Project Suncatcher is a Google research program studying whether machine-learning computing could one day run on clusters of solar-powered, laser-linked satellites carrying Google's Tensor Processing Unit (TPU) chips. Google announced it on November 4, 2025, with the Earth-imaging company Planet Labs as its spacecraft partner.[1][2] The first prototype, a single Planet-built satellite, lifted off on October 1, 2026, aboard a SpaceX Falcon 9 from California's Vandenberg Space Force Base, one of 130 payloads on the Transporter-18 rideshare,[3] and Google confirmed contact with it the same day.[4]

The program's design paper, first released as a preprint,[5] was published after peer review in the journal Joule on October 1, 2026.[6] Its worked example is a cluster of 81 satellites packed inside a radius of 1 km in a dawn-dusk sun-synchronous orbit.[6] In proton-beam tests reported in the paper, Google's Trillium TPUs survived a radiation dose equivalent to a five-year mission without permanent failures, and a learning-curve analysis suggests that launch prices to low Earth orbit could fall to about $200 per kilogram by the mid-2030s.[6] Google plans to fly two satellites in 2027 to test laser links between them.[7] The company calls Suncatcher a long-term research moonshot; it is one of several proposals for orbital data centers.[7][8]

Status as of October 2026

The first Suncatcher satellite, listed in SpaceX's deployment timeline as Project Suncatcher M1, separated from the Falcon 9 upper stage about an hour after the Transporter-18 launch at 11:32 a.m. PDT on October 1, 2026.[3] Planet made initial contact with it that day and began commissioning,[9] and Google said the spacecraft was operating as expected.[4] Orbital data from October 7 placed it in a sun-synchronous orbit of 511 by 516 km, inclined 97.44 degrees.[10] Over the following weeks Google intended to collect in-orbit measurements of how the chips tolerate spaceflight loads, radiation and temperature swings; no flight results had been published by mid-October.[4] The refereed version of the design paper appeared in Joule on launch day.[6] Google's next stated milestone is a pair of satellites, due in 2027, that will try out the inter-satellite laser links.[7]

Announcement and goals

Google introduced the program with a short post on its Keyword blog, a longer Google Research post by Travis Beals, senior director of the company's Paradigms of Intelligence team, and a preprint paper.[1][11] Google said its initial analysis had found no fundamental physical limit or insurmountable economic barrier to the idea, but listed heat rejection, high-capacity links to the ground and reliability in orbit as open engineering problems.[11] Its next milestone was to be a learning mission with Planet: two prototype satellites, launching by early 2027, to test TPU hardware in orbit and check whether optical links between satellites can carry distributed machine-learning work.[11] Beals told Ars Technica that the team had already spent more than a year on the project and that Google was bringing the computing payload to Planet, which, Ars reported, would build and test the spacecraft and arrange their launch.[8] AI Wiki's Project Suncatcher article covers the computing side: the TPUs, the machine-learning workloads and the costs Google modeled.

Planet said it would deploy the two prototypes to test the TPUs in space and to see whether two spacecraft could work together, flying in tandem over high-bandwidth crosslinks, on the same bus as Owl, the next-generation imaging fleet it had unveiled a month earlier.[2][12] By December 2025 it was citing a 2027 launch target.[13] Planet's quarterly report to the US Securities and Exchange Commission for the period ended July 31, 2026 does not use the Suncatcher name, but it describes a research and development agreement signed with Google in October 2025 to test "a payload integrated on prototype satellites", paid as milestones are met. Planet had received $5.0 million under it by January 31, 2026, and $9.5 million by July 31, 2026. The report identifies Google, which held more than 10 percent of Planet's Class A shares on July 31, 2026, as a related party.[14]

System concept

The design study has nine Google authors, and its contributions statement credits Blaise Agüera y Arcas with conceiving the project. It was posted to arXiv on November 22, 2025, and revised on June 17, 2026, after a re-run of the radiation test.[5] Joule received the manuscript on January 9, 2026, accepted it on August 18 and published it open access, with "Toward" in place of "Towards" in the title. The authors present their launch analysis as proof points rather than a full economic study.[6]

Orbit and cluster layout

The concept uses a dawn-dusk sun-synchronous orbit in low Earth orbit, chosen to balance solar power, latency to ground stations and launch cost, and for a gentler radiation environment than higher orbits. A panel there can gather up to eight times as much solar energy in a year as one at a mid-latitude site on the ground. Eclipses cannot be avoided entirely below about 1,400 km, so the operating altitude would trade eclipse time against crowded orbits, the delta-v needed to get there and the inner Van Allen belt, which begins near 1,000 km.[6]

Google prefers many small satellites in formation to one giant structure that would need assembly in orbit and be clumsier at avoiding collisions.[6] Its example is a flat cluster of 81 satellites on a square grid at a mean altitude of 650 km. The cluster's radius is 1 km, and the spacing between neighboring satellites swings between roughly 100 and 200 m as the formation changes shape twice per orbit.[6][5] Because such a formation is essentially two-dimensional, the number of satellites it can hold grows with the square of its radius.[6] The authors say satellites may eventually approach the computing capacity of one or more data-center racks, typically 50 to 100 kW today, which would probably require wider spacing.[6]

Suncatcher's tight spacing follows from its network needs. Commercial optical terminals between satellites carry 1 to 100 Gbit/s over thousands of kilometers; Starlink's run at about 100 Gbit/s over distances up to roughly 5,400 km. The paper estimates that each link in a cluster would need on the order of 10 Tbit/s.[6] Off-the-shelf transceivers of the kind used in ground data centers could supply that, but only with hundreds of microwatts of received optical power, against about 1 microwatt in typical long-range links. Because received power falls with the square of distance, close formation flying closes the gap; one aperture carrying 24 channels of 400G transceivers would provide 9.6 Tbit/s of bidirectional bandwidth.[6] A bench demonstrator built from off-the-shelf parts sent 800 Gbit/s each way, 1.6 Tbit/s in total, through one pair of transceivers over a short free-space path.[11][6]

TPU radiation testing

Google exposed a Trillium (v6e) Cloud TPU and its AMD host server to 67 MeV protons from the 76-inch cyclotron at the University of California, Davis, Crocker Nuclear Laboratory, in what the authors call the first published radiation test of such a device.[6] Behind about 10 mm of aluminum-equivalent shielding, the target orbit would deliver roughly 150 rad(Si) a year, or about 750 rad(Si) over five years.[6]

The chip's high-bandwidth memory was the part most sensitive to accumulated dose, showing irregularities after 2 krad(Si). Other tests, including an end-to-end transformer workload, ran correctly to the highest dose applied to one chip, given as 10 krad(Si) in Joule and 15 krad(Si) in the preprints, with no hard failure traced to total dose.[6][5] In typical transformer workloads, silent data corruption from single particle strikes occurred about once every 17 rad, which in orbit works out to about one error in every 3 million inferences if the chip ran one inference a second. The authors consider that rate likely acceptable for inference, but its effect on training still needs study.[6]

Launch-cost assumptions

The paper calls launch critical to the system's overall cost. SpaceX price and mass data from Falcon 1, at $30,000/kg or more in inflation-adjusted terms, to Falcon Heavy, at about $1,800/kg, give a learning rate of about 20 percent: the price per kilogram falls by about a fifth with every doubling of the total mass launched.[6] Sustained, that rate could bring prices below $200/kg around 2035, after about 370,000 t more had been launched, equivalent to some 1,800 Starship flights, or about 180 a year.[6] A separate estimate of SpaceX's own Starship costs, based on its reusable rocket targets, ranges from about $460/kg without reuse to under about $15/kg if every component flies 100 times.[6] Expressed as the annual launch cost per kilowatt of solar generating capacity, a Starlink V2 Mini-class satellite comes to $14,700 at the current reusable Falcon 9 price of $3,600/kg and $810 at $200/kg, against reported US data-center power spending of $570 to $3,000 per kilowatt-year. The comparison covers power only, not buildings or chips.[6]

