Project Suncatcher explained: why Google put four AI chips in orbit

A satellite the size of a refrigerator, reported to carry four of the chips Google uses to run AI, reached orbit on Thursday, and Google says it is in contact and "operating as expected". It is the first hardware test of Project Suncatcher, the company's research effort to find out whether data centres could one day run in space on near-constant sunshine. The mission is not measuring how fast the chips are. It is measuring whether a computer can stay cool, and stay correct, in a place with no air and plenty of radiation.

The plain-English version

Start with a thermos flask. It keeps coffee hot because the gap between its walls is a vacuum, and a vacuum is a superb insulator: with no air to carry warmth away, heat can only leak out slowly as invisible infrared light. Space is the biggest thermos there is.

Now imagine switching on a hot plate inside that thermos. That is roughly what Google launched on 1 October: a satellite about the size of a refrigerator, built with the satellite company Planet, carrying four of the chips Google uses to run AI, according to NPR. On Earth, chips like these sit in huge buildings called data centres and are cooled by fans and water. In orbit there is neither. The only way to shed heat is to pipe it to flat panels called radiators and let it glow away into space. NPR reports that the chips will run a small Google AI model for just 15 minutes at a time because of heat.

Why go to the trouble? Because space has something else: sunshine that almost never stops. AI needs enormous amounts of electricity, and on the ground that means power plants, land, water and long waits. A solar panel in the right orbit is never under cloud and almost never in the dark, and Google says it can generate up to eight times as much power as on Earth. The idea, called Project Suncatcher, is to fly clusters of satellites that each carry dozens of chips and talk to one another with lasers.

This week's launch does not show that any of that works. Think of it as a first exam with three questions. Do the chips survive the violent shaking of a rocket launch? Do they keep computing correctly while radiation flips bits in their memory? Does the cooling system behave? Google says the satellite is "operating as expected" and that it will gather data over the coming weeks. The goal, NPR reports, is a year of operation, and Google plans to fly two more satellites next year to test the laser links.

Even the project's leader, Travis Beals, told NPR: "I don't see this being something where it's cheaper to do this in the next five years." Whether it ever is depends less on the chips than on two unglamorous things: how heavy the radiators have to be, and how cheap rockets become.

How it actually works

The problem. AI runs on electricity, and electricity is getting hard to find. Data centres are projected to use around 3% of the world's electricity by 2030, roughly double their share today, according to an International Energy Agency report cited by Scientific American. Google's chief executive has said he expects capital spending of $180 billion to $190 billion this year, NPR reports. On the ground, every new data centre needs a grid connection, land and usually water for cooling.

The old way is to build wherever power can be found, and wait for more.

The new idea is to send the computing to the power. Google's design paper has three ingredients:

1. Chase the sun. A "dawn-dusk sun-synchronous" orbit rides the line between day and night, so a satellite's panels are lit almost all the time. Sunlight is stronger above the atmosphere (the paper uses 1.361 kilowatts per square metre), and there is no night, weather or low sun angle to drag down the average. Together these give the paper's figure of up to 8 times the yearly energy of the same panel at mid-latitude on Earth.
2. Fly close, talk by laser. AI chips work in tightly connected groups, so the satellites need links as fast as those inside a data centre. A laser beam spreads out as it travels, so the signal weakens quickly with distance. Google's answer is to fly the satellites only 100 to 200 metres apart, close enough for ordinary data-centre laser equipment to work.
3. Fly ordinary chips. Rather than special space-grade processors, use the same TPUs (tensor processing units, Google's AI chips) as on the ground, measure what radiation does to them, and handle the errors with software and spare capacity.

The paper notes that the Sun emits more than 100 trillion times humanity's total electricity production. Older schemes for space solar power struggled with getting that energy back to Earth, as the paper acknowledges. Here only data comes down.

Where it gets hard.

What the prototype does.

1. Survive the ride. During the roughly ten-minute ascent the spacecraft takes sustained loads of up to 10 times gravity, and individual components feel 50 to 100 g, Google wrote before launch.
2. Phone home. Google's launch post says the team has confirmed contact.
3. Compute in bursts. The chips run a version of Google's Gemma model for 15 minutes at a time, answering simple queries, NPR reports. Heat is the stated reason.
4. Count the errors. Before launch, Google fired a proton beam at a TPU on the ground to predict how often radiation would corrupt a calculation or crash the system. Orbit will show whether those predictions hold.
5. Hand over. In 2027 Google plans to fly two satellites to test the laser links.

One caveat on the hardware: according to a correction NPR appended to its report, this prototype is in "a standard orbit and uses batteries". The always-sunlit orbit is the plan for later satellites, so the headline energy advantage is not something this mission tests.

The deep dive

What flew, and what did not

Google's two posts give few specifications; the details come from reporting. NPR and Scientific American both report four TPUs, and NPR adds a one-year goal. Futurum's pre-launch analysis, which draws on New York Times reporting, gives about 1 kW of solar power and compute equivalent to a Google Cloud "TPU v6e-4" slice. Liftoff came at 11:32 a.m. Pacific time on 1 October from Vandenberg Space Force Base, on a Falcon 9 carrying 130 payloads, Via Satellite reports; Spaceflight Now's launch preview had listed that time and the pad, SLC-4E. Planet, which counts 20 satellites on the flight including the Suncatcher prototype, calls it "the first-ever test of Google's Tensor Processing Units (TPUs) in space". Not part of this flight: laser links between satellites, formation flying, or the target orbit.

Power in equals heat out

Every watt a chip draws ends up as heat, and in a vacuum the only exit is radiation. A panel of area $A$, emissivity $\varepsilon$ and temperature $T$ radiates

where $\sigma = 5.67\times10^{-8}\ \mathrm{W\,m^{-2}\,K^{-4}}$ is the Stefan-Boltzmann constant. At 300 K with emissivity 0.9, that is about 410 W per square metre per face, so each kilowatt of chips needs roughly 2.4 m² of radiator before counting the sunlight and Earth-glow the panel absorbs (our arithmetic). Independent estimates land nearby: Slava Turyshev's preprint derives 2,500 m² of radiator for a 1 MW orbital data centre, and Andrew Cavalier of ABI Research, writing in IEEE Spectrum, puts a 40 kW rack at 80 m². The fourth power is the lever: a radiator at 333 K sheds about 50% more per square metre than one at 300 K (our arithmetic), but without a heat pump it cannot run hotter than the chips it cools.

On the supply side, the paper's solar assumptions (22% efficiency at 1.361 kW/m², with 90% of the array area covered by cells) give about 0.27 kW per square metre, so each kilowatt of compute needs roughly 3.7 m² of panel, plus radiator area of the same order (our arithmetic). The paper says little about thermal design beyond "heat pipes and radiators" and calls it "a critical optimization challenge"; NPR reports that radiators are among the prototype's heaviest components.

Why the satellites fly 100 metres apart

The paper's link budget is the Friis formula for free-space optics:

With 5 W of transmit power, 10 cm telescopes (105.1 dB of gain each) and a wavelength of 1.55 µm, a conventional 5,000 km link receives about 1.6 µW. Off-the-shelf 400G coherent transceivers from data centres want about −20 dBm (10 µW) per channel, or 0.24 mW for 24 wavelength channels. Because received power scales as $1/d^2$, shrinking the distance closes the gap: the paper finds 24-channel wavelength multiplexing, worth 9.6 Tbps per aperture, feasible out to about 300 km.

Below about 5 km a second effect appears. A 10 cm beam no longer spreads appreciably over such a short hop, so smaller telescopes do the same job and several independent links fit in the same footprint: a 2×2 array of 5 cm optics at 1.25 km, a 4×4 array of 2.5 cm optics at 0.32 km. Total bandwidth then scales inversely with distance. For comparison, the paper puts Starlink's laser links at about 100 Gbps over up to 5,400 km, and its own requirement at around 10 Tbps per link. What has been demonstrated is 800 Gbps each way across a short free-space path on a lab bench.

