Orbital Data Centers

Orbital data centers are computing systems placed in Earth orbit, from a processor that filters its own satellite's sensor data to a proposed optical network of server spacecraft. The gap between proposals and flown hardware is wide: filings before the US Federal Communications Commission (FCC) propose systems of up to 1,000,000 satellites,^[26] while the most advanced AI accelerator known to have reached orbit is a single NVIDIA H100 GPU, launched on November 2, 2025.^[8][9]

A useful definition depends on architecture and workload, not on whether a company calls a payload a data center. Peer-reviewed literature distinguishes orbital edge data centers, which process data where it is collected, from orbital cloud data centers intended to provide shared compute and storage across a network.^[1] The distinction matters because the evidence is uneven. Onboard inference, high-performance computing, commercial servers, and limited model training have all been demonstrated in orbit. Dedicated commercial compute payloads have also launched. As of August 14, 2026, however, no megawatt-class orbital computing facility, tightly coupled AI-training cluster, or general-purpose cloud service for terrestrial users had been demonstrated. The peer-reviewed literature still treats systems at those scales as frameworks to evaluate, while the dated evidence below consists of studies, filings, and development programs.^[1]

Status as of August 2026#

Orbital data centers remained at experiment scale on August 14, 2026. Known orbital compute hardware included Starcloud's Starcloud-1, launched in November 2025,^[8][9] two Axiom nodes launched January 11, 2026,^[7] and hosted Spaceborne Computer-2 aboard the International Space Station.^[3] SpaceX named Starmind on June 23 and said on August 4 that AI1 satellites would carry NVIDIA Rubin GPUs and Vera CPUs.^[34][35] Its million-satellite application remained pending,^[26][36] alongside Starcloud's 88,000-satellite filing^[37] and Blue Origin's 51,600-satellite request.^[38] Starcloud says Starcloud-2, its first commercial spacecraft, will be fully operational in sun-synchronous orbit during 2027.^[39] Musk forecast on August 14 that orbital compute would become "the only way to scale AI" probably in 2029 because of power availability and permitting constraints on land.^[43] This was a forecast, not a result; SpaceX's prospectus calls the program early-stage and unproven.^[44] The American Astronomical Society petitioned March 6 to deny SpaceX's application,^[40] environmental groups asked the FCC on July 8 to pause licensing,^[41] and a June analysis found at least a tenfold GPU-year cost penalty even with optimistic launch pricing.^[42] Government and technical assessments identified major power, cooling, networking, and cost limits.^[1][45][46]

Definitions and taxonomy#

Peer-reviewed work sorts orbital computing into five categories, each with a defining function and a boundary that separates it from the next larger claim.^[1]

Conventional spacecraft computers are not automatically orbital data centers. Navigation computers, instrument controllers, and processors that execute fixed mission software are ordinary spacecraft avionics. The term becomes useful when the payload provides substantial data reduction, reprogrammable acceleration, storage, or shared computing as a mission function. Lunar surface archives are another adjacent idea, but they are off-world storage rather than orbital data centers.^[1]

The clearest near-term category is edge computing. ESA's PhiSat-1, launched on September 3, 2020, used an Intel Movidius Myriad 2 processor to classify cloudy hyperspectral images and avoid downlinking unusable data.^[2] That was AI inference on data collected by the same spacecraft, not AI training and not a cloud service.

What has been demonstrated#

The HPE Spaceborne Computer program established that recent commercial hardware can perform useful high-performance computing in a protected orbital environment. Spaceborne Computer-1 reached the International Space Station in 2017 as an unmodified commercial system. NASA reports that software monitored abnormal behavior and adjusted performance, and that the system completed more than 18 months of operation before returning to Earth. Because the computer was hosted on the station, the trial did not validate a standalone spacecraft's power, cooling, communications, or servicing architecture.^[3]

Spaceborne Computer-2 moved from endurance testing toward applications. An upgraded third configuration arrived in 2024 with more than 130 TB of flash storage. HPE reported one 2022 life-sciences workload in which 2.8 GB of source data, normally requiring about 18 hours to transmit, was processed into a 92 KB result that took about two seconds to downlink.^[4] The example shows why reducing data before transmission can be valuable; it does not establish the economics of a standalone satellite.

