Starcloud wants to build a data center satellite measuring 4 by 4 kilometers.
star cloud
Can A.I. insatiable thirst fix colossal data centers by launching them into space? Technology companies are looking at low-Earth orbit as a potential solution, but researchers say it's unlikely in the near future due to many complex and unsolved engineering problems.
huge demand Investments in generative AI products like ChatGPT have created an unprecedented demand for computing power that requires both massive space and gigawatts of power equivalent to that used by millions of homes. As a result, data centers are increasingly powered by intermittent sources such as natural gas, and technology companies argue that renewable energy can neither produce the amount of power needed nor the stability needed to use it reliably.
To solve this problem, technology executives such as Elon Musk and Jeff Bezos have proposed launching data centers into orbit, where they could be powered by solar panels with constant access to higher levels of sunlight than on Earth. Earlier this year, Bezos, who along with the Amazon founder also owns the space company Blue Origin, said that he envisions gigawatt-scale data centers in space within 10–20 years.
Google has more specific and accelerated plans to build data centers in space. The pilot program, called Project Suncatcher, aims to launch two prototype satellites with TPU artificial intelligence chips in 2027. However, perhaps the most advanced data processing experiment in space to date has been the launch of a single H100 GPU this year by Nvidia-backed Starcloud.
This computing power is clearly not enough to run modern artificial intelligence systems. For example, OpenAI is believed to have a million of these chips at its disposal, but achieving that scale in orbit will require tech companies to address a number of outstanding challenges. “From an academic research perspective, [space data centres] are far from production levels,” says Benjamin Lee at the University of Pennsylvania, USA.
According to Lee, one of the biggest challenges with no obvious solution is the sheer physical size posed by AI's computing needs. This is due to both the amount of energy that would be required from the solar panels, which would require a huge surface area, and the need to radiate the heat produced by the chips, which is the only cooling option in space where there is no air. “You can't cool them through evaporation like you can on Earth by blowing cool air over them,” Lee says.
“Square kilometers of the site will be used independently for both energy production and cooling,” says Lee. “These things get pretty big pretty quickly. When you're talking about 1,000 megawatts of power, that's a lot of power in space.” Indeed, Starcloud says it plans to build a 5,000-megawatt data center that will cover an area of 16 square kilometers, about 400 times the area of the solar panels on the International Space Station.
There are several promising technologies that could reduce this requirement, he said. Krishna Muralidharan at the University of Arizona, USA, for example, thermoelectric devices that can convert heat back into electricity and improve the efficiency of chips operating in space. “It's not a problem, it's a challenge,” he says. “Right now we can solve this problem by using these large radiator panels, but ultimately it requires much more complex solutions.”
But space is very different from Earth in other ways, including the abundance of high-energy radiation that can attack computer chips and disrupt calculations, causing errors. “This will slow everything down,” Lee says. “You'll have to restart the calculations, you'll have to recover and correct those errors, so the performance of the same chip in space will likely be lower than on Earth.”
That scale would require thousands of satellites to be launched simultaneously, which would require extremely precise laser systems to communicate between data centers and Earth, where light would be partially encrypted by the atmosphere, Muraleedharan said. But Muraleedharan hopes these are not fundamental problems and can be resolved over time. “It’s a question of when, not if,” he says.
Another uncertainty is whether AI will still require such massive computing resources by the time space-based data centers become available, especially if projected advances in AI capabilities do not scale with the increased computing firepower that exists some early signs from. “It’s likely that training needs will peak or stabilize, and then demand for larger, larger scale data centers will also peak and stabilize,” Lee says.
However, in this scenario, space-based data centers could still be used, Muraleedharan says, for example to support space exploration on the Moon or solar system, or to conduct Earth observations.
Topics:
