In many ways, the development of artificial intelligence infrastructure in space was a foreseeable progression. For years, Elon Musk and his inner circle have discussed space-based AI, often referencing Iain Banks’ science fiction universe of sentient, galaxy-controlling spaceships.
Now, Musk perceives an opportunity to actualize a version of this futuristic vision. His company, SpaceX, has sought regulatory approval to deploy solar-powered orbital data centers, potentially involving up to a million satellites, aiming to relocate as much as 100 GW of computational power away from Earth. It has been reported that some of these AI satellites could be constructed on the Moon.
“Within 36 months, space will undoubtedly be the most economical location for AI,” Musk stated during a recent podcast hosted by Stripe cofounder John Collison.
He is not alone in this conviction. xAI’s head of compute has reportedly wagered with their counterpart at Anthropic that 1% of the world’s computational capacity will be in orbit by 2028. Google, a significant investor in SpaceX, has unveiled Project Suncatcher, a space AI initiative slated to launch prototype vehicles in 2027. Starcloud, a startup that has secured $34 million with backing from Google and Andreessen Horowitz, recently submitted its own proposal for an 80,000-satellite constellation. Even Jeff Bezos has identified this as a future frontier.
However, beyond the fanfare, what will it truly take to deploy data centers into space?
An initial analysis suggests that, for now, Earth-based data centers remain more cost-effective than their orbital counterparts. Space engineer Andrew McCalip developed a useful calculator for comparing the two models. His baseline calculations indicate that a 1 Gw orbital data center could cost approximately $42.4 billion, nearly three times the expense of an equivalent ground-based facility, primarily due to the initial costs of satellite construction and orbital launches.
Experts contend that transforming this economic equation will necessitate technological advancements across multiple domains, substantial capital investment, and extensive development of the supply chain for space-qualified components. It also hinges on the rising costs of terrestrial resources and supply chain strains driven by escalating demand.
Satellite Design and Launch Logistics
The primary cost driver for any space-based enterprise is the expense of placing payloads into orbit. While Musk’s SpaceX is already working to reduce launch costs, analysts believe even lower prices are essential to make orbital data centers financially viable. Thus, while space AI data centers might appear to be a new venture preceding the SpaceX IPO, their feasibility depends on the completion of the company’s longest-running unifinished project: Starship.
Currently, the reusable Falcon 9 offers a launch cost of roughly $3,600 per kilogram. To make space data centers practical, Project Suncatcher’s white paper suggests prices closer to $200 per kilogram—an 18-fold improvement anticipated by the 2030s. At that price point, the energy provided by a Starlink satellite today would become competitive with a terrestrial data center.
The expectation is that SpaceX’s forthcoming Starship rocket will deliver these cost reductions; no other vehicle currently in development promises comparable savings. However, Starship has yet to become fully operational or even reach orbit, with its third iteration projected for a maiden launch in the coming months.
Even if Starship proves entirely successful, the assumption that it will immediately translate into significantly lower prices for customers might not hold true. Economists at Rational Futures present a compelling argument that, similar to the Falcon 9, SpaceX would likely avoid charging substantially less than its closest competitor to avoid leaving potential revenue on the table. For instance, if Blue Origin’s New Glenn rocket is priced at $70 million, SpaceX would likely not undertake Starship missions for external clients at a much lower rate, which would still place it above the cost projections made by space data center developers.
“We currently lack enough rockets to launch a million satellites, so we’re quite a distance from that goal,” commented Matt Gorman, the CEO of Amazon Web Services, at a recent event. “Considering the current cost of sending a payload into space, it’s exorbitant. It’s simply not economically feasible.”
Nevertheless, if launch costs are a persistent challenge for space ventures, the second major hurdle lies in production expenses.
“We tend to assume Starship’s cost will be just a few hundred dollars per kilo,” McCalip informed TechCrunch. “However, people are overlooking that satellites currently cost almost $1,000 a kilo.”
Satellite manufacturing expenses represent the largest portion of this cost, but if powerful satellites could be produced at roughly half the cost of existing Starlink units, the economics begin to align. SpaceX has achieved significant progress in satellite economics through the development of its record-breaking Starlink communications network and anticipates further savings through mass production. The rationale behind a million satellites undoubtedly includes the cost efficiencies gained from large-scale manufacturing.
Still, the satellites designed for these missions must be sufficiently large to accommodate the complex requirements for operating powerful GPUs, including substantial solar arrays, efficient thermal management systems, and laser-based communication links for data transmission and reception.
A 2025 white paper from Project Suncatcher offers a method to compare terrestrial and space data centers based on the cost of power, the fundamental input for chip operation. On Earth, data centers typically spend around $570–3,000 per Kw of power annually, depending on local energy rates and system efficiency. SpaceX’s Starlink satellites generate power from on-board solar panels, but the cumulative cost of acquiring, launching, and maintaining these spacecraft translates to an energy cost of $14,700 per Kw annually. In essence, satellites and their components need to become considerably more affordable before they can compete economically with grid-metered power.
The Rigors of the Space Environment
Proponents of orbital data centers often claim that thermal management in space is “free,” but this is an oversimplification. The absence of an atmosphere actually makes heat dissipation more challenging.