Prototype mission with Planet

Spacecraft

According to Ars Technica, the satellite that flew in October 2026 was not part of the original plan, which called for two custom satellites in 2027. Google wanted to move faster, so it put its chips into an existing Planet satellite that had been built for an early test. The pair of 2027 satellites is still planned.[15] Ars reported that the spacecraft, named MVP, is roughly refrigerator-sized, carries four TPUs and has solar panels supplying only about 1 kW, a figure it credited to The New York Times.[15] Google's own post on X on launch day confirmed the count of four TPUs.[24] Jonathan McDowell's catalog of space objects also lists the satellite as MVP.[10] Ars described a soft thermal interface material between the chips and heat pipes of aluminum and copper, which move the heat into a radiator; the TPUs can run for about 15 minutes before shutting down to let the radiator catch up.[15] Reports differ on the mission's length: Ars said a few months,[15] while Space.com, citing The New York Times, reported a planned year in orbit.[16]

Google's team shook the satellite along all three axes to mimic launch vibration and tested its heat pipe and radiator cooling approach in a thermal vacuum chamber.[7] The company noted that the roughly ten-minute climb to low Earth orbit brings sustained loads of up to 10 g, and as much as 50 to 100 g on parts such as the TPU chips.[7]

Launch on Transporter-18

Transporter-18, a dedicated SpaceX smallsat rideshare, lifted off from Space Launch Complex 4E at Vandenberg at 11:32 a.m. PDT (18:32 UTC) on October 1, 2026, carrying 130 payloads, 41 of them aboard four orbital transfer vehicles for release later.[3][17] The first stage, flying for the 25th time, landed at Landing Zone 4 at Vandenberg about seven and a half minutes after liftoff.[3][18][17] SpaceX's approximate timeline shows Project Suncatcher M1, manifested by Planet, deploying 59 minutes 55 seconds into the flight.[3] Planet had 20 satellites on board: the Suncatcher prototype, its Tanager-2 hyperspectral satellite and 18 SuperDoves.[9] The prototype was placed in a sun-synchronous orbit of 511 by 516 km at 97.44 degrees, lower than the 650 km mean altitude of the paper's example cluster.[10][6]

Contact and test objectives

Planet said it had made initial contact with the Suncatcher prototype, Tanager-2 and two of the SuperDoves and had begun commissioning; it added that the satellite would run the first-ever test of Google's TPUs in space.[9] Google said on launch day that the satellite was operating as expected.[4] The flight is meant to show how the TPUs cope with launch stresses, radiation and the temperature extremes of orbit, and to expose points of failure; Google says some of these effects can only be tested in space.[7][4]

DateEvent
October 2025Planet signs a research and development agreement with Google to test a payload on prototype satellites[14]
November 4, 2025Project Suncatcher is announced with a preprint; Planet plans two prototypes for launch by early 2027[1][2]
August 18, 2026Joule accepts the revised paper after peer review[6]
September 24, 2026Google announces that a prototype will fly on Transporter-18[7]
October 1, 2026Prototype launches and is contacted; Joule publishes the paper[3][4][9][6]
2027 (planned)Two satellites to test inter-satellite laser links[7]

Planned two-satellite test in 2027

The 2027 flight is meant to test what a single satellite cannot: the laser links. Google says future Suncatcher satellites would each carry dozens of TPUs and fly in clusters, each spacecraft needing to know its position and its place relative to the others. They would communicate over lasers built for very high bandwidth across very short distances, the reverse of today's long-range, lower-rate terminals. Google compared the pointing precision needed to striking a coin-sized target miles away while both ends move; its September 2026 update dates the test only to 2027.[7] Planet's descriptions call for two spacecraft flying in tandem over high-bandwidth crosslinks on the same bus as its Owl satellites,[2][13] and Space.com, citing The New York Times, reported that Planet would build both.[16]