Keeping 81 satellites in formation

The illustrative cluster is 81 satellites within a 1 km radius at a mean altitude of 650 km. Relative motion near a circular orbit follows the Hill-Clohessy-Wiltshire equations, the paper's starting point, with $x$ radial, $y$ along-track, $z$ cross-track and $n$ the orbital rate:

The drift-free solutions are ellipses twice as long along-track as they are tall: $x = A\sin nt$ and $y = 2A\cos nt$. That is why the cluster fits inside an ellipse of ±R along-track by ±R/2 in altitude, why neighbours oscillate between about 100 and 200 m apart, and why the number of satellites grows as $N \sim R^2$. In ideal Keplerian motion this costs no fuel. Earth's oblateness (the $J_2$ term) adds drift, which the paper says can be cut to under 3 m/s per year per kilometre from the reference orbit by nudging the ellipse's axis ratio to 2:1.0037. For control, it suggests backpropagating gradients through the orbit simulation, using a machine-learning framework such as JAX.

Radiation: the numbers

Google irradiated a Trillium (v6e) TPU and its AMD host server with 67 MeV protons at UC Davis's Crocker Nuclear Laboratory while running workloads.

Two readings, both our arithmetic. At 150 rad a year, one silent corruption per 17 rad is about nine undetected wrong results per chip per year: negligible for a chatbot reply, but a 10,000-chip training cluster would see roughly 240 a day. And the weak point for crashes is the ordinary host server, at about 0.7 per server per year, not the TPU. The paper calls the error rate "likely acceptable for inference" and says the impact on training "requires further study".

The $200 per kilogram argument

The cost case rests on Wright's law: price falls by a fixed fraction $r$ with each doubling of cumulative output $M$.

SpaceX's price history gives a learning rate of about 20%. Starting from Falcon Heavy (about $1,800/kg at 400 tonnes of cumulative mass), reaching $200/kg requires roughly 370,000 more tonnes, which is about 1,800 Starship launches at 200 tonnes each, or about 180 a year until the mid-2030s. A separate bottom-up estimate of Starship with 10× reuse gives a customer price under $250/kg.

The paper then compares like with almost-like. A Starlink v2 mini satellite (575 kg, an estimated 28 kW, five-year life) costs $14,700 per kW per year to launch at today's $3,600/kg and $810 at $200/kg, against $570–3,000 per kW per year for electricity at US data centres. The authors state that their estimate "does not constitute a full economic analysis". The comparison covers launch alone: building the satellites is left out, as are the chips.

Closest prior work

Google is not the first to fly an AI accelerator. The startup Starcloud launched a spacecraft carrying an Nvidia H100 last November, NPR notes, and SpaceX has said it expects to start deploying "orbital AI compute satellites" as early as 2028. What distinguishes Suncatcher is the swarm: the paper contrasts its clusters of small satellites with "monolithic" designs larger than any launch vehicle, which would have to be assembled in orbit.

Why it matters

Everyday users. Nothing changes in the products you use; four chips running a small model for a quarter of an hour at a time serve nobody. The stake is indirect: where the power, land and water for AI come from. AWS chief executive Matt Garman wrote this week that more than 100 data-centre moratoriums are being considered across the US. If part of future demand could move off the ground, that pressure eases. If it cannot, it stays local.

Developers and builders. There is nothing to build on yet. The useful lesson is which workloads would suit orbit: those that need a lot of computing but little data moved up and down. As Alan George of the University of Pittsburgh told Scientific American, "the best data compression is answers". The radiation work also has a use on the ground. The paper notes that silent data corruption can occur in any operating environment, and methods for catching wrong answers that raise no error flag are valuable anywhere.

Companies. Google designs the chip, the models and the cloud that would use them, and Futurum's analyst argues that this control is its edge in orbit. But the plan depends on a launch provider with its own ambitions: the paper's cost path runs through SpaceX's Starship, and SpaceX is planning orbital compute of its own. For Planet, best known for its imaging satellites, it is a different kind of customer.

