Project Suncatcher

Project Suncatcher is a long-term research "moonshot" at Google that studies whether machine learning compute could one day be hosted in space, on constellations of solar-powered satellites carrying Google Tensor Processing Units (TPUs) and linked to one another by free-space optical (laser) links.^[1][2] Google announced the project on November 4, 2025, together with a preprint paper, "Towards a future space-based, highly scalable AI infrastructure system design," which describes an illustrative 81-satellite cluster flying in a dawn-dusk sun-synchronous orbit and reports proton-beam radiation tests of Trillium (TPU v6e) chips.^[2][3] Planet Labs PBC is Google's spacecraft partner.^[5] On September 24, 2026, Google said the first Suncatcher prototype satellite, carrying TPUs, would launch "next week" on SpaceX's Transporter-18 rideshare mission; SpaceX is targeting October 1, 2026 for the launch, from Vandenberg Space Force Base.^[15][20] Google says a two-satellite mission to test the inter-satellite laser links follows in 2027.^[7]

Suncatcher is one of several efforts to build space-based data centers. Google itself frames it as early research: the company says the core concepts are "not precluded by fundamental physics or insurmountable economic barriers" but that significant engineering problems remain, including thermal management, high-bandwidth ground communications and on-orbit reliability.^[2]

Background and announcement

Google published three coordinated pieces on November 4, 2025: a short post on The Keyword, "Meet Project Suncatcher, a research moonshot to scale machine learning compute in space"; a longer post on the Google Research blog by Travis Beals, Senior Director of Paradigms of Intelligence, titled "Exploring a space-based, scalable AI infrastructure system design"; and the preprint paper.^[1][2][4] The same day, Planet announced it would build and operate two prototype satellites for the project, and Google CEO Sundar Pichai posted about it on X.^[5][6]

The pitch rests on solar power. The Research blog says the Sun emits "more power than 100 trillion times humanity's total electricity production," and that "in the right orbit, a solar panel can be up to 8 times more productive than on earth," producing power nearly continuously and reducing the need for batteries.^[2] The paper places that 8x figure against a panel on Earth at mid-latitude.^[3] Rather than beaming power down to the ground, as classic space-based solar power proposals would, Suncatcher proposes to use the power in orbit, running ML workloads on satellites and sending results down.^[3]

Google presents the project as part of a lineage of long-horizon bets. The Research blog compares it with the company's work on a large-scale quantum computer "a decade ago" and its autonomous-vehicle program begun "over 15 years ago, which eventually became Waymo."^[2] Pichai wrote on X: "Like any moonshot, it's going to require us to solve a lot of complex engineering challenges," adding that "significant challenges still remain like thermal management and on-orbit system reliability."^[6]

In an interview with Ars Technica published November 5, 2025, Beals said Google had worked on Suncatcher for more than a year, and described the motivation as demand: "We're just seeing so much demand from people for AI," he said. "So, we wanted to figure out a solution for compute that could work no matter how large demand might grow."^[11]

Organization and people

The public posts about Suncatcher are published under Google Research, and both its Research blog post (November 2025) and its update (September 2026) are written by Travis Beals, whose title is given as Senior Director, Paradigms of Intelligence.^[2][7] Ars Technica describes Paradigms of Intelligence as "a research team within Google."^[11] Google's researcher profile for Blaise Agüera y Arcas describes him as a VP and Fellow at Google, CTO of Technology & Society, and founder of Paradigms of Intelligence (Pi).^[17] The paper's author-contribution statement says Agüera y Arcas "conceived of this overall project."^[3]

The paper lists nine authors, all at Google: Blaise Agüera y Arcas, Travis Beals, Maria Biggs, Jessica V. Bloom, Thomas Fischbacher, Konstantin Gromov, Urs Köster, Rishiraj Pravahan and James Manyika, with the first eight marked as equal contributors.^[3] According to the contribution statement, Fischbacher developed the orbital dynamics and formation-flight work, Köster and Pravahan designed and ran the radiation testing, Bloom did the launch cost analysis, Gromov and Köster did the inter-satellite link analysis, Beals and Biggs developed the system design overview, and Manyika provided overall guidance and supervision.^[3] The acknowledgements thank Muon Space "for general discussions and for technical and economic feasibility analysis of the concept."^[3]

