Project Suncatcher is about to leave the lab. Google said on 24 September 2026 that it will launch a prototype satellite carrying its own AI chips, its Tensor Processing Units (TPUs), on SpaceX’s Transporter-18 rideshare mission. The launch is scheduled for 1 October from Vandenberg Space Force Base in California. The New York Times called it the first step by a major technology company towards AI data centres in space.
The satellite is small and the goals are modest. It carries four TPUs, roughly the computing power of one server in a data centre, and its solar panels supply about a kilowatt. Google wants to learn whether its chips survive the shaking of launch, the radiation of low Earth orbit and the problem of cooling electronics where there is no air. The company is clear that it is not launching a data centre, only the first experiment towards one.
This article explains what Google announced, why it thinks orbit could one day host AI computing and how the Project Suncatcher team has tested the hardware so far. It covers the economics, the competing projects from Starcloud, SpaceX and Nvidia, the sceptics’ objections and what Project Suncatcher means for organisations that buy computing today.
Table of contents
- What Google Announced for Project Suncatcher
- Why Google Wants AI Computing in Orbit
- Project Suncatcher’s First Hurdle: Surviving the Launch
- Radiation: How Project Suncatcher Tested TPUs in a Proton Beam
- Cooling Chips Where There Is No Air
- From One Satellite to a Project Suncatcher Constellation
- How Project Suncatcher Moved Faster Than Planned
- The Economics Behind Project Suncatcher
- How Project Suncatcher Compares With the Orbital Computing Race
- The Sceptics’ Case Against Orbital Data Centres
- What Project Suncatcher Means for Businesses
- What Happens After the Project Suncatcher Launch
- Project Suncatcher FAQ
- References
What Google Announced for Project Suncatcher
The news came in a post on Google’s blog, The Keyword, by Travis Beals, the senior director of product management for Project Suncatcher. The New York Times was given the first inside look at Project Suncatcher and published its report the same morning. Both describe a careful first step rather than a leap.
A prototype called MVP
Google has named the satellite MVP. According to The New York Times, it is about the size of a refrigerator. Inside are four TPUs, the custom chips Google designs to train and run its AI models, including Gemini. On Earth, a data centre runs thousands of these chips. MVP will answer simple AI queries and is planned to operate for about a year. It could stay in orbit for up to six years before it falls back and burns up in the atmosphere.
The launch and the ride
MVP will fly on Transporter-18, one of SpaceX’s regular rideshare missions, on a Falcon 9 rocket from Vandenberg. Rideshare flights carry dozens of satellites from different customers at once, which keeps the cost down. Barron’s reported that spaceflight databases list the mission for 1 October, and The Verge and The New York Times give the same date. Launch dates move, so the day could slip.
Who built the satellite
Google did not build the spacecraft from scratch. Planet, the Earth-imaging company better known as Planet Labs, supplied it. Google has invested in Planet, and the two announced a joint learning mission in November 2025. The chips, the cooling system and the software are Google’s. The satellite bus, the part that provides power, pointing and communications, comes from Planet.
| Project Suncatcher MVP | Detail | Source |
|---|---|---|
| Size | About the size of a refrigerator | The New York Times |
| Payload | Four TPUs, roughly one data centre server of computing | The New York Times |
| Power | About 1 kilowatt from its solar panels | The New York Times |
| Spacecraft | A ready-made Planet satellite fitted with Google’s chips | Planet via The New York Times |
| Launch | SpaceX Transporter-18 on Falcon 9, Vandenberg, scheduled for 1 October 2026 | Google, The Verge, Barron’s |
| Planned operation | About a year, answering short Gemini queries | The New York Times |
| Time in orbit | Up to six years before re-entry | The New York Times |
Why Google Wants AI Computing in Orbit
The reasoning behind Project Suncatcher starts with energy. AI models need enormous amounts of electricity, and data centres on the ground are running into limits on power, land, water and public tolerance.
