Google Puts Four AI Chips in Orbit, Lets Them Work 15 Minutes at a Time
Project Suncatcher's first satellite is in space, where compute is abundant, sunlight is free, and the main engineering challenge is that there is nowhere for the heat to go.
On October 1, a refrigerator-sized satellite built with Planet Labs rode a SpaceX Falcon 9 out of Vandenberg Space Force Base carrying four of Google’s Trillium TPUs. It is the first piece of hardware from Project Suncatcher, Google’s plan to someday run AI data centers in orbit, and it has already taught the industry its first lesson about space: it is very good at being cold and extremely bad at letting you get rid of heat.
Google, which has spent the last several years explaining that the limiting factor for AI is electricity, has found a place where electricity is plentiful. Satellites in low Earth orbit could generate up to eight times more solar power than comparable systems on the ground, Google says, per Interesting Engineering. The sun, famously, does not file interconnection requests.
The catch is that a data center is, at the physical level, a machine for turning electricity into warmth, and warmth needs somewhere to go. On Earth, that somewhere is air. In orbit, there is no air. There is only vacuum, which is an excellent insulator and therefore the world’s worst heat sink. The satellite’s chips will reportedly run a version of Google’s open-weight Gemma model in bursts of about fifteen minutes, then stop so the heat can bleed off. This is not a data center. This is a shift schedule.
Consider the arrangement. Four of the most advanced AI accelerators humans have built, flung into low Earth orbit at enormous expense, will work for a quarter of an hour and then sit quietly in the dark and think about what they’ve done. The labor practices are, frankly, the most humane in the industry.
The satellite is called an MVP, as in minimum viable product, which in this context means “will it survive.” According to Startup Fortune’s reporting, engineers plan roughly a year of observation, watching for radiation damage to the silicon and thermal stress from the swing between sunlight and shadow. Before launch, the chips were blasted with a proton beam at UC Davis and reportedly tolerated the radiation expected over a five-year mission, with memory irregularities showing up only at much higher doses. The chips have therefore been bombarded with particles on Earth in order to prepare them for being bombarded with particles elsewhere. It’s a character-building exercise.
Interesting Engineering notes that Google is pairing heat pipes with radiators, a system that passed thermal vacuum testing on the ground. The orbital flight is its first real-world exam. Google confirmed contact after launch and said the satellite was operating as expected, a sentence that, in the history of space hardware, has occasionally been followed by a second sentence.
The roadmap is a staircase. A two-satellite constellation with Planet is planned for early 2027, to test whether TPUs in different spacecraft can coordinate over laser links. After that, the long-term vision, described in a paper in the journal Joule, is tightly packed clusters of 81 satellites. Each step is a small, sensible increase over the last, in the way that going from a paper airplane to the Moon is a series of small, sensible increases.
Then there’s the math. Google’s own analysis, as reported by Startup Fortune, says the economics only work if launch costs fall to about $200 per kilogram by 2035, which requires on the order of 1,800 Starship launches over roughly a decade. Starship has not yet demonstrated orbital refueling or that kind of launch cadence. So the business case rests on a rocket that does not yet do the thing the business case needs it to do, launching, eighteen hundred times, a product that currently works for fifteen minutes.
To be fair to Google, this is how every infrastructure project begins: a prototype that shouldn’t work, a spreadsheet that shouldn’t balance, and a press release that says “first step.”
What makes the story endearing is how reasonable each piece is on its own. The power problem is real. The cooling problem is real. The idea that you’d solve one by taking it to a place that makes the other catastrophically worse is, in its way, deeply Silicon Valley, a field that has long believed any problem can be fixed by relocating it to a more difficult one.
There’s also a comfortable symmetry in the first payload. After years of data-center construction that has local governments, grid operators and utilities negotiating over every megawatt, Google has launched a data center so small and so temporary that it needs no zoning approval, only a rideshare slot on SpaceX’s Transporter-18 mission. It is the ultimate in permitting efficiency: nobody can object to your chips because nobody is there.
In the meantime, down here, the rest of the industry is still building data centers out of concrete and ambition. Space will have to wait, and then, apparently, work in fifteen-minute increments. Four TPUs, no breaks, a year of testing, and a sky full of free electricity they can’t quite use fast enough. Someday soon they may even be asked to answer a question. We’ll let you know how many minutes it takes.
Sources: