The future of AI is out of this world, literally! Elon Musk and his team at SpaceX are taking AI to new heights, or should we say, orbits. But is this ambitious plan feasible, or just a far-fetched sci-fi dream?
Musk envisions a network of solar-powered orbital data centers, a million satellites strong, shifting an incredible 100 GW of compute power off our planet. It's a vision inspired by the works of Iain Banks, where sentient spaceships rule the galaxy.
However, behind the hype and excitement, there are some serious challenges to overcome. The economics of this endeavor are brutal, to say the least. Today's terrestrial data centers are still more cost-effective than their orbital counterparts. According to space engineer Andrew McCalip, a 1 GW orbital data center could cost a whopping $42.4 billion, almost three times more than a ground-based one, due to the initial costs of satellite construction and launch.
To make this vision a reality, significant technological advancements, massive investments, and an efficient supply chain for space-grade components are required. It's a tall order, but Musk and his competitors aren't backing down.
The key to success lies in reducing the cost of getting to orbit. Musk's SpaceX is already driving down these costs, but more needs to be done. The reusable Falcon 9 rocket currently offers an orbit cost of around $3,600 per kg, but to make space data centers economically viable, prices need to drop to $200 per kg. This is where SpaceX's next-gen Starship rocket comes into play, promising significant savings, although it's yet to prove its operational capabilities.
But even if Starship delivers, will SpaceX offer competitive prices to external customers? Economists argue that SpaceX might not undercut its best competitor, leaving the ball in Blue Origin's court with their New Glenn rocket.
Matt Gorman, CEO of Amazon Web Services, sums it up: "If you think about the cost of getting a payload into space today, it's massive. It's just not economical."
The second challenge is production cost. Satellite manufacturing costs are a significant hurdle, but if high-powered satellites can be produced at a fraction of the current cost, the numbers might start to make sense. SpaceX, with its experience building Starlink, hopes to achieve economies of scale.
However, these satellites must be large enough to accommodate powerful GPUs, requiring large solar arrays, thermal management systems, and laser-based communication links.
The space environment is unforgiving. Thermal management, often touted as "free" in space, is actually more challenging due to the lack of an atmosphere. Cosmic radiation is another hurdle, degrading chips and causing data corruption. While solutions exist, they come at a cost, adding to the complexity.
Solar panels, the energy source for these satellites, offer increased efficiency in space, but they're more complex and degrade faster due to space radiation. This limits the lifetime of AI satellites to around five years, putting pressure on their return on investment.
Despite these challenges, some analysts remain optimistic. Philip Johnston, CEO of Starcloud, believes the rapid evolution of chip technology will make this a non-issue.
Danny Field, an executive at Solestial, a startup building space-rated silicon solar panels, sees orbital data centers as a key growth driver. He's working with several companies on potential data center projects, highlighting the industry's interest.
But what will these space data centers be used for? Will they be general-purpose, or for inference, or training new models? The key challenge for training is operating thousands of GPUs together, something that hasn't been achieved yet. Training in space will require coherence between GPUs on multiple satellites, a complex task.
Google's Project Suncatcher proposes an intriguing architecture, flying 81 satellites in formation to achieve the necessary connectivity. However, this presents its own set of challenges, requiring advanced autonomy to maintain each spacecraft's position.
Inference tasks, on the other hand, don't require the same level of GPU coordination. Dozens of GPUs on a single satellite might suffice, representing a minimum viable product for the orbital data center business.
Johnston believes inference workloads will dominate in space, envisioning a future where customer service agents and ChatGPT queries are computed in orbit. He claims his company's AI satellite is already generating revenue performing inference tasks.
SpaceX's orbital data center constellation aims for around 100 kW of compute power per ton, roughly twice that of current Starlink satellites. The spacecraft will operate in connection with each other and the Starlink network, leveraging its laser links for high-throughput data sharing.
With its recent acquisition of xAI, SpaceX is positioning itself in both terrestrial and orbital data centers, waiting to see which supply chain adapts faster.
As McCalip puts it, "A FLOP is a FLOP, it doesn't matter where it lives." SpaceX can scale its operations until it hits bottlenecks on the ground, then fall back to its space deployments.
The future of AI in space is an exciting prospect, but it's a long road ahead. Will Musk and his competitors succeed in realizing this vision? Only time will tell, but one thing's for sure: the journey will be fascinating to watch.
