Google’s Suncatcher project has successfully brought its experimental space-based data center online, establishing communication with Earth following its launch this month aboard a SpaceX Falcon 9, which also carried over 100 other satellites.
Equipped with four Google Tensor Processing Units (TPUs) and capable of generating up to 1 kW of power, this satellite aims to lay the groundwork for upcoming constellations of tightly grouped satellites that will execute substantial inference tasks in space. Nevertheless, the high power consumption of the TPUs limits their operation to approximately twenty minutes each, adding complexity to the already challenging task of space-to-ground communication.
Following a three-week commissioning period during which its subsystems were thoroughly tested, the spacecraft became fully operational. This week, ground engineers issued their initial command, requesting Google DeepMind’s Gemma 4 model, pre-installed on the satellite, to compose a haiku reflecting on space. The result was:
Blue marble below
Data streams in silent grace
Starlight feeds my core. ✨
Google stated that it evaluated the model using these space-themed poetic prompts (perhaps a poet would have been more suitable?) and, crucially, tasked the spacecraft with providing updates on the surrounding orbital environment.
A company spokesperson informed TechCrunch that the satellite was functioning as expected.
The current testing phase will assess the chips’ performance under thermal strain and within the harsh space radiation environment. While Google has conducted extensive terrestrial testing of its TPUs in particle accelerators, it now needs to observe the actual impact of space radiation on the processors. Radiation exposure can induce “bit flips,” altering stored values in computer memory and potentially corrupting data without detection by standard monitoring systems.
Previous ground-based tests indicated that radiation’s effect would be minor for inference tasks but more significant for model training. Now, the Google team will evaluate these dynamics in an authentic space environment, rather than a simulated one.
Should the TPUs on this satellite demonstrate sufficient capability, the subsequent phase involves deploying two additional satellites with these advanced chips in 2027. These future spacecraft will feature enhanced power generation and showcase optical links for inter-satellite communication.
Among the concepts for space-based data centers, Google’s Suncatcher project stands out for its emphasis on high-speed inter-satellite links to facilitate cohesive distributed workloads. While SpaceX’s data centers are also anticipated to employ laser links, their greater orbital separation might hinder effective shared processing. Conversely, companies such as Starcloud and Cowboy Space intend to handle inference tasks on individual satellites.
The aspirations of Google and SpaceX are deeply intertwined. Alphabet, Google’s parent company, possesses a substantial investment in SpaceX. For both entities to deploy their orbital data centers, they will depend on Starship, the heavy-lift rocket central to Elon Musk’s space strategy, which is currently under development and must achieve full and rapid reusability.
According to a peer-reviewed study released this month, Suncatcher researchers projected that SpaceX would need to launch Starship 1,800 times to attain the same cost efficiencies through learning as observed with its previous launch vehicles.
The Suncatcher initiative epitomizes a quintessential Google moonshot, with its primary objective being the development of cost-effective infrastructure for the 2030s. Nonetheless, achieving this goal might hinge on how the broader company navigates the immediate challenges and prospects presented by AI. This week, Alphabet, Google’s parent company, experienced a significant stock decline after disclosing that its efforts to integrate AI into its search operations are negatively impacting ad revenue growth.
Correction: The Suncatcher prototype satellite is equipped with four TPUs, but is currently designed to operate only one at a time.