Google has put a refrigerator-sized satellite into orbit, carrying four computer chips made for artificial intelligence. The effort is known as Project Suncatcher, and it serves as an early move toward data centers placed in orbit that run on power from the sun.
The satellite holds four Tensor Processing Units, or TPUs, which Google designed for machine learning. These chips already run in data centers on Earth. For now, they will run a version of Google’s Gemma AI model for just 15 minutes at a time, due to heat limits. Google said the mission will measure how the TPUs handle “the physical stress of spaceflight and the radiation and thermal extremes of space.”
Why Google Is Looking Up
In recent years, the demand for AI services has risen quickly. Sundar Pichai, the CEO of Google, made that point at his company’s annual developer gathering in May. He noted that demand for AI services now outstrips supply. He also said he expects capital spending this year to reach $180 billion to $190 billion — roughly six times larger than it was in 2022 — as the company expands its computing capacity and works on new chip designs.
Data centers on our planet are increasingly coming under fire from those who object to their demands for space and electricity. Up above the atmosphere, though, solar power is nearly limitless and there is nobody there to protest against their presence.
The Sun As Power Source
Project Suncatcher’s senior director and lead, Travis Beals, offered an explanation for why the initiative exists.
“The sun puts out almost all of the power in our solar system. All of the other power sources that humanity has tapped into are just a tiny fraction of a percent,” Beals said. “So in some sense, this project is about tapping into the best way to use solar power to run AI compute.”
The spacecraft will occupy a sun-synchronous orbit, which means its solar arrays will nearly always face the Sun. Because of that, the vehicle will hardly ever lack sunlight, so there is no need for large batteries or a secondary power source.
What Happens Next
In November last year, Starcloud sent a spacecraft into orbit, carrying an Nvidia H100 chip inside it. Once aloft, that vessel operated a variant of Google’s Gemini AI from up above the Earth.
SpaceX, run by Elon Musk, has its own plan for space data centers. The company anticipates beginning deployment of those satellites as soon as 2028, with “orbital AI compute satellites” being the name of the system.
The satellite imaging firm Planet joined forces with Google for this launch. The idea behind it is clusters of satellites circling the planet together, serving as space-based data centers. Every one would hold dozens of TPU chips and stay in touch with the rest of the group, along with Earth itself, through lasers. By next year, the plan calls for two additional satellites to be placed in orbit, to check how they connect.
Testing The Chips
Today’s launch sends a prototype satellite into orbit. Beals referred to it as “a very minimal test”, meant to verify whether the chips can operate in space. A SpaceX rocket delivered it there, and Planet will prepare it for service. Google will switch it on and put the TPUs through their paces. The plan calls for the mission to endure a full year.
A professor of electrical and computer engineering at Carnegie Mellon University, Brandon Lucia, said the concept sounds “very sci-fi,” though much of the underlying technology has already been demonstrated at a research level.
The Hard Part Is Cooling
High temperatures pose a serious issue for hardware, especially for computer chips that generate significant warmth during operation. On our planet, natural cooling methods such as air currents and liquid flow can help manage this excess heat. But those same approaches fail to function in the empty, airless environment of space.
“In a satellite in particular, using more power to do the computations means dumping more heat into the confined environment inside of the satellite,” Lucia said.
In its blog post, Google referred to cooling the chips as “a crucial research challenge”. The company detailed several methods for achieving it, among them pipes and radiators designed to pull heat away from the chips.
The radiators are among the heaviest components of the present mission, according to Beals, and that weight comes with a price: heavier items cost more to send into space. His team is looking for better ways to cool the chips instead.
Upkeep And The Environment
Maintenance concerns loom large as well. A human worker can enter a data center on Earth and repair a broken chip. The task grows much more difficult when that data center sits hundreds of miles above the planet’s surface.
Lucia spoke the words herself, saying “Those all have their cost and complexity amplified by a factor of 10, maybe a factor of 100. And so there has to be a big payoff,”.
Another issue worth raising concerns the environment. The rockets that send satellites into space rely on carbon-based fuel. In time, those satellites will drift back down through Earth’s atmosphere and burn up.
The Economics Of Orbit
Elon Musk has said that space is the “only way to scale” data centers to meet future demand for AI. But the economics are still unclear. Launches are costly, repairs are difficult, and sending signals by laser across hundreds of miles and through Earth’s atmosphere is slower. It will take many satellites to match the computing power of a single data center on Earth.
Beals said his small team thinks about these questions a lot.
“It doesn’t help a lot if this is technically possible, if it’s always going to be too expensive to be practical,” he said.
It is difficult to say when space-based data centers will become cost-effective, he said.
“I don’t see this being something where it’s cheaper to do this in the next five years. I think it will take longer than that,” he said.
There is no expectation that space data centers will take the place of data centers on Earth anytime soon.
“I think that there are going to be use cases that make more sense to do on Earth and use cases that make more sense to do in space,” Beals said. “In the really-long run, perhaps the set of things you do in space gets larger and larger until that’s where most of the compute is happening. But I think for a long time you’ll have both coexisting.”
The whole project, he said, is a “moonshot.”
“It’s going to take a long time,” he said.
Key facts
– TPUs: 4 per satellite
– Mission runtime: 15 minutes per Gemma AI session
– Capital expenditures: $180B to $190B this year, more than 6x 2022
– SpaceX orbital AI compute satellites: expected as early as 2028
– Mission duration: expected to be operational for a year
Source material: “Google launches Project Suncatcher, a step towards AI data centers in space,” NPR.
Get the Notebook.
The day's best stories and every fresh verdict, in plain English, in your inbox by seven. One email a day, no more.