Sunday, 27 September 2026 News CMS
Cloud & Edge Computing 27 September 2026

Google's Suncatcher Goes to Orbit: Four TPUs, One Kilowatt, and the Hardest Problem in Cloud

By: Tech Desk
Google's Suncatcher Goes to Orbit: Four TPUs, One Kilowatt, and the Hardest Problem in Cloud

Google's Suncatcher Goes to Orbit: Four TPUs, One Kilowatt, and the Hardest Problem in Cloud

Introduction

On 1 October 2026, a refrigerator-sized satellite carrying four of Google's Tensor Processing Units is scheduled to ride a SpaceX Falcon 9 rocket out of Vandenberg Space Force Base into low Earth orbit. It is called MVP, it is built by Planet Labs, and it is the first hardware to fly as part of Project Suncatcher — Google's long-gestating effort to put artificial intelligence data centres in space.

It would be easy to write this as science fiction, and plenty of the coverage has. The reality is more interesting, because Suncatcher is really a story about a thermal engineering problem that the entire cloud industry has been pretending to have solved.

MVP carries four TPU chips. Together they hold roughly the computing power of a single data centre server. The satellite's solar panels will supply about one kilowatt of power — approximately what a hair dryer draws — and that is enough to keep the chips answering short Gemini queries for about fifteen minutes at a time before they must be powered down to cool. Google has ordered the chips to run in bursts. In space, with no air to carry heat away, that is not a design preference. It is the entire operating envelope.

For readers tracking the collision between AI compute demand and physical infrastructure — the same pressures documented in our coverage of rack power climbing to 26 kW and of cities and states pausing data centre permits — Suncatcher is the extreme case of a problem that is already unsolved on the ground.

Four Chips and a Kitchen-Sink Radiator

The interesting engineering is in how Google gets heat off the boards, because a server rack in a data hall sheds heat by pushing hot air into a room that is deliberately over-cooled by chilled water. Vacuum offers none of that. Radiators are the only exit, and radiation is a far worse conductor than convection.

Google's answer, described in detail by The New York Times, is a layered structure that is essentially a refrigerator's cold plate repurposed for the opposite task. The bottom layer holds the TPUs on a green motherboard. Above it sits a "thermal interface material" — a pale green putty sold in sheets — that bonds the chips to layers of aluminium and copper. Those layers conduct heat away from the boards, and the whole stack is capped by a radiator panel that faces deep space and dumps its heat into the vacuum.

The image below is not Suncatcher, but it is the same physical idea at terrestrial scale: dense accelerator boards whose processors sit under copper cold plates, with thick coolant lines running in and out to carry heat to a facility-level chiller. In a data hall, that loop is closed by air and pumps. In orbit, every stage of the chain has to be passive or self-contained, because nothing outside the spacecraft is available to help.

Copper cold plates and coolant lines on a dense accelerator board, the terrestrial equivalent of the heat path Suncatcher must carry by radiation alone

There is a second thermal question that sits underneath the first: how do you power a machine that can only run fifteen minutes at a time? A satellite passes through orbital night on a regular cycle, during which its solar arrays produce nothing at all. A constellation doing useful work has to either accept that duty cycling, or fly five years of stored energy per satellite — which is why the Suncatcher architecture depends on clusters rather than single craft.

The Tests That Happened Before the Rocket

MVP is not flying on faith. Google spent roughly eighteen months trying to break the hardware on the ground, and the results are worth reading closely because they establish the limits the company believes it is working within.

Radiation was the first concern. In February 2025, Google began taking Trillium TPUs to the Crocker Nuclear Laboratory in Davis, California, where a cyclotron — a particle accelerator built with magnets once used in a machine of the Manhattan Project — dosed the chips to the equivalent of five years in orbit. Radiation causes bit flips, the silent corruption of a bit from zero to one, and the test results were encouraging: restarting the chip usually cleared the fault. That recovery property, not raw tolerance, is what makes flight plausible.

Vibration was the second. During ascent, a spacecraft endures loads reaching ten times normal gravity, and individual chip components have been measured at fifty to a hundred times. Technicians shook the satellite along three separate axes before flight, painting a small line across each screw to reveal any movement. 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," said Travis Beals, Google's senior director for the project — a remark that captures how the team itself viewed the result.

Before either of those, the cooling stack went into a thermal vacuum chamber, a chamber that reproduces both the temperature cycling and the near-total absence of convection that a spacecraft experiences. Google has been careful to characterise what that chamber cannot tell you, saying it will learn more from the new cooling system in orbit and refine its designs accordingly.

What none of these tests can simulate is one-year-plus of cumulative degradation, micrometeoroid strikes, or the accumulated fatigue of repeated thermal cycles over thousands of eclipses. That is the honest gap between a passing vibration test and an operational service.

Why Not Just Stay on the Ground

The motivation is arithmetic, and it starts with energy. Blaise Aguera y Arcas, the Google vice-president and AI researcher who is credited with originating Suncatcher, first pitched the idea after attending a gathering whose central topic was the escalating energy appetite of AI systems. He walked away convinced that compute had to move somewhere with unfailing sunlight, and he took the notion to James Manyika, Google's senior vice-president for research.

Manyika's initial reaction, by his own account, was that it would not work — but he agreed to tests that could settle whether AI chips could be cooled in orbit and survive the radiation. The programme was greenlit by Sundar Pichai and Sergey Brin, who attended a May 2025 pitch meeting unannounced.