The paper

"Towards a future space-based, highly scalable AI infrastructure system design" was first published as a PDF on Google's own site (the version Ars Technica linked in its November 5, 2025 coverage), then posted to arXiv as 2511.19468 on November 22, 2025 (cs.DC).^[3][4][11] A second arXiv version followed on June 17, 2026; its comments note cleaned-up references, "improved rough estimates," fixed typos, and that the team "re-ran radiation test with improved methods."^[3] The paper is a Google preprint rather than a peer-reviewed journal article. Figures in this article are from version 2 unless stated otherwise.

System design

Working backward from an eventual future in which the majority of AI computation happens in space, the paper sets an intermediate milestone of showing that a space-based system could perform roughly comparably to a terrestrial data center.^[3] It proposes many smaller networked satellites rather than a single "monolithic" orbital data center; the authors argue that monolithic structures would need in-space assembly, would make collision avoidance harder, and would add structural mass.^[3]

The paper notes that, for a given minimum spacing between satellites, this approach makes the number of satellites grow roughly with the square of the cluster radius.^[3] It also argues that a modular design leaves room to scale "to the terawatts of compute capacity that could fit within the dawn-dusk sun-synchronous low-earth orbital band."^[3]

Inter-satellite links

The paper states that commercially available optical inter-satellite links run at 1 to 100 Gbps over distances of thousands of kilometers, far below what tightly coupled ML clusters need; for comparison, it cites Starlink's links at about 100 Gbps over up to about 5,400 km.^[3] Google estimates that each link needs aggregate bandwidth "on the order of 10 Tbps."^[3] Its answer is to fly satellites very close together. Because received power in the far field falls with the square of distance, short links can deliver the hundreds of microwatts of received optical power that data-center-style transceivers need, compared with about 1 microwatt for typical long-range links.^[3]

The link analysis assumes 400G coherent transceivers using PM-16QAM modulation on a 100 GHz ITU grid, giving 24 channels (half the C-band) and 9.6 Tbps of bidirectional bandwidth through a single aperture; a 75 GHz grid could reach 12.8 Tbps per aperture.^[3] At very short range, spatial multiplexing adds more: a 10 cm total aperture could hold a 2x2 array of 5 cm optics at 1.25 km or a 4x4 array of 2.5 cm optics at 0.32 km, each carrying its own DWDM stream.^[3]

Google reports that a bench-scale demonstrator built from off-the-shelf components achieved 800 Gbps each way (1.6 Tbps total) with a single transceiver pair across a short free-space path.^[2][3]

Orbital dynamics

To model the formation, the team started from the Hill-Clohessy-Wiltshire equations and refined the results numerically; the Research blog describes "a JAX-based differentiable model for the numerical refinement," and the paper's methods section uses SciPy's eighth-order Runge-Kutta DOP853 integrator.^[2][3] At 650 km, the dominant non-Keplerian effects are Earth's oblateness (the J2 term) and potentially atmospheric drag.^[2] In the illustrative free-fall solution the cluster goes through two shape cycles per orbit while staying bounded.^[3] The paper estimates that adjusting the formation's axis ratio slightly can hold J2-driven drift under 3 m/s per year per kilometer of distance from the reference orbit, and concludes that formation flight should need only modest delta-v beyond normal station-keeping.^[3]

Radiation testing

Google tested a Trillium v6e Cloud TPU, with its AMD host server, in a 67 MeV proton beam from the 76-inch cyclotron at the UC Davis Crocker Nuclear Laboratory.^[3] The paper describes this as "the first published radiation-testing results for such a device."^[3] Chips were irradiated from the underside through the chassis, a heatsink and the circuit board, so the energy reaching the die was lower and spread out, and the team stressed the chip with HBM memory tests, compute-centric tests, and an end-to-end transformer workload.^[3]

For a shielded satellite in the target orbit, the paper estimates a dose of about 150 rad(Si) per year, and a five-year mission dose of about 750 rad(Si).^[3]