The earliest well-documented onboard machine-learning training occurred before the commercial orbital-data-center wave. In autumn 2022, University of Oxford researchers uploaded code to D-Orbit's ION SCV004 spacecraft and trained a small cloud-detection classifier in orbit. An encoder had been trained on the ground, while the classifier was trained onboard from compressed image representations.^[5] The associated 2023 paper is a preprint, not a peer-reviewed journal article; its authors describe the result as the first few-shot training of a machine-learning model onboard a satellite to their knowledge.^[6] This makes the later claim that Starcloud-1 trained the first AI model in space too broad.

Axiom Space describes several progressively more independent demonstrations. An AWS Snowcone device operated on the station during Ax-1 in 2022, and its AxDCU-1 prototype was deployed there in fall 2025. Axiom says two dedicated data-center nodes launched on January 11, 2026 with the first tranche of Kepler Communications' optical relay network. Its public description does not identify the nodes as independent spacecraft and does not provide enough telemetry to establish uptime, customer use, or delivered computing performance independently.^[7]

Starcloud launched Starcloud-1 on November 2, 2025. The company identifies it as the first spacecraft to carry an NVIDIA H100 GPU, a narrow and verifiable hardware first.^[8][9] Starcloud reports that it later ran the Gemma language model and trained NanoGPT onboard.^[8] That was an early language-model demonstration, not the first training of any machine-learning model in orbit.

Dated program matrix#

The program matrix below separates flight evidence from plans across orbital computing efforts. "Launched" does not mean that a commercial service or its advertised performance has been independently verified.

ESA's 2022 to 2024 generic study simulated an architecture rather than building one. It found a workload-dependent tipping point at which processing data onboard could be preferable to downloading raw data.^[10] ASCEND studied a much larger European cloud concept. Its official project report defined a 10 MW minimum viable product and a 1 GW objective for 2050, while calling for further work on hardware compatibility, architecture, business planning, launch capacity, and orbital assembly.^[11] Thales Alenia Space also said ASCEND's environmental case required a future launcher about ten times less emissive over its lifecycle than the baseline used in the study.^[12] These were conditional study outputs, not proof that the proposed system would be carbon-neutral or economical.

Google announced Project Suncatcher as a research moonshot, not a product. Its June 2026 revised preprint models one illustrative cluster of 81 satellites at a mean altitude of 650 km, within a 1 km radius and with neighbor separations varying roughly from 100 to 200 m. The same paper reports ground radiation tests and a short-path optical bench test, while identifying thermal control, high-bandwidth ground communications, reliability, and repair as unfinished work.^[14] Google and Planet target two learning satellites for early 2027, so no orbital TPU-cluster result existed by the article date.^[13]

Power and illumination#

Solar power in orbit begins with the total solar irradiance near Earth, about 1,361 W/m^2.^[16] If a panel were held normal to the Sun and converted 30 percent of that flux to electricity, the ideal photovoltaic area for 1 MW would be:

P / (eta x S) = 1,000,000 W / (0.30 x 1,361 W/m^2) = about 2,450 m^2.^[16]

That is a physics floor, not a spacecraft design. It excludes pointing losses, wiring and conversion losses, payloads other than the computers, end-of-life degradation, temperature effects, shadow periods, battery charging, and redundancy. The 2026 Google preprint's "up to eight times" annual energy comparison is between a panel in a favorable orbit and the same panel on the ground at mid-latitude. It is not an eightfold advantage for the complete power system.^[14]

A dawn-dusk sun-synchronous orbit can maximize illumination and give a comparatively stable thermal geometry, but "sun-synchronous" does not mean "always in sunlight." Eclipse duration depends on altitude, local solar time, season, orbital beta angle, and the size of the formation. ESA's Aeolus occupied a dawn-dusk orbit at about 320 km, yet passed through Earth's shadow for as long as 20 minutes per orbit over the winter hemisphere.^[17] A real design therefore needs an orbit-specific eclipse analysis and either storage, workload throttling, redundant power, or some combination.

Orbit selection trades power against other constraints:

No orbit is preferred for every design. Edge processing should usually stay close to the sensor it serves. A cluster optimized for solar availability may be poorly placed for low-latency access to a particular ground market.^[14][17]

Heat rejection#

Space is a vacuum, so an external radiator cannot remove heat by convection. Internal heat can move through conduction, heat pipes, or pumped fluid loops, but the spacecraft ultimately rejects it as thermal radiation. For an ideal surface looking only at deep space, where a roughly 3 K sink is negligible beside a 300 K radiator, the rejected heat is approximately:^[18]

Q = epsilon x sigma x A x T_r^4.^[18]

Here Q is waste heat, epsilon is surface emissivity, sigma is the Stefan-Boltzmann constant, A is emitting surface area, and T_r is radiator temperature.^[18] The fourth-power temperature term makes a warmer radiator smaller, but chips and coolant loops limit how warm it may run.

