Let's build AI data centers in space | Philip Johnston

This time next week, I will be at the Cape Canaveral Space Force Base, nervously awaiting the launch of our first spacecraft. Much about that setting will be different from this one, but there are a few things that will remain the same. For example, like now, in a week's time, I will also be hoping not to crash and burn. (Laughter) And like now, in a week's time, I will again be wearing my same lucky underpants. That's not a joke. (Laughter) So what about the spacecraft next week makes it so special that it warrants a rare second weekly appearance of said undergarment? (Laughs) It's because it will be the first time that anybody has tried to launch an AI data center to space. More specifically, we will be the first to launch the Nvidia H100 chip to space, which is about 100 times more powerful GPU, or AI compute, than has ever been in space before. In fact, this spacecraft is the first step in a much larger vision to build almost all data centers in space. And today I will make the case for this idea. The idea that the abundant energy and cold temperatures in space will soon mean that it makes much more sense to build data centers in space than it does to build them on Earth. But first, why do we need more data centers? Well, there's both a carrot and a stick. As many of the speakers have mentioned, the carrot is the promise of AI. So that's things like new treatments for cancer, self-driving cars, a superintelligent companion in everybody's pocket. In order to realize the promise of AI, we will need many more new data centers and many more new energy projects to power them. So that's the carrot. But there is also a stick, which people don't like to talk about as frequently. And that is the very real risk in being left behind by competitor nations. This comes with very, very real national security implications for the US. So we know we need more data centers, but why not just carry on building them on Earth as we have been? Well, some forecasts suggest that over the next just three years alone, we will need an additional 50 to 100 gigawatts or 50 to 100 new nuclear power stations, in just the US alone to meet the coming demand for AI. Permitting and other constraints mean that this is not possible. And this is basically leading to a dramatic rise in energy prices. And we also are sucking up rivers and reservoirs to keep these data centers cool. A few weeks ago, I was invited down to Tucson, Arizona, to meet with community leaders there. Tucson recently became the first city in the US to unanimously vote to reject a new data center proposal, to a gigawatt scale-data center proposal, in their community, primarily due to concerns about energy and water usage. I spoke with many members of the community. They had been shocked to find out that this new data center proposal would have put a strain on energy and water in the community that would have been a detriment for generations to come. And it's not just Tucson, this is happening in towns and cities across the US and across the world. Now, if you've ever seen a data center, you will know that they are massive. So it can be pretty difficult to imagine what this kind of data center could look like in space. And so I would like to show you a short 15-second clip of what a huge five-gigawatt data center in space might look like. So here you see a Starship-sized spacecraft, and I'll talk about this later, with a 100-ton module of chips connecting to a five-gigawatt cluster with a four-kilometer solar panel with a one-kilometer radiator. And don't worry, this won't block out the sun or anything. (Laughter) So I do know this looks a little bit wacky, by the way. When we first released this video a year and a half ago, as part of our Y-Combinator launch, we got roundly pilloried from all corners. I thought I would share with you a few of the more polite X comments that we got a year and a half ago. "This gotta be the dumbest shit ..." (Laughter) "This gotta be the dumbest shit I ever seen or heard." Fair enough, fair enough. Each has their own opinion. "I thought you nerds were supposed to be good at science." (Laughter) And just to be clear, I literally first got this message when I was inside a particle accelerator testing chips. "Good luck trying to run an h100 in space." To be honest, this is actually very generous. I really appreciate the good wishes. (Laughter) And my personal favorite, "This will never get off the ground, in a very literal sense." (Laughter) It's true that revolutionary ideas often sound stupid at first, but sadly it also happens to be true that most terrible ideas also often sound stupid at first. So how can you be certain of which camp you're in and be sure that you're not in the second camp? Well, it's important to reason from first principles, but I mean, to be very honest, in order to know which camp you're in, most oftentimes it is going to involve taking some risk. If you don't take some risk, you're likely not doing anything consequential. So for us, to know definitively if we can run an h100 in space, the best way to do that is to build a spacecraft and launch it to space. And that's exactly what we're doing with our first spacecraft launching next week. So given how controversial and audacious this idea seems to be, I thought I would share a little bit of context about how this idea came about, and then share some of the pros and some of the cons. A few years ago, I quite randomly on a weekend, decided to take a trip down to Starbase, Texas, where SpaceX is building their new Starship rocket. This wasn't a work trip or anything, I'm just a huge space nerd and this is my idea of a fun weekend. The first stop on the trip is to what they call the Rocket Garden, and it was the first time I had seen the massive new Starship launch vehicle that SpaceX is building. This was impressive, but to be honest, what really impressed me — besides the Texas barbecue, which is amazing — what really impressed me was that across the road from where I'm standing, behind the camera in this image, they are building, or they were building then, by far the largest factory I have ever seen. They're building these two Starship gigafactories that are essentially pretty similar to Tesla production lines, in that they're designed such that within two years they will be rolling one Starship per day off that production line. This is interesting because while the launch cost with reusability might come down by 50 to 100 times, the launch capacity, i.e. how many tons per year you can get to space, might go up by 1,000 times or more. And how is that possible? It's because Starship is the first-ever fully reusable rocket. So, for example, with the current launch vehicle, Falcon 9, if you build a new one every day for a year, at the end of the year, you still only have one Falcon 9 upper stage because it's dispensable. Whereas with Starship, if you build a new one every day for a year, at the end of the year you have 365 starships because they're reusable. And it's not just SpaceX. There's new rockets coming from Blue Origin, Stoke Space, Relativity Space and Rocket Lab. Even if you have a pretty healthy dose of skepticism on their forecasts, the coming capacity is just mind-blowing and will be the real game changer. So this coming capacity reminded me, or got me thinking about the concepts from sci-fi that I remember reading about as a kid or as a teenager under the torch light at night. In the 1940s, Isaac Asimov first wrote about the idea of space-based solar, which is where you have a huge solar panel in space, and then you somehow beam that power down. The problem with space-based solar, though, has always been that you lose most of the energy in transmission from space to Earth. But with low-cost launch, we now have the ability to move things like data centers to space close to the energy source, to consume the energy there and just beam down the results so we don't lose the energy in transmission. I ran the numbers with my cofounders, Ezra Feilden and Adi Oltean, and we came to the conclusion that if the launch cost gets to around 500 dollars a kilo, which is well within Starship launch prices, then the data centers in space will be economically viable. So that's how the idea came about. But where does this 500 dollars a kilo number come from? And how is it that energy for data centers in space will be cheaper than on Earth? Well, let's run a comparison with a data center on Earth, with a solar project on Earth to power a data center, because solar on Earth is the cheapest form of energy we have right now. So you have three big costs with a solar project on Earth. Number one is the cost of permitted land. This is actually the biggest cost or can be the biggest cost, particularly in North America, where it's linked to the energy price. Number two, you have the cost of battery storage because you need to charge the batteries during the day so that you have power at night. And number three, you have the cost of the solar cells themselves. So how does that compare with a solar project in space? Well, number one, on the plus side, we don't need permitted land. We don't need battery storage because we're 24/7 in the sun, and we need six times less solar cells. Since one square meter of solar panel in space produces six times the energy of one square meter of solar panel on Earth. But there's a few additional costs in space. So I mentioned we're going to need six times less solar. But there's one big cost, which is we need the launch cost. The launches to get the chips, solar panels and radiators to space. But now you can see very clearly that there is clearly a breakeven point where the launch cost is below the cost of permitted land, batteries and six times the solar. And we see that launch cost to be around 500 dollars a kilo, as I mentioned, which is well within range of the Starship launch price. I'll finish with this. Earlier, I mentioned the competition between nations for AI, and over the last few years, we've seen rising tensions with the first hot war in Europe in generations. At its root, the biggest driver of large-scale war is competition for resources. And over the coming century, the biggest competition between nations will be for energy and water for data centers. Some of you might know that in 1945, the United Nations Charter was first signed in this very room, on this very stage where I'm standing right now. Look at the curtains in this image and the chandeliers, they are the same curtains that you will see. It happened in this room. We might take it for granted right now, but the United Nations has helped keep us safe since the second World War until now. And as the grandchild of two British World War II veterans, the words of the charter, the promise of their generation to ours, still move me every time I think of them: "We, the people of the United Nations, determined to save succeeding generations from the scourge of war, which twice in our lifetime has brought untold sorrow to mankind." The most effective way that we can save our own children and grandchildren from the scourge of war will be to stop competing over the fundamentally finite resources of Earth, and to start utilizing the near limitless energy of our solar system and eventually of our galaxy. In a week's time, I'll be watching the first AI data center launch to space. In ten years' time, most new data centers will be being built in space for the energy. And who knows, maybe in 50 years' time, if I keep wearing my lucky underpants -- (Laughter) We may have started work on a Dyson sphere to help harness the full power of our sun. Thank you. (Applause)