative to their electricity and water demands, terrestrial hyperscalers face mounting land-use moratoria, transformer shortages, and water-basin restrictions.

Moving computational workloads above the atmosphere offers three foundational physical advantages over ground facilities. First, a spacecraft operating in a dawn-dusk sun-synchronous orbit (SSO) receives nearly continuous, unfiltered solar irradiance of approximately 1,361 watts per square meter, achieving capacity factors up to 99 percent and generating five to nine times more annual energy per square meter than terrestrial solar arrays. Second, spacecraft reject heat directly into the 2.7 Kelvin cosmic background through infrared thermal radiation, consuming zero evaporative cooling water. Third, orbital deployments bypass municipal zoning approvals, terrestrial land acquisition, and multi-year utility interconnection queues. As detailed in foundational analyses of space-based data center market dynamics and orbital digital infrastructure architectures, those advantages have attracted venture-backed startups, cloud hyperscalers, reusable launch providers, and state-backed research laboratories across 15 countries.

Evaluating which orbital data center programs are operational, which are under active hardware development, and which remain unflown paper concepts requires a rigorous classification framework. Industry tracking compiled across 327 primary sources by the SpaceNexus orbital data centers tracker divides global activity into seven distinct readiness stages based strictly on verified hardware milestones rather than executive announcements. The table below outlines how the 53 active programs and six statements of intent (59 total tracked entries) distribute across those seven readiness tiers as of November 2026.

Separating demonstrated orbital hardware from unflown regulatory ambition clarifies the current state of the industry. Although 21 programs have operated their own computing or storage payloads beyond Earth, no operator yet runs a megawatt-scale hyperscale data center in orbit. The only commercial compute service currently sold to external customers runs on edge-class modules drawing tens of watts, and all three multi-satellite pathfinder constellations operate in China as state-backed or corporate testbeds. At the same time, two American startups, Starcloud and Cowboy Space, hold 94 percent of the $803.3 million in confirmed dedicated startup equity, and five United States companies have filed applications with the Federal Communications Commission (FCC) seeking authorization for up to 1,259,600 orbital compute satellites.

Operational Networks and Pathfinder Constellations Running Workloads in Orbit

Only four orbital data center programs worldwide have reached Stages 6 and 7 by operating multiple networked computing satellites in low Earth orbit (LEO). One Canadian operator sells distributed edge-computing capacity commercially, and three Chinese programs operate multi-satellite pathfinder constellations validating optical crosslinks, onboard artificial intelligence inference, and spacecraft thermal loops.

Kepler Communications and Commercial Optical Relay Computing

Toronto-based Kepler Communications holds the distinction of operating the only Stage 7 commercial orbital compute service as of November 2026. On January 11, 2026, Kepler launched 10 300-kilogram-class optical relay satellites aboard a SpaceX Falcon 9 rocket from Vandenberg Space Force Base in California. Each spacecraft in this Tranche 1 deployment carries at least four optical inter-satellite communication terminals built by Airbus subsidiary Tesat, terabytes of solid-state data storage, hosted-payload interfaces, and multi-GPU computing modules. Backing the constellation is more than $233 million in cumulative capital across seven funding rounds, including a $92 million Series C led by IA Ventures.

On March 16, 2026, Kepler commissioned its distributed on-orbit computing layer across all 10 spacecraft, integrating 40 NVIDIA Jetson Orin modules into an internet-protocol mesh network. Operating over 2.5 gigabits-per-second (Gbps) optical crosslinks compatible with United States Space Development Agency (SDA) standards, the network routes data dynamically between satellites and shifts processing workloads across neighboring nodes if an individual unit becomes unavailable. Kepler opened the platform to commercial customers on April 13, 2026, reporting 18 compute and relay customers, and plans a second constellation tranche in 2028 featuring 100 Gbps optical links across multiple launches, including a dedicated Rocket Lab flight. As explored in analyses of orbital data center hub architectures, Kepler’s mesh demonstrates how relay constellations can double as shared in-space processing hubs for client sensing satellites.

China’s Tiansuan, ADA Space, and Three-Body Computing Constellations

Whereas Western startups focused on single-satellite GPU demonstrators, Chinese research institutes and commercial satellite builders deployed the world’s largest fleets of networked computing spacecraft. In late August and September 2026, the Tiansuan Constellation, initiated by the Beijing University of Posts and Telecommunications (BUPT), began regular in-orbit cloud computing services. Having launched its first primary satellite (BUPT-1) in January 2023 and its initial Phase 2 spacecraft (BUPT-2 and BUPT-3) aboard a LandSpace Zhuque-2E rocket on May 17, 2025, the Tiansuan network reached 16 operational LEO satellites connected to seven terrestrial ground stations by September 2026. The platform serves more than 60 universities, research institutes, and industrial partners conducting 6G core-network tests, cloud-native satellite orchestration, and onboard autonomous decision-making, with BUPT outlining an undated Phase 2 expansion to 24 satellites and a Phase 3 target of 300 spacecraft.

In parallel, Chengdu-based commercial operator ADA Space (Guoxing Aerospace) and the Hangzhou-based Zhejiang Lab Three-Body Constellation deployed an interconnected space-computing network designed for high-capacity artificial intelligence workloads. On May 14, 2025, a Long March 2D rocket launched from the Jiuquan Satellite Launch Center placed the first 12 computing satellites into sun-synchronous orbit, forming both ADA Space’s Group 01 batch and the inaugural tranche of Zhejiang Lab’s Three-Body constellation. According to state and company disclosures, each of the 12 satellites delivers 744 tera-operations per second (TOPS), providing roughly five peta-operations per second (POPS) of aggregate compute and 30 terabytes of onboard storage across the group.

