Starmind

Starmind is SpaceX's proposed orbital artificial-intelligence computing program. Its first spacecraft design, AI1, is currently specified at 30 m high with a 75 m wingspan and carries a computing payload rated at up to 250 kW at peak and 175 kW on average, linked to the Starlink network by laser.^[1] SpaceX presents the system as a way to run AI workloads in sun-synchronous orbit and return processed results to Earth; the published material describes a design and a production plan, not a demonstrated orbital service.^[1] No Starmind satellite had been launched by mid-October 2026. SpaceX says it plans to begin launches in 2027 and to scale the constellation to as many as 1 million satellites,^[25] and Elon Musk has set the fourth quarter of 2027 as the launch target for a space-optimized NVIDIA Vera Rubin NVL72 system.^[22]

AI1 is older than the Starmind name. SpaceX first showed the satellite in a technical update video posted on June 8, 2026, in which its director of satellite engineering described a starting design of about 150 kW of peak power and about 120 kW of average compute.^[27] Musk confirmed on June 23, 2026 that Starmind would be the name of the company's AI-satellite constellation, one day after X.AI LLC filed a US trademark application covering satellites, orbital computing, and data-center hardware.^[2][3] On August 4, SpaceX announced that it was working with NVIDIA on the AI1 payload and that the satellites would include NVIDIA Rubin GPUs and Vera CPUs,^[4] and archived copies of the program page show its AI1 figures rising to their current values between August 3 and August 4.^[28][29]

The regulatory record began under a generic name. In January 2026, Space Exploration Holdings applied to the US Federal Communications Commission for a "SpaceX Orbital Data Center System" of up to 1,000,000 satellites. The application was still pending review, with no grant date, on November 14, 2026.^[5] Neither the application nor the FCC public notice uses the names Starmind or AI1,^[6] but SpaceX's own later statements do. Its September 16, 2026 answers to FCC staff questions about the application refer to "its Starmind system of space-based compute", and its Starlink website said on October 6 that SpaceX plans to operate up to 1 million Starmind satellites.^[24][25] The 1,000,000 figure is still the ceiling of an authorization request rather than a deployment commitment, and no SpaceX filing says that every proposed satellite would use the AI1 design.

Status as of November 2026#

Starmind had no satellites in orbit in mid-November 2026, and SpaceX's October 6 space-safety page said deployment would begin in 2027.^[25] NVIDIA's August 24 announcement said the first generation of Starmind satellites would be based on an optimized form of its Vera Rubin NVL72 rack,^[21] and Musk wrote the same day that this space version was due to launch in the fourth quarter of 2027, with "significant scale in 2028".^[22] At the FCC, the Space Bureau asked SpaceX on August 28 for information it said the application lacked, including orbital-debris showings.^[23] SpaceX replied on September 16 with a mass ceiling of 4,000 kg per satellite and a plan for targeted reentry over the ocean; the application was still pending, with no grant date, on November 14.^[24][5] SpaceX's program page still listed AI1 at up to 250 kW peak with a 210 kW solar array on that date, although Musk wrote on October 1 that each spacecraft would have more than 250 kW of solar power.^[1][26]

Naming and corporate context#

SpaceX was already discussing orbital AI compute before the Starmind name appeared. Its 2026 prospectus says that the company completed its acquisition of X.AI Holdings on February 2 and describes orbital computing as part of the combined group's AI strategy.^[7] The prospectus does not use the Starmind name. Neither does SpaceX's first quarterly report as a public company, filed on August 4, 2026, although it discusses the AI business at length.^[19] AI Wiki's Starmind article covers the computing side of the program, including the NVIDIA hardware and SpaceX's AI business.

The naming record developed in seven steps:

The trademark record is an application, not a registration. As of August 12, the US Patent and Trademark Office listed it as live and awaiting examination.^[3]

AI1 design#

SpaceX's Starmind page publishes the following values for AI1. On November 14, 2026 they were the same figures the page first showed on August 4. They are company design figures, not measurements from flight hardware.^[1][29]

Taken together, the published figures describe a satellite that generates 210 kW from its solar array, draws up to 250 kW at peak in the computing payload, and rejects waste heat through a 160 m² deployable liquid radiator.^[1]