The squeeze on the ground
We have covered that squeeze from several angles. Google, Nvidia and Anthropic are backing Emerald AI to find spare grid capacity for new data centres. California has tightened its rules on data centre energy and water use. China is building AI data centres across the plains of Inner Mongolia, where land and wind power are plentiful.
Barron’s cited Morgan Stanley analyst Stephen Byrd, who sees AI power demand growing to almost 100 gigawatts by 2028. Total US generating capacity is roughly 1,300 gigawatts.
On those figures, AI alone would need about 7.7% of today’s US capacity, which Barron’s rounded to 8%.
Eight times the solar output
Space offers a different energy source. In its 2025 research post, Google said that in the right orbit a solar panel can be up to eight times more productive than on Earth and can produce power almost continuously. That reduces the need for heavy batteries. The company also noted that the Sun emits more than 100 trillion times humanity’s total electricity production.
The dawn to dusk orbit
The design Google has published places its satellites in a dawn to dusk sun-synchronous orbit. A satellite in this kind of orbit rides the line between day and night, so it stays in sunlight almost all the time. That is what makes the power close to constant. The name Project Suncatcher reflects this idea: catching the Sun’s energy where it is strongest and most reliable.
Project Suncatcher's First Hurdle: Surviving the Launch
Before any chip can do useful work in orbit, it has to get there in one piece. Beals described the launch as the first of three physical tests the Project Suncatcher mission must pass.
Ten minutes of shaking
A rocket trip to low Earth orbit takes about ten minutes. During that time the spacecraft feels intense vibration and sustained acceleration of up to 10 times the force of gravity. Individual parts can feel much more. Google said the TPU chips themselves can experience forces of 50 to 100 g.
How the team tested it
To copy those conditions, the team shook the satellite hard on all three axes, matching the frequencies of a rocket launch. The New York Times watched the test in a San Francisco lab in August. Technicians painted a small line over each screw so that any loosening would show. The screws stayed put and the chips stayed intact. “Tests like this rarely go as planned, so we were pleasantly surprised that the hardware held up to the force,” Beals wrote.
Why shaking matters for chips
Data centre hardware is designed for a quiet, climate-controlled room. It is not built to survive a rocket. Solder joints, heat spreaders and memory packages can crack or shift under heavy vibration. A satellite also cannot be opened and repaired after launch, so any damage is permanent. Passing a ground test does not guarantee a clean flight, which is one reason Project Suncatcher needs a real launch.
Radiation: How Project Suncatcher Tested TPUs in a Proton Beam
The second test is radiation. Outside the protection of Earth’s atmosphere and magnetic field, solar events and cosmic rays hit electronics constantly.
Crocker Nuclear Laboratory
Google began taking AI chips to Crocker Nuclear Laboratory in Davis, California, in February 2025, according to The New York Times. There, a cyclotron fired a proton beam at the chips while they ran AI workloads. The Project Suncatcher team watched how errors affected the work. Beals wrote that Google’s Trillium TPUs “hold up remarkably well” and can survive a total radiation dose greater than they would receive in five years in space.
Bit flips and restarts
Radiation can cause bit flips, where a stray particle changes a stored zero into a one or the reverse. In a calculation, a single flipped bit can produce a wrong answer or crash a program. The New York Times reported that restarting the chips could usually reset the bit flips. That matters because it suggests many errors can be handled with software rather than with heavy shielding or special radiation-hardened chips.
The numbers from the paper
Google’s preprint paper, “Towards a future space-based, highly scalable AI infrastructure system design”, gives the detail. The team tested Trillium, Google’s v6e Cloud TPU, in a 67 MeV proton beam. The most sensitive part was the High Bandwidth Memory. It began showing irregularities only after a total dose of 2 krad(Si). The expected dose over a shielded five-year mission is 750 rad(Si). No hard failures were seen up to the highest dose tested, 15 krad(Si) on a single chip.
On those figures the memory held out to about 2.7 times the expected mission dose, and the chip survived 20 times that dose without a hard failure. Google has not said which TPU generation flies on MVP. Its published radiation data so far covers Trillium.