The first version of the paper reported the SEE picture differently: it named HBM as the most SEE-sensitive component and gave an inference failure estimate "on the order of 1 per 10 million inferences," while stating that core logic and SRAM failed at about 150 rad per event.^[4] Version 2, with the re-run radiation test, shifts the main SEE concern to silent data corruption in logic and SRAM.^[3] Both versions report the same TID thresholds.^[3][4] The paper says SEE rates are likely acceptable for inference, but that their effect on training jobs, and the value of system-level mitigations, need further study.^[3]

Launch cost and economics

The paper is explicit that its launch analysis "does not constitute a full economic analysis."^[3] Using SpaceX pricing and mass data from Falcon 1 through Falcon Heavy, it finds a learning rate of about 20%: the price per kilogram falls about 20% for each doubling of cumulative mass launched.^[3] If that rate holds, which the paper says would require about 180 Starship launches per year, prices could fall below $200/kg by around 2035; if the launch rate were about 70% lower, prices could still reach about $300/kg in the same period.^[3] A second method, based on public Starship specifications and reuse targets, suggests SpaceX's own costs could fall to $60/kg or less with 10x component reuse, and $15/kg or less with 100x reuse.^[3]

To compare space with ground, the paper converts launch prices into a "launched power price," meaning the cost of lifting a kilowatt of solar power generation to LEO, amortized over satellite lifetime.^[3]

Source: Table 1 of the paper; $3,600/kg is the paper's current Falcon 9 reusable price.^[3] The paper compares these figures with reported US data-center power spending of about $570 to $3,000 per kW per year, and concludes that at launch prices of about $200/kg the annualized cost per unit of power in space "could be approximately comparable to terrestrial spend."^[3] The comparison covers power only; building and chip costs are excluded because both kinds of facility incur them.^[3]

Prototype missions

Planet partnership

Planet's November 4, 2025 announcement said it would "deploy two prototype satellites for Google, targeting a launch by early 2027," to test the performance of TPUs in space and "the ability of two such spacecraft to work in concert, flying in tandem with high bandwidth cross link communications."^[5] Planet said the program shares a satellite bus with its Owl mission.^[5] Beals told Ars Technica that Google provides the TPUs and compute payload and was paying Planet to handle the rest of the preparation; Ars also noted that Alphabet holds an equity stake in Planet.^[11]

First test satellite on Transporter-18

On September 24, 2026, Google announced that Suncatcher was "scheduled to embark on its first test in orbit," with a prototype satellite flying on "the upcoming Transporter-18 rideshare mission with SpaceX," developed in partnership with Planet.^[7] The blog said the mission is designed "to gather in-orbit data on how our TPUs handle the physical stress of spaceflight and the radiation and thermal extremes of space," and that "putting our first TPUs in orbit next week will help us get data and learnings to inform future launches."^[7] Pichai wrote on X: "Can our TPUs survive and operate in space? Well, we're going to find out. Project Suncatcher is hitching a ride aboard @SpaceX's Transporter-18 mission, testing a prototype satellite built in partnership with @planet," and added, "One small step for TPUs...."^[9] Google's own post on X described the flight as a test mission "designed to gather data exploring how we can one day host machine learning infrastructure in space."^[8]

Additional details come from press reports rather than Google's blog. Ars Technica reported that the satellite is called MVP, is about the size of a refrigerator, and contains four TPUs; citing The New York Times, it said the solar panels supply only about one kilowatt.^[10][13] According to Ars, the cooling system allows the TPUs to run in bursts of about 15 minutes before they must shut down so the radiators can catch up, Google plans to run Gemini models on them, and the satellite will operate for only a few months.^[10] Ars also reported that the original plan was two custom satellites in 2027, but that Google wanted to move faster and chose to integrate its chips into a satellite Planet had already built for an early test; the 2027 launches remain planned.^[10] Reuters reported that Google said the first mission is meant to gather in-orbit data and find failure points "rather than demonstrate an operational orbital data center."^[12]

As of September 25, 2026, the satellite had not launched. SpaceX says it is targeting Thursday, October 1, 2026 for the Falcon 9 launch of Transporter-18 from Space Launch Complex 4E (SLC-4E) at Vandenberg Space Force Base, California, with a 58-minute window opening at 11:18 a.m. PT (18:18 UTC) and a backup opportunity on October 2; its deployment timeline lists "Project Suncatcher M1," manifested by Planet Labs, among the mission's 130 payloads.^[20] The tracker Space Launch Schedule also lists the mission for no earlier than October 1.^[15]

Timeline

Engineering challenges

Google's September 2026 update, which accompanied a short video series on YouTube, set out the problems the first flight is meant to probe.^[7][18]

• Launch loads. Google says the ride to low Earth orbit takes about 10 minutes, with vibration and sustained accelerations up to 10 g, and that individual components such as the TPUs can see 50 to 100 g. The team shook the satellite on all three axes to mimic launch frequencies and wrote that it was "pleasantly surprised that the hardware held up to the force."^[7]
• Radiation. Google says its proton-beam results show Trillium TPUs "can survive a radiation total ionizing dose greater than what they would receive during a five-year space mission," while adding that "some things can only be tested in space."^[7]
• Cooling. With no airflow in a vacuum, heat can only leave through radiators. Google is working on "a combination of heat pipes and radiators" and has tested it in a thermal vacuum chamber.^[7] Ars Technica described the flight design as a malleable thermal interface material connecting the chips to aluminum and copper heat pipes that feed a radiator.^[10]
• Satellite interconnect. Future satellites would each carry "dozens of TPU chips" in orbiting clusters and talk over lasers built for very high bandwidth at very short range. Google compares the pointing precision to "hitting a coin-size target from miles away while both points are in motion," and says it will test this in 2027 with two satellites.^[7]

The paper adds longer-term issues: failed TPUs cannot be swapped by technicians in orbit, so the simplest answer is redundant provisioning; optical ground links must overcome atmospheric turbulence and beam-tracking problems; and eventually gigawatt-scale systems may need satellites in which power collection, compute and heat rejection are tightly integrated.^[2][3]

Reception and criticism

Press coverage of the 2026 announcement stressed how early the work is. Ars Technica wrote after the September 2026 announcement that "the team expects it will be years before Suncatcher evolves from 'project' to 'product.'"^[10] Reuters, reporting the same announcement, wrote that experts say the concept "remains years from being commercially viable given high launch costs, engineering constraints and satellite production bottlenecks."^[12]

In a December 2025 Scientific American article on orbital data centers generally, University of Pennsylvania computer architect Benjamin Lee said orbital platforms need large radiators that add launch mass, and that "we would still require a very large number of launches to build orbital data centers that are competitive with those on Earth."^[16] The same article cited Saarland University researchers whose paper "Dirty Bits in Low-Earth Orbit" estimated that a solar-powered orbital data center could have an order of magnitude greater emissions than one on Earth once launch and reentry are counted, and quoted University of Regina astronomer Samantha Lawler on reentry pollution and space debris.^[16]

Ars Technica's 2025 coverage noted Google's mixed record with moonshots, contrasting Waymo with the shuttered Project Loon balloon-internet effort.^[11] SpaceX founder Elon Musk, whose company is pursuing its own orbital compute plans, responded to a post about Google's launch on September 24, 2026 by writing: "The amount of compute in space will obviously round up to 100% of all compute."^[19]

Related projects

Suncatcher's clustered small-satellite approach differs from single-large-structure concepts. Ars Technica contrasted it with Starcloud's plan, with NVIDIA as a partner, for a very large orbital data center with kilometer-scale solar and cooling panels.^[11] Starcloud launched a 60 kg satellite carrying an NVIDIA H100 GPU in November 2025.^[14][16] SpaceX's orbital compute effort is covered in SpaceX Starmind, and SpaceX also flies the Transporter rideshare carrying Google's first test satellite.^[7] Scientific American also pointed to a Chinese "space data center" constellation that had begun launching and the European Union's ASCEND study.^[16]