Both Chinese partners expanded their orbital testing throughout late 2025 and 2026. In November 2025, ADA Space deployed Alibaba’s Qwen3 large language model onto its Group 01 satellites, executing uplinked inference queries and returning answers to Earth in under two minutes. By February 2026, Zhejiang Lab reported establishing a six-satellite laser communication chain and validating 10 artificial intelligence models in orbital testing. On July 24, 2026, Zhejiang Lab added two partner-built spacecraft (Jitianxing A-04 and Xiguang-2 03) aboard a CAS Space Kinetica-1 Y15 rocket to run orbital weather AI models, bringing the Three-Body fleet to 14 satellites and demonstrating a continuous inter-satellite laser link lasting 192 hours, eight minutes, and 45 seconds at 99.99 percent availability over distances up to 1,000 kilometers. On October 14, 2026, ADA Space launched eight additional computing test satellites aboard a Zhuque-2E rocket to evaluate 10-kilowatt liquid-cooling loops and laser routing under load in an October 2026 assessment, advancing toward a long-term target of 1,000 networked satellites by 2030 and 2,800 satellites by 2035.

Dedicated Compute Demonstrators and Hosted Orbital Nodes

Beyond the four multi-satellite networks, 10 programs have reached Stage 5 by flying their own dedicated compute satellites, and 17 programs have reached Stage 4 by operating computing or storage payloads hosted aboard third-party spacecraft, lunar landers, or the International Space Station.

Standalone Compute Satellites Testing Data-Center Silicon in Low Earth Orbit

Redmond, Washington-based Starcloud (formerly Lumen Orbit) leads Western startups in single-payload processor power and dedicated venture capital. On November 2, 2025, Starcloud launched its 60-kilogram Starcloud-1 demonstrator built on an Astro Digital Corvus-Micro+ bus aboard SpaceX’s Bandwagon-4 rideshare mission into a 495-by-503-kilometer orbit. As examined in New Space Economy’s profile of Starcloud’s orbital GPU architecture, the refrigerator-sized spacecraft carried an NVIDIA H100 Tensor Core GPU (a 700-watt data-center accelerator roughly 100 times more powerful than earlier space-flown processors) cooled via solid-state thermal conduction paths linked to deployable infrared radiators. In December 2025, Starcloud reported that Starcloud-1 trained Andrej Karpathy’s nanoGPT small language model on the complete works of Shakespeare, ran inference queries on Google’s Gemma open-weights model, and processed synthetic aperture radar imagery from Capella Space.

Investor backing for Starcloud surged in 2026: Starcloud raised a $170 million Series A round in March 2026 at a $1.1 billion valuation, followed by a $250 million financing extension in August 2026 led by Manhattan West with participation from NVIDIA and Cisco at a $2.3 billion valuation, bringing confirmed equity to $433.4 million ($450 million total reported).

Hyperscale cloud providers entered orbital hardware testing on October 1, 2026, when Google Project Suncatcher and satellite partner Planet Labs launched their first prototype spacecraft aboard SpaceX’s Transporter-18 rideshare mission. Prior to launch, Google engineers subjected commercial Trillium (v6e) Tensor Processing Units (TPUs) to 67-mega-electron-volt proton-beam irradiation at the University of California, Davis, Crocker Nuclear Laboratory, confirming in June 2026 that shielded TPUs survived a 15-krad(Si) dose (roughly 20 times the expected 750-rad(Si) five-year LEO mission dose) without hard hardware failures.

Carrying four Trillium TPUs and roughly one kilowatt of solar power, the Transporter-18 prototype operates in 15-to-20-minute compute bursts. As detailed in coverage of Project Suncatcher’s orbital AI response and the hurdles facing Google’s space computing, Google reported on October 30, 2026, that the onboard TPU payload executed its Gemma 4 model in orbit and returned a generated haiku and payload telemetry responses during a pass over Svalbard. Planet separately operates NVIDIA Jetson Orin edge processors aboard its Pelican Earth-observation satellites, demonstrating onboard aircraft detection with 80 percent accuracy on Pelican-4 in March 2026.

San Francisco-based startup Aethero has compiled the strongest record of repeatable edge-computing flights among leanly funded Western startups, completing two orbital missions on $8.4 million in disclosed capital. Following its August 2024 Deimos mission (100 TOPS) on Transporter-11, Aethero launched its Phobos spacecraft on SpaceX’s Transporter-16 mission on March 30, 2026, operating a radiation-tolerant NVIDIA Jetson Orin NX module at up to 157 TOPS for multiple commercial software customers.

Six additional international operators have flown their own Stage 5 computing satellites across Asia, the Middle East, and Europe:

• Orbital Chenguang (China): Launched Chenguang-1 on July 24, 2026, aboard a CAS Space Kinetica-1 Y15 rocket, claiming 25 petaFLOPS of compute capacity and beginning orbital validation of flexible graphene radiative cooling panels in August 2026, backed by a consortium of 24 to 30 organizations and RMB 57.7 billion in non-binding bank credit intentions.
• Space String (China): Launched its STAR.VISION satellite pair on August 5, 2026, aboard a Jielong-3 Y12 rocket and signed RMB 5 billion credit lines with both the Agricultural Bank of China and the Industrial and Commercial Bank of China in September 2026 toward a planned 1,000-satellite network.
• Supersmartcompute (China): Launched Super Intelligent Computing-1 on September 20, 2026, aboard a Kinetica-1 Y18 rocket from Jiuquan to validate in-orbit intelligent processing.
• Altair (United Arab Emirates and United States): A $1 billion, 50-satellite artificial intelligence constellation program formed by Loft Orbital and Abu Dhabi’s Marlan Space in September 2026; launched Altair-1 (built by Orbitworks with NVIDIA GPUs) on SpaceX’s Transporter-18 on October 1, 2026, acquiring signal and entering commissioning.
• TakeMe2Space (India): Hyderabad-based startup ($5 million raised) that flew its MOI-TD hosted payload on ISRO’s PSLV-C60 POEM-4 fourth stage on December 30, 2024, lost MOI-1 during the PSLV-C62 third-stage anomaly on January 12, 2026, and successfully launched MOI-1A on SpaceX’s Transporter-18 on October 1, 2026.
• KP Labs (Poland): Launched the Intuition-1 6U CubeSat in November 2023 carrying its three-TOPS Leopard Data Processing Unit (DPU) for autonomous hyperspectral image segmentation in orbit.

