Why are the economics of orbital AI so brutal?

In a sense, this whole thing was inevitable. Elon Musk and his cohorts have been talking about artificial intelligence in space for years, especially in the context of Ian Banks’ sci-fi series about a far-future world where sentient spaceships roam and rule the galaxy.

Now, Musk sees an opportunity to realize a version of this vision. His company SpaceX has requested regulatory clearance to build solar-powered orbital data centers, distributed across up to 1 million satellites, that could transmit up to 100 gigawatts of computing power off-planet. He has reportedly suggested that some of his AI satellites will be built on the moon.

“The cheapest place to put AI will be in space in 36 months or less,” Musk said last week on a podcast hosted by Stripe co-founder John Collison.

He’s not alone. The head of computing at xAI reportedly bet his counterpart at Anthropic that 1% of global computing will be in orbit by 2028. Google (which has a significant ownership stake in SpaceX) announced a space AI effort called Project Suncatcher, which will launch prototype vehicles in 2027. Starcloud, a startup that has raised $34 million with backing from Google and Andreessen Horowitz, has presented its own plans for a satellite An artificial satellite that includes 80,000 satellites. Constellation last week. Even Jeff Bezos said this is the future.

But behind the hype, what does it actually take to send data centers into space?

In the first analysis, today’s ground-based data centers remain cheaper than those in orbit. Aerospace engineer Andrew McCaleb has built a handy calculator to compare the two models. His baseline findings show that a 1-gigawatt orbital data center could cost $42.4 billion, nearly three times as much as its terrestrial counterpart, thanks to the upfront costs of building the satellites and launching them into orbit.

Experts say changing this equation will require technology development in many areas, huge capital spending, and a lot of work on the supply chain for space components. It also depends on rising costs on the ground as resources and supply chains are strained by increasing demand.

Design and launch of satellites

The main driver of any space business model is how much it costs to get to anything there. Musk’s SpaceX is already working to lower the cost of getting to orbit, but analysts looking at what it would take to make orbital data centers a reality need lower prices to close their business case. In other words, while AI data centers may seem like a story about a new line of business before SpaceX’s IPO, the plan hinges on completing the company’s longest unfinished project — Starship.

Consider that a reusable Falcon 9 rocket today costs about $3,600 per kilogram. Making space data centers feasible, according to the Suncatcher Project white paper, would require prices approaching $200 per kilogram, an 18-fold improvement that is expected to be available in the 2030s. However, at this price, the power provided by today’s Starlink satellite would be cost-competitive with a terrestrial data center.

SpaceX’s next-generation Starship rocket is expected to achieve these improvements, and no other vehicle in development promises similar savings. However, that vehicle has not yet become operational or even reached orbit; The third iteration of Starship is expected to debut sometime in the coming months.

However, even if Starship is completely successful, assumptions that it will immediately provide lower prices to customers may not pass the smell test. Economists at consulting firm Rational Futures make a compelling case that SpaceX, as with the Falcon 9, would not want to charge significantly less than its best competitor, or else the company would be leaving money on the table. For example, if Blue Origin’s New Glenn rocket sold for $70 million, SpaceX wouldn’t fly Starship missions to outside customers at a much lower price, which would leave it higher than the numbers assumed by space data center builders.

“There aren’t enough rockets to launch a million satellites yet, so we’re a long way from that,” Amazon Web Services CEO Matt Gorman said at a recent event. “If you think about the cost of sending a payload into space today, it is prohibitive. It is not economical.”

However, if launch is the bane of all space companies, the second challenge is the cost of production.

“We always consider, at this point, that the cost of Starship is going to be hundreds of dollars per kilo,” McCaleb told TechCrunch. “People don’t take into account that the price of satellite is about $1,000 a kilo right now.”

Satellite manufacturing costs account for the bulk of that price, but if high-powered satellites can be manufactured for about half the cost of current Starlink satellites, the numbers start to make sense. SpaceX has made significant progress in satellite economics while building Starlink, its record-setting communications network, and the company hopes to achieve more with broadband. Part of the reason for producing a million satellites is undoubtedly the cost savings resulting from mass production.

However, the satellites that will be used for these missions must be large enough to meet the complex requirements of operating powerful GPUs, including large solar arrays, thermal management systems, and laser-based communications links to receive and deliver data.

A 2025 report from Project Suncatcher offers one way to compare terrestrial and space-based data centers with the cost of energy, the basic input needed to run chips. In reality, data centers spend approximately $570 to $3,000 per kilowatt of power over the course of a year, depending on local energy costs and the efficiency of their systems. SpaceX’s Starlink satellites get their power from onboard solar panels instead, but the cost of acquiring, launching and maintaining these spacecraft saves energy of $14,700 per kilowatt over the course of a year. Simply put, satellites and their components must become much cheaper before they can become cost competitive with metered power.

