Elon Musk predicts space will be cheapest place for AI data centers within three years

What's the deal with space-based data centers for AI?

Starcloud has already begun running and training a large language model in space, so it can speak Shakespearean English Terrestrial data centers are so 2025. We're taking our large-scale compute infrastructure into orbit, baby! Or at least, that's what Big Tech is yelling from the rooftops at the moment. It's quite a bonkers idea that's hoovering up money and mindspace, so let's unpack what it's all about - and whether it's even grounded in reality. Let's start with the basics. You might already know that a data center is essentially a large warehouse filled with thousands of servers that run 24/7. AI companies like Anthropic, OpenAI, and Google use data centers in two main ways: AI companies need data centers because they provide the coordinated power of thousands of machines working in tandem on these functions, plus the infrastructure to keep them running reliably around the clock. To that end, these facilities are always online with ultra-fast internet connections, and they have vast cooling systems to keep those servers running at peak performance levels. All this requires a lot of power, which puts a strain on the grid and squeezes local resources. So what's this noise about data centers in space? The idea's been bandied about for a while now as a vastly better alternative that can harness infinitely abundant solar energy and radiative cooling hundreds of miles above the ground in low Earth orbit. Powerful GPU-equipped servers would be contained in satellites, and they'd move through space together in constellations, beaming data back and forth as they travel around the Earth from pole to pole in the sun-synchronous orbit. The thinking behind space data centers is that it'll allow operators to scale up compute resources far more easily than on Earth. Up there, there aren't any constraints of easily available power, real estate, and fresh water supplies needed for cooling. There are a number of firms getting in on the action, including big familiar names and plucky upstarts. You've got Google partnering with Earth monitoring company Planet on Project Suncatcher to launch a couple of prototype satellites by next year. Aetherflux, a startup that was initially all about beaming down solar power from space, now intends to make a data center node in orbit available for commercial use early next year. Nvidia-backed Starcloud, which is focused exclusively on space-based data centers, sent a GPU payload into space last November, and trained and ran a large language model on it. The latest to join the fold is SpaceX, which is set to merge with Elon Musk's AI company xAI in a purported US$1.25-trillion deal with a view to usher in the era of orbital data centers. According to Musk's calculations, it should be possible to increase the number of rocket launches and the data center satellites they can carry. "There is a path to launching 1 TW/year (1 terawatt of compute power per year) from Earth," he noted in a memo, adding that AI compute resources will be cheaper to generate in space than on the ground within three years from now. In an excellent article in The Verge from last December, Elissa Welle laid out the numerous challenges these orbital data centers will have to overcome in order to operate as advertised. For starters, they'd have to safely wade through the 6,600 tons of space debris floating around in orbit, as well as the 14,000-plus active satellites in orbit. Dodging these will require fuel. You've also got to dissipate heat from the space-based data centers, and have astronauts maintain them periodically. And that's to say nothing about how these satellites will affect the work of astronomers or potentially increase light pollution. Ultimately, there's a lot of experimentation and learning to be gleaned from these early efforts to build out compute resources in space before any company or national agency can realistically scale them up. And while it might eventually become possible to do so despite substantial difficulties, it's worth asking ourselves whether AI is actually on track to benefit humanity in all the ways we've been promised, and whether we need to continually build out infrastructure for it - whether on the ground or way up beyond the atmosphere.

Musk predicts more AI capacity will be in orbit than on earth in 5 years, with SpaceX becoming a 'hyper-hyper' scaler | Fortune