Cooling Chips Where There Is No Air
The third test may be the hardest. Beals called cooling orbital data centres “a crucial research challenge”. TPUs produce a great deal of heat in a small area, and that heat has to go somewhere.
Why a vacuum changes everything
On Earth, fans blow air over heat sinks, or liquid carries heat away to cooling towers. In space there is no air to blow. In a vacuum, heat can only leave by radiating away as infrared light, from surfaces called radiators. That requires a completely different approach to cooling electronics, and radiators get large quickly as power rises.
Layers from chip to radiator
The New York Times described the system on MVP. The chips sit on a green motherboard. Above them is a thermal interface material, a pale green putty that comes in sheets. It connects the chips to layers of aluminium and copper that spread the heat away from the board. At the top is a radiator panel that sends the heat out into space. Google also said it is combining heat pipes with radiators, and it has tested the design in a thermal vacuum chamber that copies the heat and vacuum of space.
The 15-minute duty cycle
The limits show how early this is. Beals told The New York Times that the chips can run for about 15 minutes before they must shut down to cool off. In that time they can process short queries so that Gemini can respond. A data centre on the ground runs around the clock, so a working orbital version would need far better cooling than MVP carries.
| Stress | Ground test | Result so far | What only orbit can show |
|---|---|---|---|
| Launch loads | Three-axis vibration test of the whole satellite | Hardware held, screws stayed put | A real Falcon 9 ride to orbit |
| Radiation | Proton beam at Crocker while running AI work | Survived more than a five-year dose, bit flips usually reset | Months of solar particles and cosmic rays |
| Heat | Thermal vacuum chamber, heat pipes and radiators | Worked in the chamber | Radiator performance through real day and night cycles |
| Service | None possible | Not yet tested | Whether short Gemini queries can be served from orbit |
From One Satellite to a Project Suncatcher Constellation
MVP is a single satellite. The long-term vision of Project Suncatcher is a fleet of them working as one computer.
Laser links over short distances
Future Google satellites would each carry dozens of TPU chips and fly in clusters. To share AI work, they must pass data between each other very fast. Google plans to use lasers. Beals wrote that laser links already exist in space, but most are built for low bandwidth over long distances. Project Suncatcher needs the reverse: very high bandwidth over very short distances. He compared the required precision to hitting a coin-sized target from miles away while both points are moving.
Tens of terabits per second
The research post said links between satellites would need to carry tens of terabits per second to match a ground data centre. Google’s bench-scale demonstrator has already reached 800 gigabits per second each way, 1.6 terabits per second in total, with a single pair of transceivers. Flying the satellites close together is what makes this possible, because received power falls with the square of the distance.
The 81-satellite cluster model
The paper modelled an illustrative cluster of 81 satellites at a mean altitude of 650 kilometres. The cluster has a radius of 1 kilometre, and neighbouring satellites sit roughly 100 to 200 metres apart. Google’s models suggest only modest station-keeping manoeuvres would be needed to hold that formation. The New York Times reported that Google has designs for fleets of more than 80 satellites flying in close formation. It is also talking to designers about a custom satellite the length of a football pitch.
The 2027 laser test
The laser links will be tested in 2027, when Google and Planet put two satellites in orbit. That was the original plan announced in November 2025, when Project Suncatcher was scheduled to launch two prototypes by early 2027. MVP came first because Google wanted to be in orbit this year.
How Project Suncatcher Moved Faster Than Planned
The history of the project explains why a quick, simple satellite is flying before the more ambitious pair.
An idea from a gathering on AI energy
The New York Times reported that the project was the idea of Blaise Agüera y Arcas, a Google vice president and AI researcher who leads a team studying intelligence. About three years ago he attended a gathering where the main topic was the rising energy needs of AI. He left convinced that data centres needed to move to space. “There are long-standing ideas in science fiction about using stars for computation,” he said.