“You have to rely on very large radiators to simply dissipate that heat into the cold vacuum of space, which requires a significant amount of surface area and mass to manage,” explained Mike Safyan, an executive at Planet Labs, which is developing prototype satellites for Google Suncatcher slated for launch in 2027. “This is acknowledged as one of the primary challenges, especially for long-term operations.”
Beyond the vacuum, AI satellites will also need to contend with cosmic radiation. Cosmic rays gradually degrade chips and can induce “bit flip” errors, which can corrupt data. Protection can be achieved through shielding, the use of radiation-hardened components, or redundant error checking in series, but all these options involve costly trade-offs in terms of mass. Nevertheless, Google has utilized a particle beam to assess the effects of radiation on its Tensor Processing Units (chips specifically designed for machine learning). SpaceX executives have indicated on social media that the company has acquired a particle accelerator for this very purpose.
Another hurdle arises from the solar panels themselves. The project’s fundamental premise is energy arbitrage: placing solar panels in space makes them five to eight times more efficient than on Earth, and if positioned in the correct orbit, they can maintain continuous exposure to the sun for 90% or more of the day, further boosting efficiency. Electricity is the main power source for chips, so more energy equates to more affordable data centers. However, even solar panels present added complexities in space.
Space-rated solar panels made from rare earth elements are durable but prohibitively expensive. Silicon-based solar panels are inexpensive and increasingly common in space, used by Starlink and Amazon Kuiper, but they degrade much more rapidly due to space radiation. This degradation will limit the operational lifespan of AI satellites to approximately five years, necessitating a faster return on investment.
Still, some analysts believe this is not a major impediment, given the rapid pace of new chip generations. “After five or six years, the dollars per kilowatt-hour don’t yield a return because the technology is no longer state-of-the-art,” Philip Johnston, CEO of Starcloud, told TechCrunch.
Danny Field, an executive at Solestial, a startup developing space-rated silicon solar panels, notes that the industry views orbital data centers as a crucial growth driver. He is currently in discussions with several companies regarding potential data center projects and states that “any sufficiently ambitious player is at least contemplating it.” As an experienced spacecraft design engineer, however, he acknowledges the significant challenges inherent in these models.
“You can always extrapolate physics to a larger scale,” Field remarked. “I’m eager to see how some of these companies reach a point where the economics are sound and the business case is closed.”
The Role of Space Data Centers
A key unanswered question about these data centers is their intended application. Will they be general-purpose, used for inference, or for training? Based on current use cases, they may not be entirely interchangeable with ground-based data centers.
A major hurdle for training new models is coordinating thousands of GPUs concurrently. Most model training is not distributed but performed within individual data centers. While hyperscalers are striving to change this to enhance model power, it remains an unachieved goal. Similarly, training in space will demand coherence among GPUs distributed across multiple satellites.
The team behind Google’s Project Suncatcher highlights that the company’s terrestrial data centers connect their TPU networks with throughput in the hundreds of gigabits per second. The fastest commercially available inter-satellite communication links, which rely on lasers, can currently only reach about 100 Gbps.
This disparity has led to an innovative architecture for Suncatcher: it involves flying 81 satellites in a tightly controlled formation, allowing them to be close enough to utilize the transceivers typically found in terrestrial data centers. This, of course, introduces its own set of challenges, particularly the autonomy required to ensure each spacecraft maintains its precise position, even when maneuvers are necessary to avoid orbital debris or other spacecraft.
Nonetheless, the Google study includes a caution: inference tasks can withstand the orbital radiation environment, but further research is needed to fully comprehend the potential impact of bit-flips and other errors on training workloads.
Inference tasks do not require thousands of GPUs working in perfect synchronicity. The operation can be performed with dozens of GPUs, possibly on a single satellite, an architecture that represents a minimum viable product and the probable starting point for the orbital data center industry.
“Training is not the ideal activity for space,” Johnston asserted. “I believe almost all inference workloads will eventually be performed in space,” envisioning everything from customer service voice agents to ChatGPT queries being processed in orbit. He claims his company’s first AI satellite is already generating revenue by performing inference in orbit.
Although specific details remain sparse, even in the company’s FCC filing, SpaceX’s orbital data center constellation is projected to deliver approximately 100 kW of computational power per ton, roughly double the power of current Starlink satellites. These spacecraft will operate in conjunction with each other and leverage the Starlink network for information sharing; the filing asserts that Starlink’s laser links can achieve petabit-level throughput.
For SpaceX, the recent acquisition of xAI, which is developing its own terrestrial data centers, will enable the company to establish positions in both ground-based and orbital data centers, observing which supply chain adapts more quickly.
This strategy benefits from having fungible Floating Point Operations Per Second (FLOPS)—if it can be made to work. “A FLOP is a FLOP, regardless of where it resides,” McCalip stated. “[SpaceX] can simply scale until they encounter permitting or capital expenditure bottlenecks on the ground, and then pivot to their space deployments.”
Got a sensitive tip or confidential documents about SpaceX? Reach out to Tim Fernholz at tim.fernholz@techcrunch.com. For secure communication, you can contact him via Signal at tim_fernholz.21.