Data centers won’t be located in space any time soon

Can AI’s insatiable thirst for massive data centers be fixed by launching them into space? Technology companies are eyeing low Earth orbit as a potential solution, but researchers say that is unlikely to happen in the near future due to a mountain of difficult unsolved engineering problems.

The massive demand for and investment in generative AI products, such as ChatGPT, has created an unprecedented need for computing power, requiring massive amounts of space and gigawatts of power, equivalent to that used by millions of homes. As a result, data centers are increasingly fueled by unsustainable sources, such as natural gas, with technology companies claiming that renewable energy can neither produce the amount of energy needed nor the consistency required for reliable use.

To solve this problem, tech CEOs like Elon Musk and Jeff Bezos have proposed launching data centers into orbit, where they could be powered by solar panels and have constant access to a higher level of sunlight than on Earth. Earlier this year, Bezos, who along with the Amazon founder owns the space company Blue Origin, said… It envisions gigawatt data centers in space within 10 to 20 years.

Google has more realistic and accelerating plans for data centers in space, with an experimental program called Project Suncatcher that aims to launch two prototype satellites carrying TPU AI chips in 2027. Perhaps the most advanced experiment in data processing in space to date, was the launch of a single H100 GPU this year by an Nvidia-backed company called Starcloud.

This is nowhere near enough computing power to run modern AI systems. For example, OpenAI is believed to have a million such chips at its disposal, but getting to that scale in orbit will require technology companies to address a number of challenges that have yet to be solved. “From the point of view of academic research, [space data centres] “It’s nowhere near production level,” he says. Benjamin Lee At the University of Pennsylvania in the United States.

One of the biggest problems that doesn’t have a clear solution, Lee says, is the sheer physical scale involved in the computational demands of AI. This is due to the amount of energy required from the solar panels, which would require a large surface area, and the necessity of radiating the heat produced by the chips, which is the only option for cooling in space, where there is no air. “You’re not able to cool it by evaporation like you can on Earth, by blowing cold air over it,” Lee says.

“Square kilometers of space will be used independently for power generation, as well as for cooling,” says Li. “These things get very big, very quickly. When you’re talking about 1,000 megawatts of capacity, that’s a very large area in space.” In fact, Starcloud says it plans to build a 5,000-megawatt data center spanning 16 square kilometers, or about 400 times the area of ​​solar panels on the International Space Station.

He says there are some promising technologies that could reduce this requirement Krishna Muralidharan At the University of Arizona in the United States, such as thermal devices that can convert heat back into electricity and increase the efficiency of chips operating in space. “It’s not a problem, it’s a challenge,” he says. “For now, we can solve the problem with these large radiator panels, but in the end it will require more complex solutions.”

But space is a very different environment from Earth in other ways as well, including an abundance of high-energy radiation that can hit computer chips and upset calculations by introducing errors. “It will slow everything down,” Lee says. “You’re going to have to rerun the computation, you’re going to have to recover those errors and correct them, so there’s probably going to be a reduction in performance for the same chip in space compared to what’s deployed on Earth.”

The range will also require thousands of satellites flying together, which will need very precise laser systems to communicate between data centers and with Earth, where the light will be partly mixed by the atmosphere, Muraleedharan says. But Muraleedharan is optimistic that these are not fundamental problems and can eventually be resolved. “It’s a question of when, not if,” he says.

Another uncertainty is whether AI will still need such massive computational resources by the time space-based data centers become available, especially if expected advances in AI capability do not match increased computational firepower, of which there are some early signs. “There is a distinct possibility that training requirements will peak or stabilize, and then the demand for large-scale and large-scale data centers will also peak and then stabilize,” says Lee.

However, there could still be uses for space-based data centers in this scenario, Muralidharan says, such as supporting space exploration on the moon or in the solar system, or to conduct Earth observations.