A pitch to Pichai and Brin
Agüera y Arcas took the idea to James Manyika, Google’s senior vice president for research. Manyika was convinced it would not work but agreed to tests of cooling and radiation. In May 2025 they pitched Sundar Pichai, and Sergey Brin was also in the room. According to Agüera y Arcas, Brin said: “OK, so this is a good idea. Let’s talk about how we’re doing it.”
Swapping in a ready-made satellite
Google announced the project publicly on 4 November 2025, alongside the preprint paper. Planet had agreed to launch two satellites for Google in 2027. But Eric Stevens, a director of systems engineering at Planet, told The New York Times that Google wanted to be in orbit this year and was willing to take risks to get there. To speed up, Google installed its chips in a satellite Planet had already built.
The Economics Behind Project Suncatcher
Physics is only half the problem. The other half is cost, and there Project Suncatcher depends heavily on the price of launch.
The $200 per kilogram threshold
Google’s analysis of past and projected launch prices suggests that, with a sustained learning rate, the cost to reach low Earth orbit may fall below $200 per kilogram by the mid-2030s. At that point, Google argued, the cost of launching and running a space-based data centre could become roughly comparable to the energy costs of an equivalent ground data centre, measured per kilowatt per year. Google told The New York Times it expects that rough parity in the mid-2030s.
What Google has not said
Google declined to tell The New York Times how much it is spending on Project Suncatcher. The comparison in its research post is with the reported energy costs of ground data centres, not their full cost of construction, cooling and staffing. Launch is also only one part of an orbital bill. Satellites wear out, cannot be repaired and must be replaced, which ground hardware does not demand in the same way.
| Cost claim | Launch price assumed | When | What it is compared with |
|---|---|---|---|
| Google, Project Suncatcher paper | Below $200 per kg | Mid-2030s | Energy cost of an equivalent ground data centre per kilowatt per year |
| Starcloud, Starcloud-3 | About $500 per kg on Starship | Commercial Starship access expected 2028 or 2029 | A projected price of about $0.05 per kilowatt hour |
Manyika’s tempered expectations
Google’s research lead was frank about timing. “We don’t expect, to be perfectly frank, that we’ll have anything usefully operational in the next few years,” Manyika told The New York Times. He compared Project Suncatcher to Google’s early work on driverless cars. “Remember how Google was researching for like 15 years before anything showed up? I think this is going to look like that,” he said.
How Project Suncatcher Compares With the Orbital Computing Race
Google is not alone. Several companies want to put AI computing in orbit, and some are already there.
Starcloud
Starcloud, a start-up from Redmond, Washington, launched Starcloud-1 in November 2025. The 60 kg satellite carried an Nvidia H100 graphics processor. It went on to train NanoGPT, a small language model, on the complete works of Shakespeare, and to run Google’s open Gemma model. Starcloud raised $170 million at a $1.1 billion valuation in March 2026 and a further $250 million at a $2.3 billion valuation in August, according to DatacenterDynamics. It has said Starcloud-2, with an Nvidia Blackwell chip, is due in October 2026.
SpaceX and xAI
SpaceX, which acquired Elon Musk’s AI company xAI in February 2026, has asked US regulators for permission to run a network of up to a million computing satellites. Barron’s reported that SpaceX plans to launch its own AI computing satellites as soon as 2027 on its Starship rocket. It also noted that SpaceX is giving a rival a ride here. Alphabet is paying almost $1 billion a month to use SpaceX’s AI computing capacity, Barron’s added.
Nvidia and the rest
Nvidia has announced a space-hardened Space-1 Vera Rubin module, due with Starcloud in late 2028. DatacenterDynamics reported that the chip has yet to be built, and that Nvidia also plans to fly it with Cowboy Space (formerly Aetherflux), Axiom Space, Kepler Communications, Planet and Sophia Space. Jeff Bezos, Sam Altman and former Google chief executive Eric Schmidt have all spoken in favour of orbital data centres.
