Google is set to send its own AI chips into low Earth orbit for the first time, exploring how to extend AI computing power from ground-based data centers into space.
Next week, an experimental satellite jointly developed by Google and satellite imaging company Planet will launch aboard SpaceX's Transporter-18 rideshare mission. The satellite carries four of Google's self-developed Trillium TPU chips, which will run Google's Gemini model in orbit and receive and respond to commands from the ground.
This marks the first in-orbit test of Google's "Project Suncatcher." Google's long-term goal is to deploy an AI computing cluster composed of multiple satellites in Earth's orbit to obtain more abundant solar energy and alleviate the power and land constraints facing ground-based data centers.
However, this mission remains an early-stage technology demonstration. James Manyika, Google's Senior Vice President of Research, said the current focus is on verifying whether the chips and related hardware can withstand launch, radiation, and the thermal environment of space, rather than immediately rolling out a commercial system. A more practical issue is that running AI chips in space has yet to fully solve even heat dissipation 鈥?the satellite can only operate at full power for about 15 minutes before it must shut down to cool.
Why AI Computing Is Aiming for Space
One of the core reasons driving Google's exploration of space-based data centers is the growing energy demand of AI infrastructure. As data centers continue to expand, power supply, land, and grid capacity are all facing increasing constraints, while low Earth orbit can capture more abundant solar energy, with theoretical power generation reaching up to 8 times that of the ground.
This concept was first proposed by Google Vice President Blaise Ag眉era y Arcas. Three years ago, he formed the idea during a closed-door meeting discussing AI's energy bottleneck, and subsequently reported it to CEO Sundar Pichai and co-founder Sergey Brin. The project gained support and was launched.
Google estimates that as rocket launch costs continue to decline, by the mid-2030s the overall cost of operating data centers in space could approach that of ground-based data centers. But to achieve this goal, the basic engineering challenges of running chips in orbit must first be resolved.
Chips in Space: Three Hurdles to Clear
The first hurdle is launch impact. During rocket ascent, the satellite must withstand approximately 10 times the force of gravity, and the internal chip components may endure impact forces reaching 50 to 100 times the force of gravity. Google's team subjected the satellite to full three-axis severe vibration testing, and after the test, none of the fastening screws had loosened.
The second hurdle is space radiation. After leaving the protection of the atmosphere, cosmic rays and high-energy particles can penetrate electronic components and cause computing errors. Google sent the Trillium TPU to the Crocker Nuclear Laboratory at the University of California, Davis, using proton beams to bombard the chip while running AI workloads to simulate the cumulative radiation dose of a five-year orbital mission. Tests showed that the chip can withstand this dose, and after computing errors occur, it can recover through a system reboot.
The third hurdle, and currently the most challenging, is heat dissipation. In a vacuum environment, there is no air convection, so the heat generated by the chips can only be expelled through conduction and radiation. Google designed a purely conductive heat dissipation structure that transfers heat to a metal layer, which then radiates it into the vacuum through radiator panels.
But this solution still limits the chip's ability to operate continuously: the satellite can only run at full power for about 15 minutes each time before it must shut down to cool. For a data center that needs to continuously process AI tasks, this means thermal management remains a key issue that must be resolved for large-scale deployment.
Space Data Center Race Heats Up: Google Plans Dual-Satellite Validation in 2027
The experimental satellite being launched this time is roughly the size of a household refrigerator. Its unfolded solar panels provide only about 1 kilowatt of power, and its designed operational lifespan is one year, after which it will eventually fall into the atmosphere and burn up. Compared with a full data center, it is more like an orbital test platform for validating key technologies.
Google's long-term vision is to deploy a computing cluster composed of multiple satellites. The current concept is for 81 satellites to fly in formation at an altitude of about 650 kilometers within a radius of 1 kilometer, sharing computing power and data through ultra-high-bandwidth inter-satellite laser links.
This places new demands on laser communication. Traditional space laser communication is mainly aimed at long-distance data transmission, while the satellites envisioned by Google need to maintain precise alignment within a distance of several kilometers under high-speed relative motion. Google compares this difficulty to "hitting a moving coin during high-speed motion."
The first in-orbit test of the related technology is planned for 2027, when Google will launch two satellites simultaneously to verify the satellites' laser networking capability.
Google is not the only company betting on this direction. According to The Next Web, Nvidia-backed startup Starcloud sent a satellite carrying H100 chips into orbit in November last year; SpaceX plans to launch its first batch of Starmind AI satellites carrying Nvidia chips as early as the end of 2027.
However, space-based AI infrastructure still faces clear obstacles before large-scale application. Currently, about 44,870 objects have been tracked in low Earth orbit, most of which are debris, and the orbital space near 650 kilometers is especially crowded. At the same time, launch costs, in-orbit engineering, and the ability to manufacture satellites at scale will all affect whether this concept can truly move from a technology experiment to commercial infrastructure.
Comments