Engineering constraints

Power and heat

Google calls cooling a central research challenge: TPUs pack a lot of heat into a small area, and with no air in space the heat can leave only through radiators.[7] The Joule paper designs its cooling around keeping the high-bandwidth memory, the limiting part, below about 90 C. It notes that near a radiator temperature of 50 C, radiated power drops by about 1.2 percent for each kelvin of cooling, so the temperature step from the compact chips to the radiators must be small.[6] The prototype is far from the concept's scale: by the reported figures it has about 1 kW of solar power and radiator-limited TPU runs of about 15 minutes,[15] while the paper foresees satellites that might approach the 50 to 100 kW of a current AI rack.[6]

Radiation

Google's ground tests leave open how its TPUs will fare over years in orbit. The test beam reached the die degraded in energy and mixed with secondary particles, the effect of silent data corruption on training still needs study, and because failed TPUs cannot be swapped in orbit, the paper proposes redundant hardware as the simplest answer.[6] Domenico Vicinanza of Anglia Ruskin University wrote in The Conversation that sustaining reliable performance for years, through solar storms and swings in temperature, would be a much tougher test than the laboratory exposure.[19]

Formation flying

The Suncatcher paper says its system would need much closer formation flight than any constellation so far. It starts from the Hill-Clohessy-Wiltshire equations for relative motion, adds Earth's oblateness, which is also what keeps the orbit sun-synchronous, and refines the result numerically.[6] Changing the cluster's axis ratio from 2:1 to 2:1.0037 cuts oblateness-driven drift to under 3 m/s per year for each kilometer of maximum distance from the reference orbit. A simulation added in the Joule version, which includes navigation errors, differential drag, thruster limits and stormy space weather, finds that holding neighbor distances within 20 percent of their targets would take under 30 m/s of delta-v a year, given careful implementation.[6] Mojtaba Akhavan-Tafti, a space scientist at the University of Michigan, argued in December 2025 that drag and space-weather changes in air density leave almost no margin for error at spacings under 200 m.[20]

Debris

Space debris drew early criticism. Akhavan-Tafti called sun-synchronous orbit the most congested part of low Earth orbit, estimated that a Suncatcher cluster would meet a fragment larger than a grain of sand about every five seconds, and warned that one impact could spread through a cluster and scatter new debris. He wrote that the design did not yet include the autonomous avoidance such a cluster would need.[20] The Joule paper, published ten months later, says a working constellation would need automated avoidance of hazardous debris based on conjunction alerts and the ability to deorbit satellites as regulations require. It names the cross-track direction as the natural axis for dodging and says very large clusters might have to stack one section temporarily behind another.[6] The US Government Accountability Office's April 2026 spotlight on data centers in space listed collision risk and interference with astronomical research among the challenges of adding satellites to orbit.[21]

Astronomy and environment

The Joule paper does not estimate how bright the satellites or clusters would look from the ground, and it leaves a full life-cycle assessment to future work, noting that launch, reentry and satellite demise have no terrestrial counterpart.[6] Suncatcher's dawn-dusk orbit follows the line between day and night.[8] Astronomer Samantha Lawler of the University of Regina, interviewed by Scientific American in December 2025 for an article on orbital data centers in general, noted that some observers depend on twilight to search for near-Earth asteroids.[22]

Reception

Writing the day after the announcement, Ars Technica's Stephen Clark noted Google's mixed record with moonshots, from Waymo, now an operating company, to the failed Project Loon internet balloons. He contrasted Google's fleet of small, laser-linked satellites with the 5 GW orbital data center that Starcloud and NVIDIA had described, whose solar and cooling panels would measure about 4 km on a side.[8] Vicinanza judged a two-satellite test by 2027 plausible but wrote that even a successful demonstration would not show that large orbital data centers are feasible, and that any adoption would probably unfold over decades.[19]