Hosted Payloads, International Space Station Testbeds, and Cislunar Storage

Seventeen programs sit at Stage 4, having demonstrated computing or data-storage payloads hosted aboard space stations, orbital transfer vehicles, lunar landers, or partner satellites. Houston-based Axiom Space operates the broadest multi-node hosted infrastructure in low Earth orbit. After delivering its AxDCU-1 prototype running Red Hat Device Edge and MicroShift Kubernetes to the ISS on SpaceX CRS-33 on August 24, 2025, Axiom launched two free-flying Orbital Data Center (ODC / ARC) nodes hosted aboard Kepler’s Tranche 1 satellites on January 11, 2026. On October 6, 2026, Axiom confirmed active 2.5 Gbps optical node-to-node and node-to-ground links between the two free-flying units and completed a ground demonstration migrating a post-quantum-encrypted workload.

Florida-based Lonestar Data Holdings ($15.5 million reported funding across seed rounds) focuses on sovereign data storage, backup, and disaster recovery rather than raw AI training. Lonestar ran an Ubuntu virtual-machine test on an ISS server in December 2021, conducted a store-and-return data test from the lunar surface with its Independence payload aboard Intuitive Machines’ IM-1 Odysseus lander in February 2024, and launched its eight-terabyte Freedom storage server (built by Spacebilt with a Microchip RISC-V processor) aboard Intuitive Machines’ IM-2 Athena lunar lander on February 26, 2025. Although Athena tipped onto its side after touching down near the lunar south pole on March 6, 2025, Lonestar completed in-transit data upload, encryption, and edge-processing tests for named customers including the State of Florida, the Isle of Man government, and Valkyrie AI.

Other Stage 4 hardware and software demonstrations include Hewlett Packard Enterprise’s Spaceborne Computer-1 (2017) and Spaceborne Computer-2 (launched to the ISS in 2021 and upgraded in January 2024 with more than 130 terabytes of Kioxia flash storage), Voyager Space and LEOcloud’s Space Edge multi-cloud ISS server (launched September 14, 2025, and updated with Red Hat Enterprise Linux 10.1 in May 2026), Luxembourg’s Edge Aerospace (flown on Transporter-16 in March 2026), and Sweden’s Unibap (SpaceCloud iX5 in 2021 and iX10 in November 2025 aboard Italian carrier D-Orbit’s ION vehicles). Hosted software and hardware-enabler programs at Stage 4 also encompass Amazon Web Services (Snowcone on Axiom Mission 1), Microsoft Azure Space (IMAGIN-e on the ISS), IBM and Red Hat, Booz Allen Hamilton (which partnered with Meta to run a fine-tuned Space Llama 3.2 model on SBC-2 in April 2025), India’s SkyServe (STORM edge software), Japan’s Space Compass (an NTT and SKY Perfect JSAT joint venture that cut ship-detection downlink volume by 98.1 percent on Loft Orbital’s YAM-6 in April 2025), Satlyt ($8 million raised for decentralized satellite cloud software), Astro Digital, Spacebilt, and NVIDIA.

The table below compares the nine leading operational and flown orbital data center operators across Stages 4 through 7 as of November 2026.

Funded Orbital Compute Programs Under Development for 2026 Through 2028

Schedule delays remained a defining feature of the orbital compute sector through late 2026. According to SpaceNexus promise-tracking records, every dated launch target that came due through October 2026 was either delivered five to 24 months late or pushed into 2027. No Western startup launched a new dedicated computing satellite between April and September 2026, shifting the industry’s decisive multi-kilowatt hardware demonstrations into a concentrated 2027 to 2028 deployment window.

Next-Generation Commercial Clusters From Flown Operators

Operators that have already flown single processors or small edge modules are now building multi-kilowatt spacecraft and clustered optical networks for 2027 and 2028:

• Starcloud-2, SC-1L, and Starcloud-3 (United States): Starcloud shifted the launch of Starcloud-2 (comprising two eight-kilowatt rideshare satellites carrying multiple NVIDIA H100 and Blackwell GPUs alongside the first space-deployed AWS Outposts hardware and Starcloud’s custom deployable radiators) from October 2026 into 2027. Through a partnership with Crusoe Energy, developers will rent orbital GPU instances via the Crusoe Cloud interface, managed by Quindar mission-operations software under an October 2026 agreement. On September 30, 2026, Starcloud also booked its SC-1L compute payload aboard Firefly Aerospace‘s Elytra orbital vehicle on Blue Ghost Mission 3 to lunar orbit (targeted no earlier than 2028), and the company is equipping a 100,000-square-foot factory in Woodinville, Washington, to build three-metric-ton, 200-kilowatt Starcloud-3 satellites designed around NVIDIA’s Space-1 Vera Rubin module and SpaceX’s Starship launcher.
• Google Project Suncatcher Two-Satellite Optical Cluster (United States): Following its single-satellite Transporter-18 prototype, Google Research and Planet target a 2027 two-satellite learning mission to test high-bandwidth free-space optical crosslinks in close formation. In laboratory bench tests, Google achieved 800 Gbps in each direction (1.6 terabits per second total) across a single transceiver pair, modeling an eventual sun-synchronous cluster of 81 satellites flying in a one-kilometer formation at 650 kilometers altitude.
• Aethero Titan Cluster (United States): Originally targeted for October 2026, Aethero’s Titan compute-as-a-service satellite is manifested on SpaceX’s Bandwagon-6 rideshare mission in the second quarter of 2027. Built on an EnduroSat FRAME-15 bus, Titan clusters four NxA-ECM (NVIDIA Jetson Thor) modules targeting up to 16,000 teraFLOPS at 3.4 kilowatts peak power, and Aethero signed an August 2026 agreement to purchase space-to-space beamed laser power from Star Catcher.
• Axiom Space Petabyte ISS Node and Lonestar StarVault (United States): Axiom Space, Spacebilt, Phison Electronics, Microchip Technology, and Skyloom are building AxODC Node ISS for delivery to the International Space Station in 2027, combining Microchip’s radiation-hardened PIC64-HPSC high-performance spaceflight processor, more than one petabyte of Phison solid-state storage, and a Skyloom optical terminal. Meanwhile, Lonestar Data Holdings is developing its commercial StarVault sovereign storage service; after host builder Sidus Space adjusted its manifest in September 2026, the first StarVault payload aboard LizzieSat-4 (manifested on SpaceX Bandwagon-6) is scheduled for no earlier than April 2027, alongside a second ordered Sidus payload, a May 2026 NASA Ames Space Act Agreement, and a preliminary $120 million agreement for six cislunar storage spacecraft.