The design description leaves several parameters open. SpaceX has not published AI1's exact mass, energy-storage capacity, expected lifetime, radiation-test results, laser throughput, ground-network capacity, unit cost, or a launch manifest. The webpage also does not explain the peak-mode power budget or state how long the 250 kW peak compute mode could be sustained from a 210 kW solar array.^[1] SpaceX's September 2026 FCC filing gave only outer limits for Starmind satellites: a mass under 4,000 kg, a total solar-array area under 1,500 m², a radiator area under 200 m², and a bus area under 21 m².^[24] On October 1, Musk wrote that each Starmind spacecraft would have more than 250 kW of solar power, a figure the program page had not adopted by October 16.^[26][1]

NVIDIA relationship#

NVIDIA chips were part of the AI1 plan from its first public showing. In the June 8 technical update, Musk said the current reference design used NVIDIA Rubin chips, or possibly GB300 chips, and that SpaceX would also have a reference design for TPUs.^[27] SpaceX says NVIDIA is its partner in designing the AI1 compute payload and that Starmind satellites will include Rubin GPUs and Vera CPUs.^[4] Its program page also says the wider architecture is vendor agnostic and can accept modules from any provider.^[1] Read together, those statements establish NVIDIA hardware for the announced AI1 configuration without establishing an exclusive supplier relationship for every future design. On SpaceX's earnings call of August 4, 2026, Musk went further: he described the Starmind satellite as an optimized version of NVIDIA's Vera Rubin NVL72 system, a simplification of the standard rack design, and said SpaceX had decided to build its AI compute exclusively on NVIDIA hardware.^[20]

NVIDIA described more of the arrangement on August 24, 2026. Its announcement was mainly about SpaceXAI deploying Vera CPUs for agentic AI work on the ground, but it also said the first Starmind satellite generation would use an optimized version of the Vera Rubin NVL72 rack-scale system. The two companies were adapting that platform to orbit, where power, cooling, data bandwidth, reliability and physical integration differ from conditions in ground data centers.^[21] The release gave no launch date, GPU count or power figure. Musk posted the same day that a space-optimized version of the system, designed with NVIDIA, was due to launch in the fourth quarter of 2027.^[22]

The public record still does not identify AI1 as NVIDIA's Space-1 Vera Rubin module or as a complete, unmodified Vera Rubin NVL72 rack; NVIDIA calls the planned orbital system optimized without saying what changed.^[21] NVIDIA announced Space-1 for orbital computing in March 2026, but its announcement did not name SpaceX among the participating space companies.^[8] SpaceX's AI1 announcement names processor families, not the Space-1 module. No public first-party source gives the number of GPUs or CPUs per satellite, a purchase quantity, a price, or a flight-qualified hardware configuration.^{[1][4][8][21]}

Starmind's published plans have changed#

SpaceX's first-party planning figures for Starmind changed several times between June and October 2026. In the June 8 technical update, SpaceX's director of satellite engineering said the AI1 design would start at about 150 kW of peak power and could support about 120 kW of average compute, and Musk set a goal of adding space AI compute at an annualized rate of about a gigawatt by the end of 2027.^[27] The prospectus discussed generic AI-compute satellites with about 100 kW of compute for early units, a target of about 100 kW per metric ton, and deployment beginning as early as 2028.^[7] A capture of the Starmind page from August 3 still listed AI1 at 150 kW peak, 120 kW average, and 70 kW per metric ton, with a 150 kW solar array.^[28] A capture taken at 21:02 UTC on August 4, less than two hours after SpaceX posted its NVIDIA announcement, shows the current figures: up to 250 kW of compute at peak, 175 kW on average, and 75 kW per metric ton.^[29][4] The page also says SpaceX's planned Gigasat Factory in Bastrop, Texas should let the company build and launch thousands of AI satellites from as soon as late 2027.^[1]

These are successive planning disclosures, not specifications for hardware already built. The Starmind page is still the newest full set of AI1 design figures. It gives a broad factory and deployment goal but does not publish an individual AI1 launch date, identify prototype spacecraft, or supply a flight schedule.^[1] Musk's fourth-quarter 2027 target, posted on X, is narrower than anything on the page.^[22]

FCC application#

The FCC application for the SpaceX Orbital Data Center System provides the most detailed public regulatory description of the company's orbital-computing concept, but it is not an AI1 specification. Space Exploration Holdings filed it on January 30, 2026 under call sign S00798. The Space Bureau accepted it for filing on February 4, opening a comment period. Acceptance for filing is a procedural finding that allows review; it is not permission to deploy or operate the system.^[5][6]