By the beginning of the next decade, AI will primarily become a space-based venture as the cost becomes much more advantageous to operate in orbit, according to SpaceX CEO Elon Musk. In a lengthy, wide-ranging interview with podcaster Dwarkesh Patel and Stripe cofounder and president John Collison on Thursday, the tech billionaire made some of his signature bold predictions about how the AI revolution will play out. Given the enormous energy needs of AI and limits on available land for placing massive arrays of solar panels -- not to mention all the red tape -- building new AI data centers will be much cheaper in orbit, where solar panels are five times more effective than on the ground. "In 36 months, but probably closer to 30 months, the most economically compelling place to put AI will be space," Musk said. "It will then get ridiculously better to be in space. The only place you can really scale is space. Once you start thinking in terms of what percentage of the sun's power you are harnessing, you realize you have to go to space. You can't scale very much on earth." The utility industry isn't able currently to build power plants as rapidly as needed for AI, he added. On top of that, limits on manufacturing gas turbines and wind turbines fast enough represent another bottleneck. Meanwhile, solar panels meant to be used in space are less costly than those designed for use on land because they don't need as much glass or hardening to withstand various weather events, Musk explained. In addition, the cooling needed for data centers is less of an issue in space. Considering the advantage space has over earth, he was asked where AI will be in five years. "If you say five years from now, I think probably AI in space will be launching every year the sum total of all AI on earth," Musk said. "Meaning, five years from now, my prediction is we will launch and be operating every year more AI in space than the cumulative total on earth." While he is infamous for setting incredibly ambitious targets on aggressive timelines, his next one was a whopper, even by his standards. Musk said getting all that AI and solar capacity in space will require about 10,000 launches a year -- or a launch in less than an hour every day. SpaceX is the most prolific rocket company and set a record last year with 165 orbital launches. SpaceX could pull off a 10,000-per-year launch cadence with 20-30 Starship rockets, he added, though the company will make more than that, enabling perhaps 20,000-30,000 launches a year. He pointed out the airline industry has much quicker throughput than that. The number of daily flights around the world tops 100,000. Patel then asked if SpaceX will become an AI hyperscaler. "Hyper-hyper," Musk replied. "If some of my predictions come true, SpaceX will launch more AI than the cumulative amount on Earth of everything else combined." It's already working toward that goal. SpaceX in November launched a test satellite with an AI server from start-up Starcloud. And last month, SpaceX asked the FCC for permission to launch up to 1 million solar‑powered satellites designed as data centers. Of course, there are other challenges associated with operating in space, such as protecting hardware from the sun's radiation and transmitting astronomical amounts data from orbit to earth. SpaceX's Starship rocket is also still in development. But Deutsche Bank said in a note last month the challenges of putting data centers in space are more about engineering than physics. Today's AI hyperscalers see the potential as well and are also looking to go to space. For example, Google's Project Suncatcher looks to pair solar-powered satellites with AI computer chips, and a prototype could launch as soon as next year. OpenAI CEO Sam Altman also considered buying rocket company Stoke Space to put data centers in orbit, the Wall Street Journal reported in December. For its part, SpaceX is an AI company now, after its merger with Musk's xAI. That's as SpaceX is expected to go public this year, raising tens of billions of dollars. During the podcast interview, Musk said more money is available in public markets than private markets, possibly even 100 times more. "I just repeatedly tackle the limiting factor," he added. "Whatever the limiting factor is on speed, I'm going to tackle that. If capital is the limiting factor, then I'll solve for capital. If it's not the limiting factor, I'll solve for something else."

Elon Musk says space will be the cheapest place for AI data centres in three years

Talking on the Dwarkesh Podcast, co-hosted by Dwarkesh Patel and Stripe CEO John Collison, Musk said Earth has some insurmountable power bottlenecks in scaling AI. "The availability of energy is the issue," Musk said, noting global electricity output outside China remains "pretty close flat" while chip production surges. Billionaire Elon Musk has predicted that within 36 months, or likely sooner, space will be the most economical location for artificial intelligence (AI) data centres, eclipsing Earth-based options due to vastly superior solar efficiency and the absence of a need for batteries. Talking on the Dwarkesh Podcast, co-hosted by Dwarkesh Patel and Stripe CEO John Collison, Musk said Earth has some insurmountable power bottlenecks in scaling AI. "The availability of energy is the issue," Musk said, noting global electricity output outside China remains "pretty close flat" while chip production surges. Musk quipped that he meant to wear a t-shirt that said, "It's always sunny in space". "Because you don't have a day-night cycle, seasonality, clouds, or an atmosphere in space. The atmosphere alone results in about a 30% loss of energy [on Earth]," said Musk, adding, "It will simply not be physically possible to scale power production to the scale needed for AI on Earth." "Any given solar panel can do about five times more power in space than on the ground. You also avoid the cost of having batteries to carry you through the night. It's actually much cheaper to do in space," he said. The billionaire, who is very close to a trillionaire status, said this days after his aerospace giant SpaceX acquired his AI company xAI, to create a $1.25 trillion behemoth. The company said it is acquiring xAI to "form the most ambitious, vertically-integrated innovation engine on (and off) Earth, with AI, rockets, space-based internet, direct-to-mobile device communications and the world's foremost real-time information and free speech platform." On the podcast, Musk explained the multitude of issues facing the AI scaling problem. For space to dominate economically, three conditions must come together, according to Musk: * Earth's power hits a hard ceiling as AI demand explodes * Chip fabs like Musk's planned "TeraFab" outpace energy scaling * Starship achieve thousands of launches yearly. If this happens, Musk said his ecosystem wins. SpaceX alone can launch at that cadence, powering xAI with unlimited gigawatts annually while rivals scrap over turbines and grids. "The only place you can really scale is space," he stressed, eyeing hundreds of gigawatts launched yearly via Starship. Five years out, space AI could surpass all terrestrial capacity combined, lapping US power (500 GW average) repeatedly. When asked about chances of technical issues of data centres in space, Musk said, "At this point, we find our GPUs to be quite reliable. There's infant mortality, which you can obviously iron out on the ground. "Once they start working and you're past the initial debug cycle, they're quite reliable."