| Programme | Hardware in orbit | Status in September 2026 | Next milestone |
|---|---|---|---|
| Google Project Suncatcher | None yet, four TPUs on MVP | Launch scheduled for 1 October | Two laser-linked satellites in 2027 |
| Starcloud | One Nvidia H100 on Starcloud-1 since November 2025 | $450 million raised, $2.3 billion valuation | Starcloud-2 with a Blackwell chip, October 2026 |
| SpaceX with xAI | No computing satellites yet | Seeking approval for up to a million satellites | AI computing satellites on Starship as soon as 2027 |
| Nvidia Space-1 Vera Rubin | Not built yet | Partners named, including Planet and Axiom Space | Deployment with Starcloud, late 2028 |
Where Google’s approach differs
Project Suncatcher stands out in two ways. Google flies its own chips rather than buying graphics processors from Nvidia, so it controls the whole stack from silicon to software. And it is unusually open about the research, with a public preprint and a video series. Its first satellite is also deliberately limited, which reflects its stated aim of learning before scaling.
The Sceptics' Case Against Orbital Data Centres
Not everyone is convinced that orbit is a sensible place for AI computing, and Google’s own paper lists hard problems that remain.
Heat at scale
Cooling is the objection heard most often. MVP must stop every 15 minutes or so to cool down, and it draws about a kilowatt. A useful orbital data centre would draw many megawatts. For comparison, a TNW report in March noted that SpaceX’s entire Starlink network of about 10,000 satellites generates around 200 megawatts, while US data centres with more than 25 gigawatts of capacity are under construction. Radiators big enough to shed megawatts of heat would be very large.
Scale, repair and debris
Brandon Lucia, a professor of electrical and computer engineering at Carnegie Mellon University, told The New York Times that growing from one satellite to a data-centre-sized network would take years and enormous funds. “If you do this on a large scale, there are additional engineering problems,” he said. “That is uncharted waters.” Failed satellites cannot be serviced, and thousands of closely packed spacecraft raise questions about collisions and debris.
Ground links and the obvious question
Google’s research post names thermal management, high-bandwidth links to the ground and on-orbit reliability as major challenges still to solve. Researcher Joe Bak-Coleman offered journalists a sharper version on Bluesky, suggesting they ask about heat dissipation at scale and, simply: “Doesn’t the sun come down here too?” Solar farms and batteries on Earth are getting cheaper as well, so orbit has to beat a moving target.
| Factor | Data centre on the ground | Orbital design in Project Suncatcher |
|---|---|---|
| Power | Grid connections that are slow and contested | Near-constant sunlight, panels up to 8 times more productive |
| Cooling | Air or liquid, often using water | Radiators only, with no airflow |
| Land and permits | Planning fights and local opposition | No land, but orbital and spectrum licences |
| Repair | Technicians on site | No servicing, hardware must be replaced |
| Networking | Fibre between racks and sites | Laser links between satellites, radio or laser to the ground |
| Main cost driver | Energy and construction | Launch price per kilogram |
What Project Suncatcher Means for Businesses
For most organisations, nothing changes on 1 October. The more useful question is what the Project Suncatcher experiment says about where computing is heading.
No change to your cloud plans yet
MVP will serve short test queries for a year. No major cloud provider sells orbital computing to ordinary customers, and Google’s own research lead does not expect anything usefully operational for years. Decisions about cloud infrastructure and data centre operations in 2026 and 2027 should be made on the ground, on today’s prices and today’s constraints.
Energy is the real story
Project Suncatcher is a signal of how seriously the largest AI buyers take the power problem. When a company with Google’s resources explores launching servers into space, it tells you ground capacity is tight. Organisations planning heavy AI use should expect power, location and sustainability to shape price and availability. Al Gore has argued that data centre emissions are small next to landfills and air conditioning, but he worries the rush to power them could lock in gas generation for decades.
Where orbital computing could matter first
If orbital computing arrives, it will probably start with work that already happens in space. Earth-observation satellites collect far more imagery than they can send down. Processing it in orbit, with computer vision models that pick out ships, floods or crop damage before downlink, would save bandwidth and time. Starcloud already processes data from Capella Space’s radar satellites in orbit, TNW reported. We covered a similar push for autonomy in AstroForge’s AI-commanded spacecraft.