References

1. ^1 ^2 ^3Google. "Meet Project Suncatcher, a research moonshot to scale machine learning compute in space." The Keyword, November 4, 2025. blog.google/...google-project-suncatcher
2. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10 ^11 ^12Travis Beals. "Exploring a space-based, scalable AI infrastructure system design." Google Research Blog, November 4, 2025. research.google/...ai-infrastructure-system-design
3. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10 ^11 ^12 ^13 ^14 ^15 ^16 ^17 ^18 ^19 ^20 ^21 ^22 ^23 ^24 ^25 ^26 ^27 ^28 ^29 ^30 ^31 ^32 ^33 ^34 ^35 ^36 ^37 ^38 ^39 ^40 ^41 ^42 ^43 ^44 ^45 ^46 ^47 ^48 ^49 ^50 ^51 ^52 ^53 ^54Blaise Agüera y Arcas, Travis Beals, Maria Biggs, Jessica V. Bloom, Thomas Fischbacher, Konstantin Gromov, Urs Köster, Rishiraj Pravahan, James Manyika. "Towards a future space-based, highly scalable AI infrastructure system design." arXiv:2511.19468 (v1 November 22, 2025; v2 June 17, 2026). arxiv.org/...2511.19468
4. ^1 ^2 ^3 ^4Google. "Towards a future space-based, highly scalable AI infrastructure system design" (original preprint PDF, November 2025). services.google.com/...suncatcher_paper.pdf
5. ^1 ^2 ^3 ^4 ^5Planet Labs PBC. "Planet to Build and Operate Advanced Space Platform for Google's Project Suncatcher Moonshot." Planet Pulse, November 4, 2025. planet.com/...form-for-project-suncatcher-moonshot
6. ^1 ^2Sundar Pichai. Post on X, November 4, 2025. x.com/...1985754323813605423
7. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10 ^11 ^12Travis Beals. "Behind Project Suncatcher, our moonshot to put AI in space." The Keyword, September 24, 2026. blog.google/...google-project-suncatcher-facts
8. ^1 ^2Google. Post on X, September 24, 2026. x.com/...2103229008343126519
9. ^Sundar Pichai. Post on X, September 24, 2026. x.com/...2103209164072010051
10. ^1 ^2 ^3 ^4 ^5Ryan Whitwam. "Google's first Suncatcher orbital data center test launches October 1." Ars Technica, September 24, 2026. arstechnica.com/...-center-test-launches-october-1
11. ^1 ^2 ^3 ^4 ^5 ^6Stephen Clark. "If you want to satiate AI's hunger for power, Google suggests going to space." Ars Technica, November 5, 2025. arstechnica.com/...-google-suggests-going-to-space
12. ^1 ^2Akash Sriram. "Google plans first test of AI chips in space under Project Suncatcher." Reuters, September 24, 2026 (via StreetInsider). streetinsider.com/...27100094
13. ^Mike Wheatley. "Google's first Project Suncatcher AI satellite set to blast off into orbit next week." SiliconANGLE, September 24, 2026. siliconangle.com/...blast-off-into-orbit-next-week
14. ^Bruce Gil. "Google's Project Suncatcher Is Sending AI Chips Into Space Next Week." Gizmodo, September 24, 2026. gizmodo.com/...ips-into-space-next-week-2000816985
15. ^1 ^2 ^3Space Launch Schedule. "SpaceX Transporter 18 (Dedicated SSO Rideshare) Falcon 9 Block 5 Rocket Launch." Accessed September 25, 2026. spacelaunchschedule.com/...dedicated-sso-rideshare
16. ^1 ^2 ^3 ^4Jeremy Hsu. "Space-Based Data Centers Could Power AI with Solar Energy, At a Cost." Scientific American, December 9, 2025. scientificamerican.com/...data-centers-in-space
17. ^Google Research. "Blaise Aguera y Arcas" (researcher profile). research.google/...106776
18. ^Google. "Google's latest moonshot to put machine learning in space." YouTube. youtube.com/watch
19. ^Elon Musk. Post on X, September 24, 2026. x.com/...2103188449960313173
20. ^1 ^2 ^3SpaceX. "Transporter-18 Mission." Accessed September 25, 2026. spacex.com/...transporter18