Pre-Launch Startups, Integrated Rockets, and Stage 3 Ground Prototypes

Alongside the flown contenders, several heavily funded startups and hardware specialists are advancing ground prototypes and vertically integrated launch-and-compute systems toward inaugural computing flights between late 2026 and 2028:

• Cowboy Space / Aetherflux (United States): Founded in 2024 by Robinhood co-founder Baiju Bhatt as the space-based solar power startup Aetherflux, San Carlos, California-based Cowboy Space has raised $325 million in confirmed equity ($355 million total reported, including a $50 million Series A and a $275 million Series B). On October 1, 2026, Cowboy Space launched its first spacecraft, Reason-1, on an Apex Aries bus aboard SpaceX’s Transporter-18 mission to demonstrate infrared laser power beaming from orbit to ground receivers, with Reason-2 and a first Galactic Brain orbital compute node targeted for 2027. To overcome launch capacity bottlenecks, Cowboy Space opened a 291,000-square-foot manufacturing plant in Kent, Washington, in September 2026 and unveiled its reusable stainless-steel Big Iron rocket on October 26, 2026. As analyzed in New Space Economy’s review of Cowboy Space’s Big Iron rocket, the 86-meter methalox launcher uses nine Clydesdale engines on a reusable first stage, whereas the 15,000-kilogram upper stage remains in orbit as the data center satellite itself, unfurling a 132-meter solar wingspan and transforming its payload fairing into a thermal radiator to cool five racks of GPUs (one megawatt of compute) starting with a targeted first flight in December 2028.
• Sophia Space (United States): Pasadena, California-based Sophia Space has raised $13.5 million in confirmed equity ($22 million company-stated, including a $10 million seed round in February 2026 and a $7 million SAFE in June 2026) and signed a non-binding $300 million lease-financing letter of intent with Space Leasing International (SLI) in September 2026 for a 10-satellite constellation in 2028 after raising $10 million. Co-founded by Caltech space-solar pioneers, Sophia Space secured U.S. Patent 12,679,564 in July 2026 for its modular TILE passive thermal radiator architecture, plans to run its SOOS orchestration software on Kepler satellites in the fourth quarter of 2026, and booked a 2027 in-orbit TILE hardware demonstration on an Apex Nova satellite bus.
• Orbital (United States): Founded in 2026 by Spin co-founder Euwyn Poon, Orbital raised a $5 million pre-seed round in June 2026 from Andreessen Horowitz (a16z speedrun), Antler, and Basis Set Ventures. The company plans a hosted GPU pathfinder on a Falcon 9 rideshare in 2027 and selected German manufacturer Reflex Aerospace in September 2026 under an exclusive agreement to build Orbital-1, a 250-kilowatt distributed AI inference satellite targeted for 2028.
• OrbitsEdge (United States) and Ramon.Space (Israel): Both companies sit at Stage 3 (Ground Prototype). Florida-based OrbitsEdge developed its hardened SatFrame micro-data center enclosure integrating commercial HPE servers; its Edge1 first orbital demonstration target in early 2026 passed without a confirmed launch as of October 2026. Israel-based Ramon.Space ($26 million Series B), drawing on 50 space missions of radiation-hardened digital signal processor heritage, partnered with Foxconn subsidiary Ingrasys to manufacture space-grade computing blades and targets orbital data center platform prototypes in 2027 and 2028.
• Emerging Global Builders and High-Power Bus Enablers: In India, launch startup Agnikul Cosmos partnered with cloud provider NeevCloud to develop a 300-to-350-kilogram pilot orbital edge data center unit aboard Agnikul’s Agnibaan rocket, and hyperspectral operator Pixxel partnered with Sarvam AI on a sovereign AI Pathfinder satellite. In Canada, PowerBank and Orbit AI (backed by Intellistake) plan five to eight compute and DeFi verification satellites. In China, Comospace Group targets a two-satellite high-power compute demonstration by March 2027, Xingshu Tiansuan targets its Fuxi-1 compute satellite before March 2027, and Space String plans an 11.5-kilowatt G1 verification satellite in the fourth quarter of 2027. Supporting these operators are high-power satellite bus builders K2 Space (which flew its 20-kilowatt Gravitas bus in March 2026, raised a $500 million Series D at a $6.8 billion valuation in July 2026, and pitches its 100-kilowatt Giga bus for 2028 hyperscaler deployments), Muon Space (developing the 20-to-100-kilowatt Condor-Ultra bus with 25 Gbps Starlink lasers and optional NVIDIA Space-1 modules for 2028), Atomic-6 / ODC.space (developing Hot Wing deployable radiators and brokering 1U-to-42U orbital compute capacity from $575,000 per month), and Rocket Lab (developing kilometer-scale silicon solar arrays for orbital compute platforms).

Planned Mega-Constellations, Regulatory Filings, and Institutional Roadmaps

Whereas near-term engineering teams work to cool single-kilowatt and 10-kilowatt payloads in orbit, regulatory filings submitted during the first half of 2026 proposed orbital data center constellations on a scale never previously contemplated in space law. Between January 30, 2026, and June 24, 2026, five United States companies filed applications with the Federal Communications Commission requesting authorization for up to 1,259,600 orbital compute satellites, representing more than 100 times the entire active satellite population orbiting Earth in 2026.

