Suncatcher: Google Joins a Crowded Race to Put AI Compute in Orbit
Solar panels in low Earth orbit generate up to eight times what the same panels produce on the ground, and they do it around the clock, with no clouds and no nightfall to plan around. That single fact is the entire premise of Project Suncatcher, a Google Research moonshot laid out in a new paper and blog post from Travis Beals, the company's senior director of Paradigms of Intelligence: pack Google's Tensor Processing Units into a tight constellation of satellites, wire them together with lasers, and let the sun do what a power grid can't.
Google has been here before, in spirit if not in altitude. The same post points back to the company's quantum computing program, started roughly a decade ago when a large-scale quantum computer wasn't considered a realistic engineering goal, and to the self-driving project that eventually became Waymo. Both took years of unglamorous groundwork before they looked plausible to anyone outside the building. Suncatcher gets pitched the same way, as a bet with real open questions rather than a finished product on a shelf.
The timing isn't incidental, either. Data centers on the ground are running into power-grid limits in market after market, and building new grid capacity or a gas or nuclear plant to feed one takes years of permitting before a single chip powers on. A satellite doesn't stand in that line. It doesn't compete with a neighborhood for grid interconnection, and its power source doesn't dim at night or during a heat wave. That's the specific bottleneck Google is aiming at, rather than some broader claim that computing in space is simply cheaper.
What Google Actually Tested
The paper, "Towards a future space-based, highly scalable AI infrastructure system design," lists Beals alongside eight co-authors, including Google VP Blaise Agüera y Arcas, and it's organized around four things Google needed evidence for before the concept could be called plausible rather than speculative.
The first is bandwidth. Training a model across many chips means moving data between them fast, and Google's target for satellite-to-satellite links is tens of terabits per second, roughly what a terrestrial data center's internal network handles. A bench-scale demonstrator already reached 800 Gbps in each direction between one pair of transceivers, 1.6 Tbps combined, using dense wavelength-division multiplexing hardware paired with spatial multiplexing to pack more channels through the same beam. Because signal strength drops with the square of distance, hitting that speed over any real range means flying the satellites within a kilometer of each other or closer, tighter than the spacing most existing satellite constellations, including Starlink, use for routine collision avoidance.
That closeness raises a second question: keeping dozens of satellites from drifting into each other or losing their laser lock. Google built physics models, starting from the standard equations for relative orbital motion and refined with custom numerical work, and tested an 81-satellite formation at 650 kilometers altitude with a one-kilometer cluster radius. The models suggest Earth's own gravity does most of the work of holding the formation together, and the satellites would need only modest station-keeping, rather than constant active correction. Next-nearest neighbors in that simulated cluster oscillated between roughly 100 and 200 meters apart over a single orbit, close enough that the whole exercise reads more like choreography than traffic control.
Radiation was the third question, and the one where Google seems to have gotten a pleasant surprise. Firing a 67 MeV proton beam at a Trillium v6e TPU, the most sensitive component, its onboard memory, didn't show trouble until 2 krad(Si), nearly three times the roughly 750 rad(Si) Google expects a shielded five-year mission to accumulate. Nothing failed outright up to the maximum tested dose of 15 krad(Si). Google calls the chips "surprisingly radiation-hard," a result the paper is careful to note wasn't something the hardware was specifically engineered for.
The fourth question is money. Launch prices have already fallen sharply as reusable rockets replaced expendable ones, and Google's modeling extends that curve, projecting costs under $200 per kilogram by the mid-2030s. At that price, the company estimates a space-based data center's cost per kilowatt-year could land close to a terrestrial one's, once savings on cooling, land, and batteries get factored in. That comparison rests entirely on Google's own extrapolation of launch-cost trends and a footnoted, published estimate of terrestrial energy costs, rather than on a price anyone has actually paid for operating a data center in orbit.
The Rest of the Sky
None of this is happening in a vacuum, so to speak. By the time Google's two Planet-built prototypes fly in early 2027, at least one competitor will already have years of orbital operating data. Starcloud, a Redmond startup that went through Y Combinator in 2024, put its first satellite in orbit in November 2025 carrying an Nvidia H100 GPU, the first data center-grade GPU to run in space. The company has since raised $170 million at a $1.1 billion valuation and filed with the FCC for a constellation of up to 88,000 satellites.
The rest of the field is moving too. SpaceX, which acquired Elon Musk's AI company xAI in February 2026, has filed for permission to build a distributed-computing satellite network of its own. Jeff Bezos has said publicly that he expects gigawatt-scale data centers in orbit within a decade or so for Blue Origin. Former Google CEO Eric Schmidt has gone a different route entirely, acquiring the rocket company Relativity Space in 2025 and steering it toward orbital data centers under his own banner rather than inside a hyperscaler.
Next to all of that, Google's posture reads as the more deliberate one in the group, a peer-reviewable paper backed by proton-beam data and orbital-dynamics modeling, published well before any hardware of its own has flown. Whether that patience pays off once Starcloud and the others have more flight hours in the log is a question the market will settle over the next few years.
Beals is candid about what the paper doesn't answer. Thermal management in a vacuum, where there's no air to carry heat off a chip, hasn't gotten a bench demonstration the way the optical link and the radiation test have. Neither has ground-to-orbit communication at the bandwidth a real workload would need, or basic reliability over a mission that runs several years rather than a few months. Those gaps put the harder parts of the problem a step behind the results Google is actually reporting today.
The paper closes on a longer-term idea rather than a promise: satellites built from the ground up so solar collection, compute, and cooling are one integrated system instead of separate parts bolted together, the way a smartphone's chip design eventually grew out of tight integration rather than assembled components. That stays speculative until the 2027 flights happen, and Beals doesn't offer a date for when the idea might move past the concept stage. The full paper and Google's own technical detail are available directly from Google Research for anyone who wants to check the modeling themselves, rather than take the summary on faith.
Author Bio
David Borish is the author of The Tony Hawk Paradox: When Video Games Predict Reality and publisher of The AI Spectator. More of his writing is at davidborish.com
