太空数据中心：AI算力的未来解决方案？

When you ask AI to answer a question, generate an image, or write a line of code, somewhere, a machine kicks into gear. A neural network of high-performance computing in data centers that consume huge swathes of land. There are now thousands of them around the world. And they’re getting harder to sustain. These AI data centers just require an enormous amount of power in order to run them, also to cool them. And so they’ve essentially hit what is an “energy wall.”

And yet, demand keeps accelerating. Globally, electricity use for data centers is expected to double by 2030. By 2050, it will likely represent a tenth of all electricity consumed on Earth. Some forecasts suggest that over just the next three years alone, we will need an additional 50 to 100 gigawatts of new energy for new data center capacity just in North America alone. For context, one nuclear power station is about a gigawatt so you’d be looking at 50 to 100 new nuclear projects in the US.

So what happens when artificial intelligence outpaces the power and land needed to sustain it? 3, 2, 1. The answer may be above Earth. Engines full power. And liftoff. In late 2025, space startup Starcloud, sent a satellite into orbit carrying an Nvidia H100 chip — the most powerful processor ever deployed in space. Starcloud 1. Separation confirmed. A machine only the size of a small fridge, representing a big ambition: to move data centers off Earth.

Until we launched Starcloud 1, people thought that you couldn’t run Earth-based, terrestrial, state-of-the-art GPUs or AI chips in space so we did a lot of engineering work based around the thermal system and around the radiation tolerance and shielding in order to make these chips work in space. This is the Starcloud 1 satellite, so you can see here we’re just flying over Africa. It takes about 90 minutes to orbit the globe once.

Starcloud 1 is not a data center — not yet. It’s a single prototype in orbit. The future that Starcloud imagines is tens of thousands of satellites, each carrying chips like this one, working together to do complex, large-scale computing in space — also known as “orbital compute.” For many, the vision begins with a prompt from Earth, beamed into space through laser links. The request is then processed by AI chips inside a satellite bus, which is powered by massive solar arrays stretching as much as four square kilometers.

Together, these satellites form a linked computing network that then sends results back to Earth in milliseconds. Two years ago, when we first started talking about building data centers in space, people, you know, frankly thought we were a bit crazy. And it does sound a little bit sci-fi. But what sounds like sci-fi is a response to a very real problem. On Earth, data centers require vast amounts of land, water and electricity.

And as the AI systems inside them grow more powerful, the heat they generate is becoming harder and costlier to contain. In space, on the other hand, there are no land constraints or permitting battles, not to mention access to almost unlimited solar energy — if it can be made to work. It’s still very much in development and theoretical at the moment. There are certainly a lot of engineering challenges that will come with operating data centers in space.

The solar panels that are going to be needed for these data centers are going to be massive. And we just don’t have the experience with operating solar panels of that size in space. At Nanyang Technological University in Singapore, where the idea of an AI chip in space is also being tested, researchers are working on a solution for exactly that. We need to be able to launch the solar panel in a small configuration so that when we reach space, we can deploy it to cover as much as possible of the area.

As part of Singapore’s Space Technology Development Programme, the university is testing a new type of solar cell that could do this. Unlike most silicon-based solar panels, this one is made from a material called perovskite, which is cheaper, lighter and flexible enough to be rolled up for launch and unfurled in orbit. It’s a chemical ink, coated onto a surface, then crystallized in a vacuum to become a solar cell. But the challenges don’t stop there.

The panels will need to track the sun at all times to maximize absorption. And with all those extra machines in space, it’s becoming far more crowded. The team at NTU is testing a low-power propulsion system that could give small satellites the ability to adjust their orbit and avoid collisions. Then comes the question of thermal management, with no air or water to carry away heat in the vacuum of space. You want to make sure that they have a way to pull heat away from the computer, and so you’ll need radiators on your spacecraft as well.

It’s kind of a complicated dance of all of these different engineering marvels that you’ll need to perfect in order for them to work. Every part of that dance depends on communication. For this to really work, satellites will need to transmit enormous amounts of data across orbit and back to Earth. Like Starcloud, Singaporean startup Transcelestial believes lasers, beamed from rooftop ground stations like this one, are the answer.

On Earth, 99% of the data that we send and receive goes through undersea cables. However, in space, we don’t have that luxury. So essentially, what we’re doing is using the same laser that we have inside a fiberoptic cable, taking it out and sending it through free space. So that can go from one satellite to another satellite or, in other words, inter-satellite links. Or we could also have them go from a satellite to the ground.

With over a thousand times more bandwidth compared to traditional radio frequency communication systems, laser links could become the connective tissue of an orbital internet. When it comes to inter-satellite links, you’re basically connecting through vacuum. There’s no air, there’s no rain, there’s no clouds, there’s no environment that can block your links. So lasers are actually a very natural fit for inter-satellite links.

Transcelestial’s long-term vision is something they call The Ring — a constellation of approximately 40 satellites that are sitting around Earth’s equator. As we grow and this technology becomes more and more accessible, I can imagine, at some point in the future, a lot of day-to-day users will be having their information sent and received via laser links to data centers in space, being processed there and getting their information back to the ground.

But perhaps the biggest obstacle of all, as it’s been since the dawn of global space exploration, is the cost of getting there. Partially reusable rockets like SpaceX’s Falcon 9 have changed the economics of reaching orbit entirely. 3, 2, 1. Liftoff of the Falcon 9. SpaceX has really led the way in terms of bringing reusable rockets to the market with their Falcon 9 rocket, but it’s only a partially reusable rocket. With Starship, they’re trying to move just a step further.

What SpaceX has been trying to achieve is a rocket that can go all the way to space, deploy satellites into orbit and then come back and land intact and need minimal refurbishing before it can launch again. That can dramatically lower the cost of launch by potentially 50 or 100 times. We see that breakeven cost to be around $500 a kilo. The Starship launch vehicle, which is coming online over the coming years, is designed to reduce launch costs of between $10 to $20 a kilo.

In fact, SpaceX has plans to launch its own data centers in space. The company has declared that they’re going to put up to a million satellites into space for this AI data center project. But they’re not the only ones that are talking about this. Blue Origin, which is helmed by Jeff Bezos, has also discussed making data centers. Then you also have other space startups like Axiom Space, Relativity Space. This is definitely something that has caught the attention of the space community.

Another p

原文超出正文长度上限，此处截断——上游还有内容，完整版见上方「原文 ↗」。