The orbital and communications details in the original filing come from the FCC public notice, SpaceX's application narrative, and its technical attachment.^[6][9][10] The waiver document explains why SpaceX asked not to be bound by the usual requirement to deploy half the authorized satellites within six years and the remainder within nine years.^[11] None of those documents contains AI1's published height, wingspan, radiator area, solar output, or computing hardware. SpaceX's later supplements set outer limits for the satellites but still do not name AI1 or its computing hardware.^[30][24]

The application as filed did not assign a population to each precise shell or disclose a final spacecraft mass, orientation, material set, service life, replacement cadence, or disposal split. SpaceX instead asked for a broad design space in which hardware can vary by shell.^[9][10] Supplements requested by FCC staff have since narrowed some of those gaps. On May 29, SpaceX listed six groups of shells with a maximum satellite count for each, said it would start with the shells below 600 km, and said it planned controlled reentry for satellites below that altitude while asking for a waiver to use other disposal methods above it.^[30] Its September 16 answers added the mass and area limits, put each listed shell's probability of collision with large debris below the FCC's 1-in-1,000 threshold, and estimated that propulsive deorbit from 585 km would take less than 60 days.^[24] Accordingly, the one-million figure is the upper bound of a pending authorization request. It is not an authorization, a manufacturing order, a launch commitment, or evidence that one million identical AI1 satellites will be built.

The September answers responded to an August 28 letter from the Space Bureau, which said the additional information was needed "to render the Application suitable for further processing". Its questions covered how many satellites would use the backup Ka-band radio links at any one time, whether first- and second-generation Starlink satellites could support optical links for a million-satellite system, satellite dimensions, collision and casualty-risk showings, and whether reentry would be targeted or uncontrolled.^[23] SpaceX met bureau staff on September 14 and filed its answers two days later. It said Starmind would use radio links only as a last resort if its optical links were unavailable, and it did not supply the requested Starlink optical-link analysis, arguing that the FCC has no jurisdiction over optical inter-satellite links. SpaceX also asked the commission to "swiftly authorize" the application.^[24]

Dependencies and engineering limits#

Starmind relies on three SpaceX systems that are themselves still being scaled. The company says Starship is critical because AI satellites are large and heavy, Starlink's laser mesh would carry results to ground stations, and the planned Gigasat Factory would manufacture satellites at high volume.^[1] SpaceX's prospectus similarly says commercial orbital compute depends on Starship's payload capacity, reusability, and flight rate. It treats delays in Starship or satellite production as business risks, not solved assumptions.^[7]

Sun-synchronous orbit is the power strategy. SpaceX says high-altitude sun-synchronous shells could remain illuminated more than 99 percent of the time, reducing battery use and power cycling.^[9] That is a company projection for parts of the FCC orbital envelope. It does not mean every allowed orbit between 500 and 2,000 km has the same illumination, and the application does not select a final AI1 altitude.

Heat rejection is a separate constraint. SpaceX's 160 m² radiator, active fluid loops, and micrometeoroid protection are a concrete design response, but no on-orbit AI1 thermal result has been published.^[1] A 2026 engineering preprint on orbital data centers concluded that feasibility depends on closing power generation, eclipse storage, radiative cooling, communications, utilization, lifetime, launch cost, and spacecraft-build cost together. It identified edge processing of space-generated data as an easier early use case than general computing for terrestrial users.^[12] The paper is a preprint, not a peer-reviewed validation of AI1, and it did not analyze the current Starmind design.

SpaceX's prospectus is unusually direct about the remaining uncertainty. It says neither SpaceX nor anyone else had previously operated or attempted orbital AI compute and lists radiation, thermal cycling, micrometeoroids, space debris, component failure, and the difficulty of repair or upgrade among the risks.^[7] SpaceX has extensive experience with satellite avionics and laser links, but that experience is not a flight demonstration of a 175 kW average AI payload.