The strategic case for orbital compute is not really about orbits. It is about two constraints converging on the ground at once. The first is the interconnection and grid build-out documented in the cloud and edge infrastructure coverage: feeding large campuses requires transmission capacity and generation that takes years to permit. The second is local politics, which we have covered in detail as municipal and state-level pushback grows. An operator facing a five-year queue for a substation and a county council that has just refused a permit has, at least on paper, a third option.

A second incentive is the sun itself. Satellites in low Earth orbit can generate up to eight times more solar power per unit of panel area than equivalent arrays on the ground, because they are continuously above atmospheric attenuation. On a constrained launch manifest, power density per kilogram is the real currency.

There is also a decidedly non-technical incentive. Put the compute far enough away and the familiar tools of local resistance — zoning hearings, water usage permits, noise complaints, NIMBY campaigns — stop reaching it. SpaceX has pitched investors on a million-satellite constellation of solar-powered AI satellites, and a trade projection cited this month puts the space data centre market at $28.16 billion. It is worth holding those figures loosely; they are venture-adjacent forecasts attached to programmes that have flown no compute payload between them.

What Orbital Compute Actually Breaks

The appeal is real, and so is the list of things that orbital compute changes about the economics of running a workload. The most obvious is latency, and it is the one most often waved away. A round trip from a low Earth orbit to a ground station is not a short distance, and no amount of inter-satellite laser links makes it one. A constellation in low orbit trades away the property that made terrestrial regions attractive in the first place: a compute node close to the user.

The image below shows the hardware that has to close that gap. Maintaining a link between two spacecraft in formation is, in Google's own words, "similar to hitting a coin-size target from miles away while both points are in motion." The company plans to test that in 2027 with a pair of satellites, and the answer matters more than any single chip test, because without it a cluster of dozens of craft is a collection of islands rather than one machine.

A compact satellite bus with its deployable solar array wings spread on a laboratory bench during preflight testing

Then there is maintenance. Terrestrial operators swap a failed drive in a hot-aisle or a cold-aisle in minutes. Nothing in orbit is serviceable, and a failed cold plate is not a failed drive: it takes the whole compute payload with it. Every reliability figure for a data centre assumes a human with a torque wrench, and the entire Suncatcher architecture has to be re-derived without that assumption.

Radiation hardening, meanwhile, is not a one-time cost. Radiation-tolerant silicon is slower and more expensive than its commercial equivalent, and error correction has to run constantly, stealing cycles from the payload doing the paying work. The bit-flip recovery Google demonstrated in the cyclotron still has to work unattended, in orbit, for years, without anyone available to reboot the machine.

And the economics depend on a launch cost curve that is a promise rather than a history. Google estimates orbital data centres will reach rough cost parity with terrestrial ones in the mid-2030s, as the price of sending satellites to orbit falls. That estimate is doing enormous load-bearing work. It is the difference between an interesting experiment and a business.

Satellite operations staff monitoring telemetry across a bank of workstations, the kind of ground segment that has to exist before any orbital cluster can be operated commercially

A Launch, Not a Milestone

The most useful thing Google has said about Suncatcher was said by Manyika, and it was deliberately deflationary. "We don't expect, to be perfectly frank, that we'll have anything usefully operational in the next few years," he said, comparing the effort to Google's early driverless car work, where roughly fifteen years of research passed before anything usable appeared. He suspects this will look the same way.

That framing is a corrective to a genre of coverage that has treated every orbital data centre announcement as a milestone on a timeline. By Google's own account this is a long-horizon research programme with a one-year prototype that runs its chips for fifteen minutes per cycle. The useful milestones to watch are unglamorous: whether the 2027 pair of satellites establishes a working optical link, whether the cooling stack survives a full year of eclipse cycles, and whether the launch cost curve bends as steeply as the mid-2030s parity estimate assumes.

There is a telling line from Beals, Google's senior director of product management for the project. "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. And on the long run: "If we're really successful with this in the long run, this will ultimately be boring, and people won't think anything of the fact that their Gemini query might be getting served in space."

That is the actual goal, and it is worth being clear-eyed about what it would mean. A Gemini query served from orbit is a piece of infrastructure treated as unremarkable — the same way terrestrial regional cloud capacity and edge nodes are now unremarkable. The measure of success is not whether a refrigerator-sized satellite reaches orbit on 1 October. It is whether anyone notices, in five years, when it does.

Conclusion

On 1 October 2026, a Falcon 9 lifts off from Vandenberg carrying four TPUs, one kilowatt of solar capacity, and a radiator designed to dump heat into a vacuum that will not take it. If the hardware survives, Google will have answered the two questions it set out to answer: whether AI silicon tolerates orbit, and whether it can be cooled there. Both are engineering problems with knowable answers.

Everything after that is not engineering. It is launch economics, inter-satellite link performance, unattended reliability without a technician, and whether a mid-2030s cost parity arrives on schedule. The orbital data centre is not a solution to terrestrial grid constraints. It is an experiment in whether compute can be moved to where the power and the permission both happen to be, in quantities that would justify the moving.

The rocket is scheduled for 1 October. The useful data will not arrive for years.

Images

  • A rocket lifts from a coastal launch complex, the kind of ascent MVP will make from Vandenberg. Photo: NASA, via Wikimedia Commons.
  • Dense accelerator boards under copper cold plates — the terrestrial version of Suncatcher's heat path. Photo: via Wikimedia Commons.
  • A compact satellite bus with deployable solar arrays, on a test bench. Photo: via Wikimedia Commons.
  • Satellite operations staff monitoring telemetry in a mission control room. Photo: Planet Labs, via Wikimedia Commons.

References

Tags: Google Project Suncatcher TPU orbital data center edge computing AI infrastructure SpaceX Planet Labs data center cooling low Earth orbit
← Back to Home