The space environment does not deceive

Proponents of orbital data centers often say that thermal management is “free” in space, but this is an oversimplification. Without an atmosphere, heat dissipation is more difficult.

“We rely on very large radiators to be able to dissipate that heat in the darkness of space, and so you have to manage a lot of surface area and mass,” said Mike Sufian, an executive at Planet Labs, which is building prototypes of Google Suncatcher satellites that are expected to launch in 2027. “That’s admittedly one of the major challenges, especially in the long term.”

Besides the vacuum of space, AI satellites will need to deal with cosmic radiation as well. Cosmic rays decay chips over time, and can also cause “bit flip” errors that can corrupt data. Chips can be shielded, use ray-hardened components, or operate serially with redundant error checks, but all of these options involve expensive block trades. However, Google used a particle beam to test the effects of radiation on Tensor processing units (chips specifically designed for machine learning applications). SpaceX executives said on social media that the company acquired the particle accelerator for just this purpose.

Another challenge comes from the solar panels themselves. The logic of the project is energy balancing: placing solar panels in space makes them five to eight times more efficient than on Earth, and if they are in the right orbit, they can be within sight of the sun for 90% of the day or more, increasing their efficiency. Electricity is the main fuel for chips, so more power = cheaper data centers. But even solar panels are more complicated in space.

Space-rated solar panels made from rare earth elements are powerful but expensive. Silicon solar panels are cheap and increasingly common in space, used by Starlink and Amazon Kuiper, but they degrade much faster due to space radiation. This would limit the lifespan of AI satellites to about five years, meaning they would have to achieve a return on investment faster.

However, some analysts believe this is not a big deal, based on how quickly new generations of chips are arriving on the scene. “After five or six years, the dollar-per-kilowatt-hour price doesn’t bring any return, and that’s because it’s not state-of-the-art,” Starcloud CEO Philip Johnston told TechCrunch.

The industry sees orbital data centers as a key driver of growth, says Danny Field, an executive at Solstial, a startup that builds space-grade silicon solar panels. He is talking to several companies about potential data center projects. “Any player big enough to dream about it is at least considering it,” he says. However, as a long-time spacecraft design engineer, he doesn’t underestimate the challenges these models face.

“You can always extrapolate physics to a larger scale,” Field said. “I’m excited to see how some of these companies get to the point where the economics make sense and close the business case.”

How do space data centers fit in?

One of the outstanding questions about these data centers: What will we do with them? Is it for general purposes, inference, or training? And based on current use cases, they may not be completely interchangeable with on-ground data centers.

The main challenge for training new models is running thousands of GPUs together Collectively. Most model training is not distributed, but performed in individual data centers. Superscale is working to change this in order to increase the power of their models, but this has not been achieved yet. Likewise, training in space will require interconnection between GPUs on multiple satellites.

The team at the Google Suncatcher project notes that the company’s ground-based data centers connect its TPU networks with a throughput of hundreds of gigabytes per second. The fastest inter-satellite communications links available today, which use lasers, can only get up to about 100 gigabits per second.

This led to an interesting design for Suncatcher: It involved launching 81 satellites in formation, close enough to use the kind of transponders that ground-based data centers rely on. This of course poses its own challenges: the autonomy required to ensure that each spacecraft remains in its correct station, even if maneuvers are required to avoid orbital debris or another spacecraft.

However, the Google study offers a caveat: inference work can withstand an orbital radiation environment, but more research is needed to understand the potential impact of bit flips and other errors on training workloads.

Inference tasks don’t have the same need for thousands of GPUs working in unison. This task could be accomplished using dozens of GPUs, perhaps on a single satellite, an architecture that represents a kind of minimum viable product and a potential starting point for an orbital data center business.

“Training is not the ideal thing to do in space,” Johnston said. “I think almost all inference workloads will be performed in space,” he said, imagining everything from customer service voice agents to ChatGPT queries being computed in orbit. He says his company’s first AI satellite is already generating revenue through inference in orbit.

While details are scarce even in the company’s FCC filing, SpaceX’s array of orbital data centers appears to expect about 100 kilowatts of computing power per ton, roughly twice the power of current Starlink satellites. The spacecraft will communicate with each other and use the Starlink network to exchange information. The filing claims that Starlink’s laser links can achieve petabit-level throughput.

As for SpaceX, the company’s recent acquisition of xAI (which is building its own terrestrial data centers) will allow the company to gain locations in both terrestrial and orbital data centers, to see which supply chain adapts faster.

That’s the benefit of having replaceable floating point operations per second – if you can make it work. “Flipping is flipping, it doesn’t matter where he lives,” McCaleb said. “[SpaceX] It can only scale up [it] It hits permissiveness or capitalist bottlenecks to the ground, then retreats to [their] Space deployments.”

Do you have sensitive information or confidential documents about SpaceX? Contact Tim Fernholz at tim.fernholz@techcrunchCOh. For secure communication, you can contact him via Signal at tim_fernholz.21.