Astronomy, debris, and atmospheric review#

SpaceX's FCC technical attachment promises redundant propulsion, automated conjunction screening, and collision-avoidance maneuvers throughout a satellite's active life. It proposes a mix of atmospheric reentry, disposal orbits around Earth, and heliocentric disposal. The attachment says exact disposal-orbit altitudes have not been selected and that a failed spacecraft could reenter without control if a targeted reentry is not possible.^[10] These are proposed mitigations. Their system-wide performance has not been demonstrated for the requested scale.

The American Astronomical Society filed a petition to deny the application on March 6, citing possible optical, infrared, radio, orbital-debris, and atmospheric effects.^[13] Its optical simulations should not be mistaken for measurements or for a model of the current AI1 design. The petition explicitly says SpaceX had not supplied exact sizes, orientations, materials, or precise orbits. Its model therefore used inferred orbits and an assumed 700 m² satellite area.^[13] SpaceX's September 2026 filing later set the outer limit for solar-array area at 1,500 m², with radiators under 200 m².^[24] Amazon and Viasat filed their own petitions to deny on March 6, and SpaceX answered all of the petitions in a consolidated opposition on March 16. That filing said brightness mitigation was a core design criterion, with the goal of meeting the visual magnitude 7 threshold advocated by the National Science Foundation, the Vera C. Rubin Observatory, and the International Astronomical Union.^[35] The petitions were still pending with the application on October 16, 2026.^[5]

The underlying concerns are supported by research on existing constellations, although those studies do not predict Starmind's exact impact. Peer-reviewed optical simulations show that interference varies with satellite altitude, inclination, brightness, time of night, season, latitude, and instrument, and that no single mitigation removes trails from every observing program.^[14] Radio observations with LOFAR detected unintended electromagnetic radiation from 47 of 68 first-generation Starlink satellites in one study.^[15] A later study found that second-generation Starlink satellites emitted unintended radiation up to 32 times stronger than the first generation at the measured frequencies.^[16] Those measurements concern Starlink hardware, not AI1; they establish a design issue to test rather than a measured Starmind emission level.

Atmospheric effects are likewise real but not yet quantifiable for Starmind. A 2023 peer-reviewed study measured spacecraft-derived metals in about 10 percent of sampled stratospheric sulfuric-acid particles larger than 120 nanometers and said the effects of increased metallic content remain unknown.^[17] A 2025 modeling study examined a hypothetical 10 gigagram-per-year alumina injection and found outcomes depended strongly on particle size and reentry latitude; it also stressed that actual reentry aerosol composition and size remain poorly characterized.^[18] Because the FCC record gives no final spacecraft mass and does not publish materials, operating life, or reentry cadence, neither result can be converted into a defensible Starmind-specific atmospheric estimate.

The environmental dispute continued after SpaceX's September filing. The Center for Space Environmentalism responded on September 22 that a million satellites at the 4,000 kg ceiling would add up to about 4 million metric tons of hardware. It asked the FCC to deny SpaceX's waiver requests, require a full environmental impact statement under the National Environmental Policy Act, and hold off on authorization until SpaceX publishes complete specifications for its optical links and collision avoidance along with verified demise testing.^[31] DarkSky International, Environment America, and Public Employees for Environmental Responsibility had already petitioned on July 8 for a programmatic environmental impact statement covering all pending orbital data-center applications.^[34] SpaceX met Space Bureau staff again on October 5 to rebut what it called misleading claims from commenters on the Starmind record. Its follow-up letter said SpaceX designs its satellites to burn up completely on reentry, that its targeted-reentry trials with Starlink satellites started in December 2024, and that more than 85 percent of those reentries had since come down over the ocean.^[32]

Achieved and planned#

Starmind's documented record separates a short list of completed steps from a much longer list of plans.