The following values are transparent lower bounds calculated from the cited radiation equation with emissivity 0.90, a 3 K sink, view factor 1, and no absorbed sunlight, Earth infrared, or albedo. Area means active emitting area, not necessarily the projected footprint of a two-sided panel.^[18]

These figures should not be quoted as flight-system sizes. A radiator partly sees warm Earth, or absorbs solar and reflected energy, rejects less net heat. Heat exchangers, pumps, working fluid, pipes, deployment hardware, micrometeoroid tolerance, margins, and redundant loops add mass. Coatings also change during a mission: NASA notes that atomic oxygen in low Earth orbit, ultraviolet exposure, and radiation can alter optical properties and degrade thermal performance.^[19]

A 2026 technical preprint, using a specified high-sunlight 1 MW case rather than the ideal table above, estimated about 2,500 m^2 of radiator and 5,640 m^2 of beginning-of-life photovoltaic area.^[20] Its similarity to the 300 K ideal radiator result should not obscure the assumptions. Different chip efficiency, coolant temperature, radiator coating, geometry, orbit, or eclipse duty cycle changes the answer.

Radiation, faults, and reliability#

Space radiation causes several different failure modes. Total ionizing dose accumulates over time; displacement damage alters semiconductor materials; and a single energetic particle can cause a bit flip, transient, latch-up, or destructive burnout. ESA treats these as distinct effects whose severity depends on orbit, shielding, mission length, component process, and circuit design.^[21]

High-performance commercial chips offer far more compute than traditional radiation-hardened processors, but they need a system-level safety case. Measures can include shielding, error-correcting memory, memory scrubbing, watchdogs, checkpoint and restart, duplicate computation, power cycling, fault isolation, and a simpler radiation-tolerant controller supervising the commercial accelerator. Spaceborne Computer-1 is evidence that software management can work in a particular hosted environment, not proof that every GPU can survive every orbit for years.^[3][14][21]

Google's Trillium tests illustrate both promise and uncertainty. The June 2026 preprint modeled a five-year, 10 mm aluminum-equivalent shielded sun-synchronous mission at about 150 rad(Si) per year. In a 67 MeV proton beam, high-bandwidth-memory stress tests began showing irregularities after 2 krad(Si), while other end-to-end and compute tests ran to 15 krad(Si) without a total-dose hard failure. The tests also observed silent data corruption and system interruptions. Google judged inference errors potentially manageable, but said the effects on long training runs and the effectiveness of system-level mitigations still require study.^[14] A proton-beam test cannot by itself validate the full particle spectrum, every component, manufacturing lot, thermal state, or years of autonomous operation.

Reliability also includes ordinary hardware failures. Pumps, storage devices, deployable structures, solar arrays, and batteries can degrade or fail. Terrestrial data centers tolerate failed components because technicians replace them. An orbital service must instead carry spare capacity, isolate faults remotely, accept declining performance, or pay for robotic servicing or replacement launches.^[14][19]

Networks, latency, and workload placement#

Orbit shortens the path to another spacecraft but adds a bottleneck for terrestrial data. A low Earth orbit radio path can have less than 10 ms of one-way propagation delay to a visible ground station, yet that physical floor is not application latency. Ground-station visibility, scheduling, routing, weather, retransmission, queueing, encryption, and the terrestrial path can dominate.^[22]

Optical links provide high rates without consuming the same radio-frequency bandwidth, but they require narrow-beam acquisition, tracking, and pointing. NASA's TBIRD demonstration achieved a 200 Gbit/s space-to-ground optical downlink and sent 4.8 TB error-free in five minutes during a 2023 pass.^[23] That result proves a burst link, not continuous global capacity. Clouds and severe atmospheric attenuation can prevent an optical downlink, so networks need geographically diverse ground stations, storage, alternate routing, or radio fallback.^[24]

Tightly coupled AI training is harder than downloading finished results. Google's preprint estimates that its proposed cluster would need aggregate bandwidth on the order of 10 Tbit/s per link, compared with roughly 1 to 100 Gbit/s for surveyed commercial optical inter-satellite terminals. A bench setup reached 800 Gbit/s in one direction and 1.6 Tbit/s bidirectionally over a short free-space path, but it was not a moving, thermally constrained flight network.^[14] Close formation improves the optical link budget while making relative navigation, collision avoidance, plume impingement, and fault containment harder.