The table below summarizes the five major United States FCC orbital data center constellation filings submitted in 2026.

SpaceX, Blue Origin, and the Million-Satellite Filing Wave

SpaceX triggered the mega-constellation filing wave on January 30, 2026, by requesting FCC authorization for an Orbital Data Center System of up to one million satellites operating in shells between 500 and 2,000 kilometers altitude and linked via optical crosslinks to the Starlink network. Three days later, on February 2, 2026, SpaceX acquired artificial intelligence developer xAI in a transaction valuing the combined company at $1.25 trillion, ahead of confidential initial public offering filings targeting a $1.5 trillion to $1.75 trillion valuation. As examined in analyses of how NVIDIA and SpaceX reshape LEO compute, SpaceX unveiled its AI1 orbital data center satellite design (150 kilowatts peak power with a deployable liquid radiator supplying roughly 1,400 watts per square meter of heat rejection) and announced a Gigasat factory in Bastrop, Texas, on June 8, 2026. In September and October 2026, SpaceX chief executive Elon Musk stated that third-generation Starlink-derived Star Mind compute satellites co-designed with NVIDIA would deliver 250 kilowatts each, carry the equivalent of one NVIDIA Vera Rubin NVL72 rack, and enter mass production in 2027, with SpaceX prospectus filings projecting initial test flights in 2027, initial service in 2028, and a long-term aspirational target of 100 gigawatts per year of orbital AI compute by 2030. Notably, SpaceX’s own May 2026 registration filings include an explicit risk factor warning investors that orbital AI computing involves unproven technology and may not achieve commercial viability.

Four additional massive filings followed SpaceX’s submission. On February 4, 2026, Starcloud filed for up to 88,000 satellites capable of delivering five to 20 gigawatts of compute capacity. On March 19, 2026, Blue Origin filed for Project Sunrise, seeking permission to deploy up to 51,600 solar-powered compute satellites across sun-synchronous orbital planes between 500 and 1,800 kilometers altitude, linked via an optical mesh network and Blue Origin’s planned TeraWave broadband backhaul system (scheduled to begin deployment in the fourth quarter of 2027). As noted in New Space Economy’s assessment of Blue Origin’s Project Sunrise filing, Blue Origin has not yet announced computing hardware or a launch date for the network, and NASA submitted a public comment to the FCC on May 5, 2026, warning that the 51,600-satellite proposal lacked the transparency, rigor, and coordination required to assure orbital safety.

Cowboy Space followed on May 14, 2026, with its Stampede application for 20,000 satellites, and startup Orbital filed on June 24, 2026, for up to 100,000 satellites (~10 gigawatts). Outside the United States, China submitted filings to the International Telecommunication Union (ITU) in late December 2025 covering roughly 200,000 LEO satellites (CTC-1 and CTC-2) and incorporated integrated communication, navigation, sensing, and space computing into its 15th Five-Year Plan in January 2026.

Institutional Feasibility Studies, Concepts, and Strategic Intent

At Stage 2 (Funded Study) and Stage 1 (Concept), government agencies and major corporations are evaluating longer-horizon architectures:

• European Union ASCEND Study (France and Europe): Led by Thales Alenia Space under a €2,047,882 Horizon Europe contract, the Advanced Space Cloud for European Net Zero Emission and Data Sovereignty (ASCEND) consortium completed its feasibility study in June 2024. The study concluded that assembling modular orbital data centers using European Robotic Orbital Support Services (EROSS) technology is technically feasible and outlines targets of 13 building blocks totaling 10 megawatts by 2036 and 1,300 building blocks totaling one gigawatt by 2050, but warned that achieving net lifecycle carbon reductions over terrestrial grids requires developing a reusable heavy launcher 10 times less emissive across its lifecycle than current rockets.
• SpaceBlast (Japan) and Madari Space (United Arab Emirates): Japan’s SpaceBlast conducts a funded Stage 2 orbital data center design study. In Abu Dhabi, Stage 1 startup Madari Space, backed by the Mohammed bin Rashid Innovation Fund, MBRSC, and the United Nations Office for Outer Space Affairs (UNOOSA), signed a November 2025 memorandum of understanding with South Korean rocket builder INNOSPACE to fly hosted LEO edge storage and processing payloads for Earth-observation operators, although its initial third-quarter 2026 pilot window closed without a confirmed launch.
• State Contractors, Defense Integrators, and Hyperscaler Intent (6 Entries): Six major organizations hold Stage 1 statements of intent. In China, the China Aerospace Science and Technology Corporation (CASC) and the China Academy of Space Technology (CAST) published roadmaps targeting an experimental 10-kilowatt, 1,000-kilogram computing satellite around 2027 and 500-kilowatt, 20,000-kilogram space-cloud modules around 2030. In South Korea, Hanwha Systems announced plans to link 64 synthetic aperture radar satellites to an orbital data center architecture by 2031. Among American artificial intelligence leaders, Meta co-developed the Space Llama ISS demonstration in 2025 and signed an April 2026 agreement with Overview Energy reserving up to one gigawatt of space-collected laser-beamed solar power (25-megawatt pilot targeted for 2028) to power terrestrial data centers, Anthropic expressed formal interest in May 2026 in procuring multiple gigawatts of orbital compute from SpaceX, and OpenAI and Relativity Space have also explored space-compute initiatives.

Physics Advantages Versus Thermodynamic, Radiation, and Networking Constraints

Translating orbital data centers from kilowatt-scale technology demonstrations into megawatt-scale commercial infrastructure requires overcoming four severe physical constraints: vacuum heat rejection, orbital illumination geometry, ionizing radiation, and inter-satellite communication bandwidth. Every engineering solution to these challenges adds spacecraft mass, and every additional kilogram of mass multiplies launch and replacement costs.