The field. The paper describes its proton-beam data as the first published radiation-testing results for such a device. More data of this kind would turn "can commercial AI chips work in space?" from an argument into an engineering table. The conversation about bottlenecks also moves: from chips, to electricity, to kilograms.

Two second-order effects are worth holding on to. The first is a feedback loop: the paper argues that launch prices keep falling only if someone buys enormous launch mass, and that orbital computing could be that buyer. Its cost forecast therefore partly depends on projects like Suncatcher going ahead. The second is the environmental ledger. Solar-powered satellites still ride rockets that burn carbon-based fuels, and they end their lives burning up in the atmosphere, NPR points out.

What to be skeptical of

A launch is not a result. Google has reported contact and a healthy satellite, not data. The prototype is not in the target orbit, runs on batteries, has no laser links and, per NPR, computes in 15-minute bursts. Beals himself called it "a very minimal test".

Basic facts are thinly sourced. Google's posts do not state the number of chips, the power or the model. Those come from press interviews, and reports conflict: NPR says the chips run Gemma, Scientific American says Gemini. Some outlets place the prototype in the 650 km dawn-dusk orbit, which is the paper's design target; NPR's correction says otherwise.

The cost claim is narrow. "$200 per kilogram makes space comparable" sets launch cost against electricity bills. Satellites, radiators and replacement hardware are extra, and the mass figure is borrowed from Starlink, which is not a data centre. Turyshev's model puts an orbital data centre at 34–59 kg per kW, against about 20 for the Starlink proxy (our arithmetic), and finds that the allowable budget for launch and construction combined is "3.4-13.5 times below" today's Falcon 9 launch price alone. Cavalier's model concludes that running a GPU in orbit for a year costs "at least an order of magnitude" more than on the ground, even at a launch price of $44/kg. Both are models with their own assumptions, not measurements. They are also the only cost estimates here that do not come from Google.

The radiation test has limits. It used one proton energy, on the ground, with the highest dose applied to a single chip, and counts of correctable memory errors "were not reliably available". The paper itself lists galactic cosmic rays as part of the orbital environment. Every radiation number so far is vendor-reported.

Scale is the whole question. Kerri Cahoy of MIT told Scientific American the jump is "from arrays that are several meters on a side to arrays that are a couple of kilometers on a side." Brandon Lucia of Carnegie Mellon told NPR that upkeep and repairs in orbit have their cost and complexity "amplified by a factor of 10, maybe a factor of 100", so "there has to be a big payoff."

We read the preprint. Google says a peer-reviewed version of the paper is now published in Joule. We could not access that version; the figures here are from the arXiv revision dated 17 June 2026.

What to watch next

• The coming weeks: first data. Google's launch post promises to "share more as the mission unfolds". Look for temperatures, measured error counts set against the ground-test predictions, and whether the 15-minute limit moves.
• 2027: two satellites. Google plans to fly a pair to test the laser links, the first check in orbit of an idea so far shown only on a bench at 1.6 Tbps.
• October 2027: the one-year mark. The goal reported by NPR is a year of operation. Whether the chips and the host computer last that long is itself a result.
• Starship's flight rate and prices. The paper's forecast needs about 180 launches a year and roughly $200/kg by the mid-2030s. Both are public numbers that can be tracked.
• Rivals. SpaceX has said it expects to start deploying "orbital AI compute satellites" as early as 2028, and Starcloud flew an Nvidia chip last November. Watch whether anyone publishes thermal and error data rather than announcements.
• Independent checks. Radiation tests of AI accelerators by groups outside the vendors, and responses to the Joule paper's cost model.

For the demand side of this story, the models that make Google want more power in the first place, see our explainer on Gemini 4 Argon.

How this was made: researched and written by an AI model (Claude) from the primary sources listed above, then checked claim-by-claim against those sources in a separate AI fact-check pass. Spotted an error? Email [email protected] and we correct it publicly. Our process.