Workload placement determines whether the communications constraint closes:

The strongest demonstrated case is data reduction at its source. General cloud computing for users on Earth gives up that advantage because both inputs and outputs traverse the space-to-ground boundary.^[2][10][20]

Economics and lifecycle#

SpaceX's June 5 EU prospectus said orbital AI-compute satellites could begin deploying as early as 2028. SpaceX said it believed about 10 GW of annual deployment could support a commercially attractive business and described up to 100 GW per year as a long-term ambition. The 100 GW case assumes 100 kW of compute power per metric ton and 100 metric tons to orbit per Starship launch, requiring thousands of launches each year.^[44] These are forward-looking goals, not demonstrated performance. The same filing says the initiative is early-stage, unproven at commercial scale or at all, may fail, and has a timeline and launch cadence that may be difficult or impossible to determine.^[44] GAO reported the Department of Energy's projection that data centers could use up to 12 percent of US electricity in 2028. It also found that large centers would need solar arrays larger than any launched and assembled as of April 2026, that cooling at this scale remained unproven, and that more advanced data-transfer systems might be needed.^[45] The peer-reviewed Nature Electronics paper provides architectural and lifecycle frameworks, not a parity date.^[1] Under its simplified roofline model, a July 2026 preprint finds single-satellite inference potentially feasible but frontier-scale LLM training unlikely to compete with terrestrial data centers.^[46] None of these sources establishes that orbit will be the only scalable option in 2029.

Launch price is only one part of delivered compute cost, but current public pricing shows the scale of the gap. As of August 2026, SpaceX advertised a Falcon 9 rideshare minimum of $350,000 for 50 kg to sun-synchronous orbit, with additional mass at $7,000 per kilogram.^[25] That is a purchasable rideshare price for compatible payloads, not SpaceX's internal cost, a universal dedicated-launch price, or a forecast for Starship.

Google's preprint does not claim that launch already costs $200 per kilogram. It applies a learning-curve model and says prices might fall below that level in the mid-2030s if its assumptions, including high launch cadence, hold.^[14] The separate 2026 Turyshev analysis is also a preprint. For its representative 1 MW system it estimated 34 to 59 kg of total spacecraft mass per delivered kilowatt. With about 40 kg/kW and a terrestrial infrastructure benchmark of $10,000 to $40,000 per kilowatt, it derived only $250 to $1,000 per kilogram for spacecraft construction and launch together, before communications, operations, utilization, and finite lifetime.^[20] The result is a necessary-condition model, not a price forecast.

A serious comparison needs at least:

• spacecraft development and qualification;
• solar, storage, conversion, radiator, propulsion, shielding, and communication mass;
• integration, launch, insurance, ground stations, and network operations;
• useful compute after radiation margins, redundancy, and thermal throttling;
• utilization and revenue during each ground or customer contact window;
• hardware lifetime, obsolescence, replacement cadence, and failed launches;
• disposal, collision avoidance, and any servicing system;
• terrestrial alternatives using low-carbon power, dry cooling, or better chips.^[14][20]

Avoided land and cooling water are real benefits only if the comparison uses the same service boundary. Moving compute to orbit does not remove manufacturing, ground terminals, terrestrial networking, launches, or replacement hardware. It also does not make maintenance free. Software can be updated remotely, but physical repairs require crewed or robotic access, a resupply architecture, or disposal and replacement. Axiom explicitly lists software updates, resupply, and replacement modules as different maintenance paths.^[7]

Lower altitude can shorten passive deorbit time but increases drag and propulsion demand. Higher altitude can lengthen life and reduce drag, but failed spacecraft persist longer. Hardware may also become economically obsolete before its solar arrays or structure wear out. Designs therefore need a full cycle from production through launch, useful compute-years, replacement, and end-of-life disposal.^[14][17][20]

Regulation, debris, and public risk#

The largest proposal in the public regulatory record is SpaceX's Orbital Data Center System. The FCC's February 4, 2026 public notice says SpaceX applied for up to one million satellites at 500 to 2,000 km, using 30-degree and sun-synchronous inclinations, optical inter-satellite links, and Ka-band Earth links that could route through Starlink.^[26] The application predates the Starmind name and does not mention Starmind or AI1. "Accepted for filing" means the agency found the application complete enough for public processing and comment. It does not grant authority to deploy the system.

