The News
The demand for AI data centers has grown so large that there isn’t enough energy on earth to power the buildout.
Google’s engineers think they may have an answer: Put some of them in space. The idea could help Google meet a portion of the soaring demand for compute, but also reduce its carbon footprint on Earth, where data centers are increasing greenhouse gas emissions and have drawn criticism from environmentalists.
In a research paper published Tuesday, Google outlined plans to launch its AI chips, known as Tensor Processing Units, into a low-earth orbit that keeps them constantly in view of the sun, allowing them to continually run off solar power.
The TPUs, attached to satellites, would be deployed as a constellation where they’d travel close enough together for high-bandwidth wireless communication, allowing the satellites to work in concert just like the AI chips in today’s earthbound data centers.
Google, which dubbed the effort “Project Suncatcher,” plans to launch two test satellites, each carrying four TPUs, in 2027 with the company Planet Labs. As space transport comes down in price, the team believes the idea could become economical by 2035.
“If things keep going down the path where we keep having more uses for AI and we keep wanting more energy to power it, this has tremendous potential to scale,” said Travis Beals, senior director for Paradigms of Intelligence, a research team within Google. “But as with all the moonshots, there’s no certainty.”
The advantage of running a data center in space is that it can be powered entirely by the sun. Terrestrial data centers that run off solar panels lose that power at night and production drops significantly during the winter.
Google’s plan is to launch the satellites into an orbit that would keep them in nearly continuous sunshine and receive up to eight times more solar power per year than a panel located on earth at mid-latitude.
But running a data center in space creates other challenges beyond the initial hurdle of getting them up there in the first place.
In space, Google’s TPUs will be exposed to much higher levels of radiation than they would on Earth, and researchers need to figure out how long they will last. They took the chips to a facility at University California, Davis and used a particle accelerator to irradiate the processors to simulate years of solar exposure in space.
“They held up quite well,” Beals said, suggesting they could easily handle a five- or six-year mission. Further testing will determine whether there are any subtle issues caused when chips operate in space, like subatomic particles colliding with transistors and causing too many “bit flips,” when binary numbers reverse in value.
In this article:
Know More
In today’s data centers, AI chips are connected using copper or optical wiring. To do that in space, someone would have to either launch an object much larger than modern rockets can handle or assemble one in space from smaller parts.
A startup called Starcloud, which has partnered with Nvidia, plans to build what look more like traditional data centers in space, with modular containers filled with server racks.
Google’s idea is to take a much simpler approach by launching fleets of small satellites and connecting them wirelessly, reducing the need to carry large, bulky server racks on rockets.
Google found that if the satellites are positioned hundreds of meters apart they can use an optical connection that achieves fast enough speeds, with current tests showing about 1.6 terabits per second.
Current satellite constellations, like Starlink, are roughly 120 km apart, by some estimates. Google used physics modeling to show how it was possible to keep them in a tight formation using modern thruster technology.
The approach is in character for Google, which has historically looked for ways to lower hardware costs in data centers with a practical engineering mindset. Its early servers were built as cheaply as possible and famously used cork board as a way to separate components stacked on top of each other. It then built customized software to work around any issues with the subpar hardware.
Even if Google’s satellite data center concept worked, it would be too expensive today to launch the equipment into space. But the price is coming down quickly: Google researchers estimated that today’s roughly $1,500 per kg on a SpaceX Falcon Heavy rocket could drop to about $200 by 2035, making a space-based data center on par with a terrestrial one on a cost-per-kilowatt basis.
“We’ve spent the past year or so trying to think through, what are all the ways this might not work? Can we prove it can’t work? And we’re still here because we haven’t seen any obvious showstoppers,” Beals said.
Reed’s view
Google is famous for its “moonshots,” or projects that are so ambitious that they have high potential reward but a low probability of success or of ever generating revenue for the company. The concept comes straight out of the brains of its eccentric founders, Larry Page and Sergey Brin, who saw the massive profit margins from search advertising as a means to build something much more interesting than an internet cash cow.
The company has embarked on many moonshots with a wide range of difficulty. There was Project Calico, tasked with solving death. Project Loon was going to use high-altitude balloons to deliver internet connection to remote areas of the world. Verily wanted to make glucose-monitoring contact lenses.
Those ideas have not yet panned out. Others, like self-driving cars and advanced artificial intelligence, are beginning to look like the future of the company.
For several reasons, the “data centers in space” idea seems like one of the lower probability ideas, like it took the term “moonshot” too literally.
In a decade, there might be enough power to reduce the cost of a terrestrial data center, making it impractical to put them into orbit. The technology inside data centers could change, creating unforeseen challenges. And then there’s the basic fact: Stuff goes wrong in space.
On the other hand, we are living in highly-capitalized, moonshot times. The idea of building data centers with millions of AI chips — even on earth — would have sounded nuts even a few years ago.
A decade from now, it might be commonplace to supplement data center supply with some cheap, green, space-based inventory.
And maybe the idea is not ambitious enough. The need is here today, and if AI is important to the national security of the US, perhaps the government should accelerate the timeline by contracting with Google and SpaceX to make this happen by, say, 2030?
Just like the first moonshot — the one where Americans landed on the actual moon — it would surely lead to new technology and economic benefits down the road.
Room for Disagreement
A European Union-funded study questioned the net-positive carbon impact of data centers in space, considering the carbon footprints of rocket launches and the upper atmosphere emissions that can be particularly harmful. “Indeed, the results of our study indicate that data centers in space become more interesting than terrestrial data centers if and only if the launcher can be reusable and that it emits less than 370 kgCO2/kg of payload on average over its entire lifespan,” the researchers wrote.
Note: SpaceX’s under-development Starship may significantly improve on the amount of CO2 and other pollutants emitted during launch.
Notable
- Lonestar Data Holdings is taking the moonshot concept to its logical conclusion and plans to put data centers on … the moon!