The Thermodynamic Bottleneck: Radiating Heat in a Vacuum

The most frequently misunderstood aspect of orbital computing is cooling. Popular accounts often describe space as “freezing cold,” implying that satellites enjoy free refrigeration. In reality, the vacuum of low Earth orbit eliminates convective and conductive heat transfer to surrounding air, turning vacuum into a thermal insulator rather than a coolant. As explained in foundational reviews of spacecraft power sources and solar radiation in the space economy, a spacecraft must reject every watt of waste heat generated by its processors, plus absorbed solar radiation and infrared Earth albedo, exclusively through thermal radiation governed by the Stefan-Boltzmann law.

Independent thermodynamic modeling published in 2026 demonstrates how severely radiator mass constrains orbital compute density. According to ABI Research analyst Andrew Cavalier (cited in IEEE Spectrum in June 2026) and NASA Jet Propulsion Laboratory physicist Slava G. Turyshev in his April 2026 technical study on orbital data center constraints, a single 700-watt NVIDIA H100 GPU operating at 60 degrees Celsius requires approximately 1.4 square meters of two-sided radiator area, and a one-megawatt orbital computing system requires roughly 2,500 square meters of thermal radiators alongside 5,640 square meters of solar panels. Together, deployable solar arrays, pumped two-phase fluid loops, and radiator panels account for 65 to 70 percent of total satellite mass, requiring roughly 34 to 59 kilograms of spacecraft mass per delivered kilowatt of computing power. Even NVIDIA chief executive Jensen Huang noted in February 2026 that orbital compute economics remain poor today because of heat dissipation constraints, just one month before NVIDIA unveiled its space-qualified Space-1 Vera Rubin module.

Orbital Geometry, Ionizing Radiation, and Hardware Lifespan

Orbital mechanics also complicates the promise of uninterrupted solar power. Because of Earth’s 23.4-degree axial tilt, a dawn-dusk sun-synchronous terminator orbit only remains completely free of Earth’s shadow throughout the entire year at altitudes above roughly 1,400 kilometers. Yet all five major U.S. FCC filings propose operating shells starting at 500 kilometers to minimize round-trip signal latency and avoid the intense trapped proton radiation of the inner Van Allen belt, which begins near 1,000 to 1,500 kilometers altitude. Satellites orbiting between 500 and 700 kilometers still experience seasonal eclipse periods requiring lithium-ion battery banks or power-throttled compute bursts.

Within low Earth orbit, galactic cosmic rays and solar proton events cause single-event upsets (bit-flips in memory) and cumulative total ionizing dose degradation. Historically, Hewlett Packard Enterprise’s Spaceborne Computer-1 lost nine of its 20 solid-state drives during its ISS mission. Although Google’s June 2026 proton-beam tests proved that commercial Trillium TPUs survive a five-year LEO equivalent dose without permanent hard failure, HBM memory irregularities began appearing at two krad(Si), requiring continuous error-correcting code (ECC) software, lockstep execution, or triple modular redundancy. Industry cost models, including Andrew McCalip’s public orbital compute calculator, incorporate an estimated nine percent annual GPU failure rate in orbit and a mandatory five-year full-satellite replacement cycle, compared with a 15-year facility lifespan on Earth, because no robotic servicing system has ever repaired or swapped a failed server blade in orbit.

Networking Bandwidth and Space-Native Versus Earth-Origin Workloads

Connectivity imposes a third structural filter on which workloads belong in space. Inside a terrestrial AI training cluster, thousands of GPUs exchange intermediate gradients over NVLink and InfiniBand fabrics operating at 3.6 to 7.2 terabits per second per node. In contrast, the fastest operational commercial optical inter-satellite links flown on orbital compute constellations in 2026, such as Kepler’s 2.5 Gbps SDA terminals or Starlink’s 25 Gbps Mini Lasers, operate three orders of magnitude slower, as detailed in New Space Economy’s guide to Starlink satellite laser communications. Even Google’s 1.6 Tbps single-transceiver laboratory bench test remains below the roughly 10 Tbps per satellite link required for tightly coupled foundation-model training across formation-flying satellites.

Consequently, independent academic studies published in October 2026 demonstrate that near-term orbital data centers deliver their strongest value on space-native data rather than ordinary terrestrial cloud jobs. As analyzed in New Space Economy’s coverage of who benefits from orbital data centers, which AI workloads belong on satellites, and when in-orbit computing sends answers, processing synthetic aperture radar, hyperspectral, and optical imagery directly in orbit reduces required space-to-ground downlink bandwidth by 85 to 99 percent and returns actionable detections to client spacecraft within two seconds. Conversely, uploading massive datasets from Earth to orbit for general-purpose cloud computing incurs severe uplink and downlink bottlenecks.

Launch Cost Parity, Hardware Refresh Cycles, and Space Governance Hurdles

Whether orbital data centers expand beyond space-native edge processing into a multi-gigawatt overflow market for terrestrial AI training depends almost entirely on launch economics, insurance capacity, and international space governance.

Independent Cost Models and the Starship Cadence Dependency

Published economic analyses from 2025 and 2026 converge on a stark conclusion: at current launch prices of $600 to $7,000 per kilogram, operating an orbital data center costs roughly 2.5 to 10 times more per unit of compute than building and powering an equivalent facility on Earth. Startup white papers from Starcloud and Lonestar quote effective orbital solar electricity costs as low as $0.002 to $0.005 per kilowatt-hour compared with $0.04 to $0.08 per kilowatt-hour for wholesale terrestrial grid power. However, financial analysts including short-seller Jim Chanos and aerospace engineers at NASA JPL and SemiAnalysis note that electricity represents only five to seven percent of AI data center revenue; capital depreciation of GPUs, launch mass, thermal radiators, radiation shielding, spare capacity, and five-year satellite replacement dominate total levelized cost.

The table below summarizes six major independent and corporate cost models evaluating orbital data center economics and launch-price parity thresholds as of late 2026.

Across these models, reaching economic parity with terrestrial data centers requires launch costs to drop into the $50 to $500 per kilogram range, a reduction that depends on SpaceX’s fully reusable Starship achieving both low marginal cost per flight and high launch cadence. On September 28, 2026, Starship Flight 14 reached orbit and deployed 26 production Starlink V3 satellites before an engine anomaly shortened the flight to roughly two orbits. Yet deploying a single 100-megawatt orbital cluster (~4,200 metric tons of hardware) requires more than 20 dedicated Starship launches, and Google’s Project Suncatcher paper estimates roughly 1,800 Starship flights to loft a gigawatt-scale constellation. As examined in New Space Economy’s analysis of the Falcon 9 to Starship transition, Starcloud chief executive Philip Johnston warned in August 2026 that securing sufficient launch capacity as Falcon 9 winds down toward retirement around 2028 is the primary bottleneck facing independent orbital data center startups, particularly because SpaceX is simultaneously their dominant launch supplier and their largest direct competitor.

Insurance Limits, Orbital Congestion, and Data Sovereignty Rules

Three institutional hurdles further constrain mega-constellation deployment:

• Space Insurance Capacity: Global space insurance underwriting capacity per single launch or in-orbit risk stands at roughly $400 million to $700 million against an annual global premium pool near $500 million, with launch premiums averaging 13 percent of insured value and annual in-orbit coverage costing one to 2.8 percent. Insuring a multi-billion-dollar, megawatt-scale orbital GPU cluster will require layered reinsurance pools or sovereign backstops that do not yet exist, following initial exploratory briefings between orbital compute startups and Lloyd’s of London underwriters in June 2026.
• Orbital Debris, Astronomy, and Reentry Risk: As highlighted in New Space Economy’s review of why orbital data center safety risks are drawing scrutiny, astronomers from the American Astronomical Society and the European Southern Observatory warn that deploying hundreds of thousands of satellites with 75-to-132-meter solar wingspans in dawn-dusk terminator orbits would severely degrade optical astronomy and near-Earth asteroid detection. Furthermore, replacing massive radiator-equipped satellites every five years via atmospheric disposal raises upper-atmosphere alumina aerosol deposition and ground casualty risks if five percent of heavy spacecraft components survive reentry.
• Licensing and Data Sovereignty: No jurisdiction on Earth has enacted a binding regulatory regime tailored specifically to orbital data centers, and the FCC has not granted any of the five mega-constellation filings as of November 2026. In July 2026, the FCC adopted its Space Modernization for the 21st Century framework (FCC 26-47) introducing modular licensing categories, followed by a voluntary certification pilot from the U.S. Department of Commerce Office of Space Commerce in August 2026 and the bipartisan House passage of the Remote Access Security Act (369 to 22) in January 2026 governing cloud export controls. Notably, an October 1, 2026 legal analysis by WilmerHale confirmed that placing servers in outer space does not create a lawless extraterritorial data haven: under Article VIII of the 1967 Outer Space Treaty, the state of registry retains jurisdiction over a spacecraft, meaning the European Union’s General Data Protection Regulation (GDPR) and the United States CLOUD Act continue to apply to personal and enterprise data processed in orbit.

Summary

The global orbital data center sector has graduated from theoretical white papers to flight-proven demonstrations, yet a wide chasm separates the hardware operating in low Earth orbit as of November 2026 from the million-satellite constellations proposed in regulatory filings. Across 53 active programs and six strategic intents in 15 countries, 21 initiatives have flown computing or storage hardware in space. Kepler Communications sells commercial edge compute across a 10-satellite optical mesh, Chinese operators BUPT Tiansuan, ADA Space, and Zhejiang Lab run multi-satellite pathfinder networks totaling more than 40 spacecraft, Axiom Space links free-flying orbital nodes over laser crosslinks, Lonestar validates cislunar disaster-recovery storage, and Starcloud and Google Project Suncatcher have proved that commercial data-center GPUs and TPUs can execute artificial intelligence models in orbit.

Near-term commercial viability favors space-native workloads (processing Earth-observation imagery, synthetic aperture radar, radio-frequency signals, and autonomous spacecraft telemetry close to the sensor to eliminate 85 to 99 percent of downlink bottlenecks) alongside sovereign disaster-recovery vaults. Offloading terrestrial hyperscale AI training into orbit remains gated by uncompromising physics and launch economics: radiating megawatts of heat in a vacuum requires thousands of square meters of heavy radiator panels, unserviced commercial silicon faces a five-year replacement cycle, and orbital compute costs roughly three to four times more per GPU-hour than terrestrial facilities until fully reusable heavy-lift rockets drive launch prices below $100 to $200 per kilogram. Between 2027 and 2030, multi-kilowatt missions from Starcloud, Google, Aethero, Axiom, Sophia Space, Cowboy Space, and Chinese constellations will determine whether orbital data centers remain a specialized edge-computing layer for the space economy or scale into an overflow power grid for global artificial intelligence.

Appendix: Useful Books Available on Amazon

• The Datacenter as a Computer: Designing Warehouse-Scale Machines
• Space to Grow: Unlocking the Final Economic Frontier
• The Space Economy: Capitalize on the Greatest Business Opportunity of Our Lifetime
• Spacecraft Systems Engineering
• Orbital Mechanics for Engineering Students
• Power and Prediction: The Disruptive Economics of Artificial Intelligence
• Prediction Machines, Updated and Expanded: The Simple Economics of Artificial Intelligence
• Chip War: The Fight for the World’s Most Critical Technology

Appendix: Top Questions Answered in This Article

How Many Orbital Data Center Programs Have Flown Hardware in Space as of November 2026?

As of November 2026, 21 out of 53 tracked orbital data center programs worldwide have flown their own computing or data-storage hardware in space, either aboard dedicated satellites, hosted payloads, lunar landers, or the International Space Station. Ten additional programs have flown software or supplied components to third-party spacecraft, and 22 programs plus six corporate statements of intent remain in ground development, feasibility study, or concept stages.

Which Company Currently Sells Commercial Orbital Compute Services?