The FCC proceeding covers an orbital communications system, specified frequency bands, telemetry, tracking and command, deployment milestones, and requested waivers. It is not blanket approval of every launch, satellite, or reentry. Optical crosslinks reduce dependence on radio-frequency links within the constellation, but SpaceX still requested Ka-band Earth links.^[26]

Post-mission disposal is a design input, not an afterthought. The FCC rule for US-licensed spacecraft ending missions in or passing through the region below 2,000 km and using uncontrolled reentry requires disposal as soon as practicable and no later than five years after mission end.^[27] NASA's active debris standard limits the calculated worldwide human-casualty risk from a single uncontrolled reentry to 1 in 10,000.^[28] Large structures may need design-for-demise, controlled reentry, or recovery. Even when each reentry meets an event-level criterion, a high replacement cadence creates an aggregate-risk and atmospheric-mass question.^[29]

Collision avoidance becomes harder as count, cross-sectional area, and differential orbital geometry increase. ESA's May 2026 environment report identifies 400 to 600 km as an altitude range with high concentrations of active and maneuverable satellites, says current global mitigation performance is insufficient for long-term sustainability, and reports 1,200 intact-object reentries during 2025.^[29] A compute constellation therefore needs maneuverability, tracking data, autonomous conjunction response, passivation, reliable disposal, and enough propulsion reserve to execute them. A failed, non-maneuvering server satellite is still space debris.

Environmental and astronomy impacts#

Claims that orbital data centers could be carbon neutral remain model-dependent. The 2025 Nature Electronics paper is peer-reviewed and provides a lifecycle accounting framework, but its title describes the objective being explored, not an operating carbon-neutral system.^[1] ASCEND similarly conditioned its favorable case on a much less emissive future launcher.^[12] A lifecycle assessment must state the terrestrial electricity mix, launch vehicle and propellant, reuse, payload mass fraction, lifetime, replacement rate, manufacturing emissions, and avoided ground infrastructure.

Launch and reentry emissions occur high in the atmosphere, where their chemistry and residence times differ from surface emissions. A peer-reviewed 2025 Journal of Geophysical Research study modeled a hypothetical 10 Gg per year injection consisting entirely of aluminum oxide from reentries. It found scenario-dependent accumulation and atmospheric changes, while stressing that the particles produced by actual vaporization remain poorly characterized.^[30] A 2026 peer-reviewed Earth's Future study likewise modeled radiative forcing and ozone effects from megaconstellation launch and reentry pathways and highlighted sensitivity to assumptions.^[31] These studies establish a risk requiring measurement and assessment, not a precise impact for an unbuilt orbital data center.

Astronomy effects depend on spacecraft number, altitude, attitude, size, reflectivity, radio emissions, and operational behavior. A 2026 peer-reviewed Monthly Notices of the Royal Astronomical Society paper modeled one hypothetical 5 GW facility with a roughly 4 km solar array. Under its assumptions, such a structure in sun-synchronous low Earth orbit would be exceptionally bright during twilight and large enough in angular extent to affect observations.^[32] Those results should not be applied directly to a small edge-compute satellite, but they show why operators of large arrays need credible brightness, occultation, radio-frequency, and debris analyses before deployment. More general peer-reviewed constellation studies find that no single scheduling or brightness mitigation protects every optical and near-infrared observation.^[33]

Open technical evidence gaps#

Nine question areas separate what orbital computing has demonstrated from what its proposed scale requires.

The evidence supports orbital edge computing today. It supports experiments with shared hosted compute and increasingly powerful commercial accelerators. It does not yet support treating proposed orbital clouds as proven substitutes for terrestrial data centers. The decisive measurements will come from openly reported power balance, radiator performance, fault rates, useful workload throughput, link availability, service utilization, replacement cadence, and lifecycle impacts from actual flight systems.^{[1][7][14][20]}