See also#

• Starcloud
• Project Suncatcher

References#

1. Starmind - SpaceX, accessed November 14, 2026.
2. Confirmation of the Starmind name for SpaceX's AI-satellite constellation - Elon Musk on X, June 23, 2026.
3. STARMIND trademark status, US serial number 99898550 - US Patent and Trademark Office, generated August 12, 2026.
4. SpaceX announcement of the Starmind AI1 partnership with NVIDIA - SpaceX on X, August 4, 2026.
5. Application summary for SAT-LOA-20260108-00016 - Federal Communications Commission ICFS, accessed November 14, 2026.
6. Space Bureau accepts for filing SpaceX's application for orbital data centers, DA-26-113 - Federal Communications Commission, February 4, 2026.
7. Final prospectus, Form 424B4 - Space Exploration Technologies Corp., filed with the US Securities and Exchange Commission, June 12, 2026.
8. NVIDIA launches space computing, rocketing AI into orbit - NVIDIA Newsroom, March 16, 2026.
9. Orbital Data Center System application narrative - Space Exploration Holdings, filed with the Federal Communications Commission, January 2026.
10. Orbital Data Center System technical attachment - Space Exploration Holdings, filed with the Federal Communications Commission, January 2026.
11. Orbital Data Center System waiver requests - Space Exploration Holdings, filed with the Federal Communications Commission, January 2026.
12. Orbital Data Centers: Spacecraft Constraints and Economic Viability - Slava G. Turyshev, arXiv preprint, April 29, 2026.
13. Petition to deny SpaceX's orbital-data-center application - American Astronomical Society, March 6, 2026.
14. Analytical simulations of the effect of satellite constellations on optical and near-infrared observations - C. G. Bassa, O. R. Hainaut, and D. Galadi-Enriquez, Astronomy & Astrophysics 657, A75, 2022.
15. Unintended electromagnetic radiation from Starlink satellites detected with LOFAR between 110 and 188 MHz - F. Di Vruno et al., Astronomy & Astrophysics 676, A75, 2023.
16. Bright unintended electromagnetic radiation from second-generation Starlink satellites - C. G. Bassa et al., Astronomy & Astrophysics 689, L10, 2024.
17. Metals from spacecraft reentry in stratospheric aerosol particles - D. M. Murphy et al., Proceedings of the National Academy of Sciences 120, e2313374120, 2023.
18. Investigating the Potential Atmospheric Accumulation and Radiative Impact of the Coming Increase in Satellite Reentry Frequency - C. M. Maloney et al., Journal of Geophysical Research: Atmospheres 130, e2024JD042442, 2025.
19. Quarterly report, Form 10-Q for the quarter ended June 30, 2026 - Space Exploration Technologies Corp., filed with the US Securities and Exchange Commission, August 4, 2026.
20. Earnings call transcript: SpaceX beats revenue estimates in Q2 2026, shares swing - Investing.com, August 4, 2026, accessed August 13, 2026.
21. SpaceXAI Adopts NVIDIA Vera CPU to Accelerate Agentic AI at Massive Scale - NVIDIA, August 24, 2026.
22. Launch target for a space-optimized Vera Rubin NVL72 system - Elon Musk on X, August 24, 2026.
23. Letter requesting additional information on ICFS File No. SAT-LOA-20260108-00016 - Federal Communications Commission Space Bureau, August 28, 2026.
24. Ex parte response to Space Bureau questions on the Starmind system - Space Exploration Technologies Corp., filed with the Federal Communications Commission, September 16, 2026.
25. Space Safety: The Need for Improved Data Sharing Between Operators in Low-Earth Orbit - Starlink, SpaceX, October 6, 2026.
26. Solar power of each Starmind spacecraft - Elon Musk on X, October 1, 2026.
27. Free writing prospectus with the transcript of SpaceX's June 8, 2026 AI-satellite technical update - Space Exploration Technologies Corp., filed with the US Securities and Exchange Commission, June 9, 2026.
28. Starmind, archived August 3, 2026 - SpaceX, captured by the Internet Archive, August 3, 2026.
29. Starmind, archived August 4, 2026 - SpaceX, captured by the Internet Archive, August 4, 2026.
30. Supplement on backup Ka-band use, orbital configuration, and orbital debris - Space Exploration Technologies Corp., filed with the Federal Communications Commission, May 29, 2026.
31. Response to ex parte presentation and supplemental comments - Center for Space Environmentalism, filed with the Federal Communications Commission, September 22, 2026.
32. Ex parte letter on targeted reentry - Space Exploration Technologies Corp., filed with the Federal Communications Commission, October 7, 2026.
33. Reply on renaming SpaceXAI as SpaceXSI - Elon Musk on X, October 4, 2026.
34. Petition to initiate a programmatic environmental impact statement on orbital data center applications - DarkSky International, Environment America, and Public Employees for Environmental Responsibility, filed with the Federal Communications Commission, July 8, 2026.
35. Consolidated opposition to petitions and response to comments - Space Exploration Holdings, filed with the Federal Communications Commission, March 16, 2026.