Questions to ask if a provider pitches orbit
Orbital computing raises questions that do not apply on the ground. Where is data held, legally, when it sits on a satellite? What latency should you expect for a round trip to orbit? What happens to your workload when a satellite fails and cannot be repaired? How are launch emissions and space debris counted in sustainability claims? If you need help weighing new infrastructure claims against your own needs, independent technology consulting can separate the hype from the options.
What Happens After the Project Suncatcher Launch
The next few months will show whether the ground tests hold up in orbit.
Milestones to watch
The first milestone is a clean launch and deployment on 1 October, or whenever Transporter-18 flies. After that, watch for Google confirming that the TPUs have powered on, that the cooling system is working and that the satellite has answered its first Gemini query from orbit. Google has promised to share what it learns, including through its video series.
| Date | Project Suncatcher milestone |
|---|---|
| February 2025 | Radiation testing of AI chips begins at Crocker Nuclear Laboratory |
| May 2025 | Pichai and Brin approve Project Suncatcher |
| 4 November 2025 | Public announcement and preprint paper, Planet partnership for two satellites by early 2027 |
| August 2026 | MVP passes its vibration test in San Francisco |
| 24 September 2026 | Google announces the first in-orbit test |
| 1 October 2026 | Scheduled launch on SpaceX Transporter-18 |
| 2027 | Two satellites test laser links in orbit |
| Mid-2030s | Google’s estimate for rough cost parity with ground data centres |
What would count as success
Google has set a low bar for Project Suncatcher on purpose. “This first launch is about seeing what works, identifying points of failure, and applying those findings to future missions,” Beals wrote. He told The New York Times that a perfect record would be a warning sign. “If five years from now, everything we’ve done has worked perfectly, it probably means we’ve not taken enough risk and we’ve not learned as much as we could,” he said.
The long view
Beals also described what success would look like in the end. “If we’re really successful with this in the long run, this will ultimately be boring, and people won’t think anything at the fact that their Gemini query might be getting served in space,” he said. That is a long way off. For now, Project Suncatcher is one refrigerator-sized satellite, four chips and a year of data.
Project Suncatcher FAQ
What is Project Suncatcher?
Project Suncatcher is a Google research moonshot, announced in November 2025, that explores whether fleets of solar-powered satellites carrying Google’s TPU chips could one day run AI computing in space.
When does the first Project Suncatcher satellite launch?
The prototype, called MVP, is scheduled to launch on 1 October 2026 on SpaceX’s Transporter-18 rideshare mission, on a Falcon 9 rocket from Vandenberg Space Force Base in California.
What is on the satellite?
MVP carries four Google TPUs, roughly the computing power of one data centre server, powered by about a kilowatt of solar panels. It was built by Planet and fitted with Google’s chips and cooling system.
What will the mission test?
It will test whether the TPUs survive launch forces, radiation in low Earth orbit and cooling in a vacuum. The chips will also answer short Gemini queries in bursts of about 15 minutes.
Is Google building a data centre in space?
Not yet. Google says MVP is an experiment, not a data centre. Two satellites follow in 2027, and the company estimates orbital computing could approach ground costs in the mid-2030s if launch prices fall below $200 per kilogram.
References
Behind Project Suncatcher, our moonshot to put AI in space
Meet Project Suncatcher, a research moonshot to scale machine learning compute in space
Exploring a space-based, scalable AI infrastructure system design
Towards a future space-based, highly scalable AI infrastructure system design
Google takes the AI data center race to outer space (The New York Times)
Google is sending an AI satellite into space next week
Why Google and SpaceX are testing AI in orbit: The earthly grid can’t keep up (Barron’s)
Google will launch its first AI chips into orbit next week
Starcloud raises $170M at a $1.1B valuation to build data centres in orbit
Starcloud closes $250m Series A extension at $2.3bn valuation
More AI coverage: explore Progressive Robot's AI Models, Tools & Releases hub — hands-on reviews, setup guides and benchmarks in one place.
