As rural areas across the U.S. increasingly push back against the establishment of AI computing data centers nearby, some major tech firms are exploring creative alternatives, including the possibility of transferring these facilities to space.
While the viability of such technology remains uncertain, Google is taking steps to investigate this concept. The tech giant is preparing to launch its inaugural satellite for Project Suncatcher, an initiative aimed at developing orbital data centers. This prototype is scheduled for liftoff on a SpaceX Falcon 9 rocket during a rideshare mission called Transporter-18, set to take place at 2:18 p.m. ET (11:18 a.m. PT) from Vandenberg Space Force Base in California.
This isn't the first time AI chips have been tested in a space environment, but Google’s satellite aims to specifically assess the performance of its tensor processing units (TPUs) after enduring the intense conditions of a rocket launch and the harsh radiation of space.
“This first launch focuses on determining what functions effectively, pinpointing points of failure, and applying those lessons to future missions,” remarked Travis Beals, Google’s senior director of intelligence paradigms.
Looking ahead, Google plans to advance its research in 2027 with the deployment of two additional satellites.
The challenges of radiation, heat management, and space congestion are significant. The demand for advanced machine learning capabilities has led various companies to explore the deployment of data centers in orbit, a concept that partially contributed to SpaceX's remarkable IPO earlier this year.
Nonetheless, many technological hurdles lie ahead. Google's recent video series on Project Suncatcher highlights several issues to be addressed, including the risk of collisions in an already congested orbital space, which could be aggravated by the addition of data centers.
Moreover, the radiation found in space can lead to data errors known as “bit flips,” and ongoing exposure to high-energy particles can shorten hardware lifespan. Preliminary tests conducted at the Crocker Nuclear Laboratory at the University of California, Davis, showed that Google's TPU chips performed “remarkably well” under simulated space conditions, but definitive conclusions can only be drawn from actual space tests.
Heat dissipation is another critical issue. On Earth, cooling data centers relies on convection and the evaporation of water, but these methods are impractical in space’s vacuum. Therefore, satellites housing AI computing will likely require extensive radiators to manage heat from their internal processing units.
Google plans to explore a combination of radiators and heat pipes in search of solutions, indicating intentions to refine their designs based on findings from these tests.
It’s worth mentioning that the prototype satellite won't significantly advance cooling technologies; it will temporarily power down every 20 minutes to prevent overheating, even though its TPU chips will consume only about 1 kilowatt of power. This is notably less than current telecommunications satellites, which typically run between 10 to 20 kilowatts, and nowhere near the potential 100 kilowatts needed for operational AI data centers, as highlighted by Caleb Henry, research director at Quilty Space.
A recent paper from Google's research team published in Joule points out that orbital data center satellites may need to provide power equivalent to Earth-based data racks, estimated between 50 and 100 kilowatts. However, it remains unclear how much power the upcoming test satellites will yield or the specifics of their heat-management systems. Those satellites are expected to operate continuously without the shutdowns required by the initial test, possibly incorporating advanced laser communication to exchange data with one another—a technology that has seen significant development only in SpaceX's Starlink satellites so far.
While many technical challenges remain, they may indeed be surmountable. Additional concerns linger, particularly regarding the cost-effectiveness of orbital data centers compared to their terrestrial counterparts.
Establishing AI chips in space complicates hardware upgrades, as AI technology evolves every 1 to 2 years while satellites are designed for a lifespan of 5 to 7 years, creating the challenge of hardware obsolescence. According to a June report from JLL Research, the appeal of orbital data centers is contingent upon significantly decreasing launch costs, which presently are on the rise.
This is where SpaceX holds an advantage as it manufactures its rockets and leads the industry in both satellite production and launch frequency. Should the concept of orbital data centers gain traction, the scale could be monumental, with SpaceX proposing a constellation of up to a million AI-computing satellites. Meanwhile, Google has mentioned the potential for operating in clusters of 81 satellites, yet the total number needed for a fully functional system is still undetermined.
Despite the uncertainties, Google asserts it’s committed to the long-term exploration of orbital data centers. "Just like the early phases of research into autonomous driving and quantum computing required years of experimentation, venturing into space computing warrants careful, methodical steps," the company stated in a release.
Planet Labs, a San Francisco-based satellite operator and Google partner, echoes this sentiment. Co-founder and CEO Will Marshall has expressed optimism about the long-term viability of the project, noting that it will require substantial research and development funding that these tech companies are poised to invest.