References#

1. The development of carbon-neutral data centres in space - Nature Electronics, peer-reviewed perspective, October 27, 2025.
2. Artificial Intelligence for Earth observation - European Space Agency.
3. Cutting-Edge Computing Goes Spaceborne - NASA Spinoff, February 11, 2025.
4. HPE Spaceborne Computer-2 returns to the International Space Station - Hewlett Packard Enterprise, January 30, 2024.
5. Researchers successfully train a machine learning model in outer space for the first time - University of Oxford, July 28, 2023.
6. Fast model inference and training on-board of satellites - arXiv preprint, July 17, 2023.
7. Orbital Data Centers - Axiom Space, accessed August 12, 2026.
8. Starcloud-1 - Starcloud, accessed August 12, 2026.
9. Bandwagon-4 - SpaceX, November 2, 2025.
10. Space based data centres - European Space Agency Nebula Public Library, 2022 to 2024.
11. Advanced Space Cloud for European Net zero emissions and Data sovereignty - European Commission CORDIS, updated January 14, 2026.
12. Thales Alenia Space reveals results of ASCEND feasibility study on space data centers - Thales Alenia Space, June 27, 2024.
13. Project Suncatcher explores powering AI in space - Google, November 4, 2025.
14. Towards a future space-based, highly scalable AI infrastructure system design - arXiv preprint, revised June 17, 2026.
15. NVIDIA Launches Space Computing, Rocketing AI Into Orbit - NVIDIA, March 16, 2026.
16. Solar Irradiance Science - NASA Goddard Space Flight Center.
17. Aeolus satellite - European Space Agency.
18. NIAC Phase I Final Report: Molecular Composition Analysis of Distant Targets - NASA Technical Reports Server, 2016.
19. State-of-the-Art of Small Spacecraft Technology: Thermal Control - NASA Small Spacecraft Systems Virtual Institute, May 7, 2026.
20. Orbital Data Centers: Spacecraft Constraints and Economic Viability - arXiv preprint, April 29, 2026.
21. Radiation-Hardness Components at scaled technology nodes - European Space Agency.
22. Delay-Tolerant Networking for Low Earth Orbit - NASA Technical Reports Server, 2009.
23. NASA's Record-Breaking Laser Demo Completes Mission - NASA, September 2024.
24. State-of-the-Art of Small Spacecraft Technology: Communications - NASA Small Spacecraft Systems Virtual Institute, 2026.
25. Smallsat Rideshare Program - SpaceX, accessed August 12, 2026.
26. Space Bureau Accepts for Filing SpaceX's Application for Orbital Data Centers - Federal Communications Commission, February 4, 2026.
27. Mitigation of Orbital Debris in the New Space Age, FCC 22-74 - Federal Communications Commission, September 29, 2022.
28. NASA-STD-8719.14C: Process for Limiting Orbital Debris - NASA, November 5, 2021.
29. ESA's Annual Space Environment Report, Issue 10.0 - European Space Agency Space Debris Office, May 1, 2026.
30. Investigating the Potential Atmospheric Accumulation and Radiative Impact of the Coming Increase in Satellite Reentry Frequency - Journal of Geophysical Research: Atmospheres, peer-reviewed article, 2025.
31. Radiative Forcing and Ozone Depletion of a Decade of Satellite Megaconstellation Missions - Earth's Future, peer-reviewed article, 2026.
32. The impact on astronomy of data centres orbiting Earth - Monthly Notices of the Royal Astronomical Society, peer-reviewed article, May 1, 2026.
33. Analytical simulations of the effect of satellite constellations on optical and near-infrared observations - Astronomy & Astrophysics, peer-reviewed article, 2022.
34. Confirmation of the Starmind name for SpaceX's AI-satellite constellation - Elon Musk on X, June 23, 2026.
35. SpaceX announcement of the Starmind AI1 partnership with NVIDIA - SpaceX on X, August 4, 2026.
36. Application summary for SAT-LOA-20260108-00016 - Federal Communications Commission ICFS, accessed August 13, 2026.
37. Applications accepted for filing, Report SAT-01982 - Federal Communications Commission, March 13, 2026.
38. 51,600 more satellites? Blue Origin adds another twist to the data center space race with Project Sunrise - GeekWire, March 21, 2026.
39. Starcloud-2 - Starcloud, accessed August 13, 2026.
40. Petition to deny SpaceX's orbital-data-center application - American Astronomical Society, March 6, 2026.
41. Coalition calls for rigorous federal review of space-based data center proposals - Earthjustice, July 8, 2026.
42. Why orbital data centers are harder than Silicon Valley thinks - IEEE Spectrum, June 11, 2026.
43. Forecast that orbital compute could become the only way to scale AI around 2029 - Elon Musk on X, August 14, 2026.
44. SpaceX EU prospectus - Space Exploration Technologies Corp., June 5, 2026.
45. Science & Tech Spotlight: Data Centers in Space - U.S. Government Accountability Office, GAO-26-109012, April 28, 2026.
46. The Cost and Network Limits of Space-Based AI Compute - Kees van Berkel, arXiv preprint, July 15, 2026.