Canadian operator Kepler Communications is the only program classified at Stage 7 (Operational Service) selling distributed in-orbit computing commercially as of late 2026. After launching 10 optical relay satellites in January 2026 and commissioning 40 onboard NVIDIA Jetson Orin modules in March 2026, Kepler opened its 2.5 Gbps optically linked mesh network to commercial compute customers in April 2026.

Which Countries Operate Multi-Satellite Computing Constellations in Orbit?

Canada and China operate the four active multi-satellite computing constellations in low Earth orbit as of November 2026. Alongside Canada’s 10-satellite Kepler network, China operates three Stage 6 pathfinder constellations: BUPT’s 16-satellite Tiansuan cloud constellation, Zhejiang Lab’s 14-satellite Three-Body Computing Constellation, and ADA Space’s computing fleet, which expanded with eight additional liquid-cooled test satellites in October 2026.

What Did Starcloud and Google Project Suncatcher Prove in Orbit?

Starcloud and Google proved that commercial data-center artificial intelligence accelerators can survive launch and execute machine-learning models in low Earth orbit. Starcloud’s 60-kilogram Starcloud-1 satellite launched an NVIDIA H100 GPU in November 2025 to train nanoGPT and run Gemma models, and Google’s Project Suncatcher prototype launched four Trillium TPUs on October 1, 2026, returning Gemma 4 responses later that month.

Why Are Companies Proposing to Build Data Centers in Outer Space?

Proponents seek to bypass terrestrial electric grid interconnection queues of seven to 12 years, land-use moratoria, and freshwater cooling constraints driven by surging artificial intelligence power demand. Satellites in dawn-dusk sun-synchronous orbits receive up to 1,361 watts per square meter of nearly continuous solar energy and reject heat into deep space without consuming water.

Why Is Cooling a High-Power Data Center Difficult in the Vacuum of Space?

Because space is a vacuum, heat cannot leave a spacecraft through air convection or conduction and must be emitted as infrared thermal radiation from large radiator panels. Engineering studies show a 700-watt H100 GPU requires roughly 1.4 square meters of radiator area and a one-megawatt cluster needs 2,500 square meters of radiators, accounting for up to 70 percent of satellite mass.

Which Companies Have Filed FCC Applications for Mega-Constellation Orbital Data Centers?

Five United States operators filed Federal Communications Commission applications between January and June 2026 seeking authorization for 1,259,600 orbital compute satellites. SpaceX and xAI requested up to 1,000,000 satellites, Orbital requested 100,000, Starcloud requested 88,000, Blue Origin requested 51,600 for Project Sunrise, and Cowboy Space requested 20,000 for its Stampede network.

What Launch Cost per Kilogram Is Needed for Orbital Data Centers to Compete With Earth Facilities?

Independent economic models from SemiAnalysis, Boston Consulting Group, Bain & Company, Google Research, and NASA Jet Propulsion Laboratory estimate that orbital compute costs 2.5 to 10 times more than terrestrial compute at current launch prices of $600 to $7,000 per kilogram. Reaching cost parity with ground data centers requires fully reusable heavy-lift rockets like SpaceX’s Starship to reduce launch prices to between $50 and $500 per kilogram.

Which Artificial Intelligence Workloads Make the Most Economic Sense in Orbit Today?

Space-native edge workloads, such as processing Earth-observation imagery, synthetic aperture radar scans, and radio-frequency signals directly aboard or near the collecting satellite, offer the strongest near-term return. Filtering and compressing raw sensor data in orbit cuts required space-to-ground downlink bandwidth by 85 to 99 percent and delivers rapid alerts to spacecraft and defense users.

Do Terrestrial Data Privacy and Sovereignty Laws Apply to Orbital Data Centers?

Yes, international space law extends terrestrial legal jurisdictions directly into orbit rather than creating a regulatory vacuum. Under Article VIII of the 1967 Outer Space Treaty, the country that registers a spacecraft retains legal jurisdiction and control over it, meaning privacy and disclosure frameworks such as the European Union’s GDPR and the United States CLOUD Act apply to data processed in space.

Appendix: Glossary of Key Terms

Orbital Data Center

A spacecraft or networked constellation of satellites equipped with high-performance processors, data storage, solar arrays, thermal radiators, and optical communications links designed to process or store digital workloads in space rather than transmitting raw data to terrestrial facilities.

Sun-Synchronous Orbit

A nearly polar low Earth orbit in which a satellite’s orbital plane precesses at the same rate that Earth revolves around the Sun, allowing spacecraft in a dawn-dusk terminator configuration to remain in almost continuous sunlight for solar power generation.

Space Development Agency

An organization within the United States Space Force responsible for deploying the Proliferated Warfighter Space Architecture and establishing open interoperability standards for laser optical communication terminals connecting military and commercial satellites in low Earth orbit.

Tera-Operations per Second

A standardized metric of processor performance that measures how many trillions of computational operations an artificial intelligence accelerator or digital signal processor can execute each second when running machine-learning inference or sensor-processing tasks.

Tensor Processing Unit

A custom application-specific integrated circuit developed by Google specifically to accelerate neural network machine-learning workloads, tested in low Earth orbit aboard the Project Suncatcher prototype satellite launched in October 2026 to evaluate radiation and thermal behavior.

Single-Event Upset

A temporary or permanent state change in a microelectronic memory cell or logic circuit caused when a single high-energy ionizing particle from cosmic rays or solar proton events strikes a semiconductor node in space, requiring error-correcting software or hardware redundancy.

Free-Space Optical Communication

A point-to-point wireless transmission technology that uses modulated laser beams in the infrared spectrum to transfer data at multi-gigabit or terabit speeds between satellites in space or between orbital spacecraft and terrestrial optical ground stations.

Stefan-Boltzmann Law

A physical principle stating that the total radiant heat power emitted from a surface is proportional to the fourth power of its absolute temperature, governing how much radiator surface area a spacecraft requires to reject waste processor heat into vacuum.