When Google announced the October flight, Ars Technica reported that the team expected it to be years before Suncatcher moved from project to product.[15] Elon Musk, whose SpaceX launched the prototype and is planning its own orbital computing constellation, Starmind,[16] wrote that day that the share of computing done in space would obviously approach all of it.[23] The Joule paper calls its results a first milestone. As of mid-October 2026, neither close formation flying nor high-rate laser links between Suncatcher satellites had been tried in orbit.[6][7]

See also

References

  1. Meet Project Suncatcher, a research moonshot to scale machine learning compute in space - Google, November 4, 2025.
  2. Planet to Build and Operate Advanced Space Platform for Google's Project Suncatcher Moonshot - Planet Labs PBC, November 4, 2025.
  3. Transporter-18 Mission - SpaceX, October 1, 2026.
  4. Our Project Suncatcher prototype satellite is in orbit - Travis Beals, Google, October 1, 2026.
  5. Towards a future space-based, highly scalable AI infrastructure system design - B. Agüera y Arcas et al., arXiv preprint 2511.19468, version 1 November 22, 2025, version 2 June 17, 2026.
  6. Toward a future space-based, highly scalable AI infrastructure system design - B. Agüera y Arcas et al., Joule, article 102678, peer-reviewed article published online October 1, 2026.
  7. Behind Project Suncatcher, our moonshot to put AI in space - Travis Beals, Google, September 24, 2026.
  8. If you want to satiate AI's hunger for power, Google suggests going to space - Stephen Clark, Ars Technica, November 5, 2025.
  9. Planet Launches Suncatcher, Tanager-2, and 18 SuperDove Satellites - Planet Labs PBC, October 1, 2026.
  10. GCAT current catalog - Jonathan McDowell, General Catalog of Artificial Space Objects, updated October 15, 2026.
  11. Exploring a space-based, scalable AI infrastructure system design - Travis Beals, Google Research, November 4, 2025.
  12. Introducing Owl: Planet's Most Advanced Satellite Mission Yet - Will Marshall, Planet Labs PBC, October 7, 2025.
  13. Planet Reports Financial Results for Third Quarter of Fiscal Year 2026 - Planet Labs PBC, exhibit 99.1 to a Form 8-K filed with the US Securities and Exchange Commission, December 10, 2025.
  14. Quarterly report, Form 10-Q for the quarter ended July 31, 2026 - Planet Labs PBC, filed with the US Securities and Exchange Commission, September 3, 2026.
  15. Google's first Suncatcher orbital data center test launches October 1 - Ryan Whitwam, Ars Technica, September 24, 2026.
  16. SpaceX launching prototype Google AI satellite next week - Mike Wall, Space.com, September 25, 2026.
  17. SpaceX launches Google AI satellite, 129 other payloads on 2nd leg of spaceflight tripleheader (video) - Mike Wall, Space.com, October 1, 2026.
  18. Falcon 9 • Transporter-18 - Will Robinson-Smith, Spaceflight Now, September 28, 2026, updated after the October 1, 2026 launch.
  19. Data centres in space: will 2027 really be the year AI goes to orbit? - Domenico Vicinanza, Anglia Ruskin University, The Conversation, December 3, 2025.
  20. Google's proposed data center in orbit will face issues with space debris in an already crowded orbit - Mojtaba Akhavan-Tafti, University of Michigan, The Conversation, December 3, 2025.
  21. Science & Tech Spotlight: Data Centers in Space - U.S. Government Accountability Office, GAO-26-109012, April 28, 2026.
  22. Data Centers in Space Aren't as Wild as They Sound - Jeremy Hsu, Scientific American, December 9, 2025.
  23. Post predicting that compute in space will round up to all compute - Elon Musk on X, September 24, 2026.
  24. Launch-day post on the Project Suncatcher prototype satellite - Google on X, October 1, 2026.