Image
The AI Race Is Moving to Space
When you think about data centers, outer space might not be the first thing that comes to mind. But for major tech companies and aerospace firms, space may be the next frontier.
With popular opinion of data centers declining amid rising concerns about power infrastructure, water scarcity, and environmental degradation, deploying data centers in orbit is becoming closer to reality.
Peter Beling is the interim associate dean of research and a professor at the University of Virginia School of Data Science, where he leads generative AI research at the National Security Data and Policy Institute.
We sat down with him to discuss why major companies have set their sights on space, what it would entail to put data centers in orbit, and what’s at stake.
What are the advantages of data centers moving to space?
It’s becoming clear that the demand for AI data centers is growing much faster than our capability to add capacity in terms of energy on the ground. That’s what’s motivating the idea of thinking about space. That, and other environmental factors associated with data centers.
The main attraction of space is unlimited energy, basically, in the form of solar. If one can get solar panels in space, they can be the source of energy that comes from sunlight hitting them without regard to weather, pollution — all the things that make it more difficult to do solar on the ground.
What would putting data centers in space entail, and what would this technology make possible?
Traditionally, it’s been a very expensive process to move a lot of mass up into space from Earth. In the past, rocket boosters would be dropped as the craft rose, and they would fall into the ocean or burn up, leaving very little that was reusable.
Now, SpaceX has gotten to the point where their small rockets are essentially fully reusable. They can fly in and land — just like in the movies — vertically on the pad and get caught by a pair of mechanical arms. And they can do that with the booster section of their biggest rocket called Starship. Starship is the biggest rocket and space launch project ever built in human history. And now the entire Starship, including the booster and spacecraft — is on the verge of being fully reusable.
If they can land the spacecraft as well as the booster on their own, we’re very close to having almost aircraft-like reusability. That would open the possibility of moving millions of tons of cargo into orbit, which is what you need to have when you think about data centers in space. That’s the goal. We may be very, very close to that. Maybe within a year or two.
If SpaceX creates a reusable rocket, one thing we’ll see is a much greater impact on communications here on the ground. They'll be able to put all kinds of communications satellites into space, above and beyond the 20,000 or so that are currently there. Eventually, there could be enough satellites to completely cover the Earth, to the point where we'll all essentially have satellite phones in our pockets. We won’t need to talk to cell phone towers anymore. Your phone will talk directly to a satellite, and internet traffic will end up floating around space through this great network carrying all of our data.
What security concerns and other issues could arise?
You might wonder what happens if something crashes into those satellites. What if someone wanted to take out internet traffic or disrupt all those data centers? Could they do it by blowing up a satellite and creating a debris field that crashes into everything?
And the answer is that with so many satellites, you get a kind of resilience against disruption, whether it's orbital debris crashing into satellites, somebody deliberately targeting and destroying them, or even a cyberattack that takes over a fraction of the network.
If you have a communication network that has a million different nodes that traffic flows through, then you have all kinds of ways to reroute traffic on the fly. If you lose some part of the network, disrupting the entire system becomes much more difficult. Cyberattacks that take over software are always a risk, but that kind of risk is also present in all our systems.
There’s one other factor to consider: nature. Space is filled with cosmic rays, little bits of radiation traveling through space. Here on Earth, we’re protected by Earth’s magnetic field, but in space, the chips doing AI work will have a lot less protection. When cosmic rays hit the chip, they can disrupt the computations and even change values stored in memory.
That creates another technical challenge, and it's one that people are actively working to solve. There are a number of solutions, and researchers in the UVA School of Data Science have proposals out right now that focus on radiation shielding and resilience of AI in orbit.
The 20th-century space race had a clear finish line: land a person on the moon and return them safely to Earth. As the AI race increasingly extends into space, what would it mean to win?
There are two schools of thought. One is that we’ll know we've won when everyone in the world is using U.S. models, we’ve made them open source, and everyone’s relying on U.S. technology. The other school of thought is: forget open source. We want to have the best stuff.
My view is that the space race will ultimately be dominated by the launch costs. Whoever has the lowest cost will be able to do more in space. And there’s more that we can do in space than we can do on the Earth, so that’s the key.
Right now there are only three entities that can think about heading down that path: SpaceX, Blue Origin, and China. Reusability of the rocket is key, and SpaceX is very, very far ahead of everyone else. Blue Origin is starting to see larger successes, but it has also had spectacular failures. China recently succeeded in landing a small rocket after at least one failure. It’s a very difficult proposition. Everyone needs to go through a lot of failures to get to success.
If the U.S. were to win the AI space race, is it possible to win by building systems that are fast and capable but not actually trustworthy? What would that cost us down the line?
The whole thing driving the AI race is that there’s no fixed point with AI. As we look at one of the current models and start thinking about all the things that we can do with it, we'll come up with something that wasn't imagined before: a new company, a new idea, a new service. We’ll use AI for that, and that will create a demand for even more and even better AI. There’s a feedback loop: the more useful AI becomes, the more we want to use it.
So how can we test a particular AI model or system and say it will work fine when we can’t even imagine all the uses people will put it to in the future? We can try to lock down those uses, but we won’t really succeed.
Researchers in the School of Data Science are studying the concept of resilience for AI. Resilience starts with the recognition that these AI models and systems are potentially untrustworthy. What we can do is observe their behavior, put guardrails in place, and develop the capability to switch them off or try to limit any damage from inappropriate behavior.
Resilience is a notion you could apply to almost any system. But for AI, the questions are: How do you anticipate how these systems might behave? How can you monitor them? How do you put guardrails and mitigations in place to limit damage if they’re not performing the way we want? That’s an active area of research with several of our faculty and labs.
As data scientists, we have to be aware that when we learn something, the very act of learning it can end up being exploited. We have a lot of researchers at the School who deal with topics like the ethics of data science and ask: Is there any harm in learning something?
The harm really comes when what is learned is acted upon, and we don’t really understand the consequences of that action. It gets into pretty deep philosophy, but it also gets to what makes data science different from other fields. We care about those questions.
