Google’s orbital data-center test exposes power and cooling limits

Google launched its first experimental orbital data-center satellite on October 1 aboard a SpaceX Falcon 9 rocket as part of the Transporter-18 rideshare mission to low-Earth orbit. The refrigerator-size spacecraft, dubbed MVP, hitched a ride rather than flying on a dedicated launch. Its aim was to validate Google’s vision for a constellation of AI satellites, and the single platform will function for only a short time. The Project Suncatcher prototype came online in late October as part of Google’s long-term initiative to establish data infrastructure and machine learning capabilities in orbit.

Google accelerated its original timeline by placing its hardware aboard an existing Planet Labs spacecraft instead of waiting for newly built custom satellites. The initial plan was to launch two custom satellites in 2027. Planet Labs built the refrigerator-size spacecraft, while Google supplied four custom tensor processing units inside, and the firm did not build the satellite from scratch. The single spacecraft will function for only a short time as a test of technologies.

Each of the four onboard tensor processing units, which are AI accelerator chips, has the computing power of one server in a terrestrial data center. These accelerators handle AI model training and inference work to produce tokens. Terrestrial data centers deploy thousands of similar chips and require enormous amounts of electricity. That scale contrasts with the single platform’s tight physical and electrical limits.

The satellite generates only enough electricity to run one of its four tensor processing units for approximately 20 minutes at a time. Solar panels supply about one kilowatt of power, roughly enough for a household appliance such as a microwave. Google needs to solve major power limits before expanding the network, and the craft operates under tight physical and electrical limits. These electrical limits are among the hurdles Google must clear as it considers a larger constellation.

Google confirmed on October 30 that the Project Suncatcher prototype ran DeepMind’s Gemma 4 model in orbit and produced a haiku using one tensor processing unit. The confirmation came roughly three weeks after the October 1 liftoff. Once ground checks were complete, ground engineers requested a space description from Gemma 4, and the system also produced limericks and orbital status reports. Google stated the equipment worked as expected.

During tests, the tensor processing units run for 15-minute stretches to run Google’s Gemini AI and process short queries, then shut down to cool off. Cooling compute hardware in space is difficult because fans do not work and sending water is impractical. Ground centers use water circulation and moving air to remove heat from processors, but neither method works effectively in orbit. Google is experimenting with pipes and radiators to shed warmth.

Because space lacks air to move heat outward, servers require extensive radiator surfaces, and Brookings cites an estimate of 2.15 million square feet of radiator area for an orbital data center. This cooling difficulty is among the barriers to commercial orbital computing. Ground centers rely on air and water circulation to remove heat, neither of which works effectively in orbit. The extensive radiator need reflects the challenge of shedding heat without an atmosphere to carry it away.

Continuous radiation exposure is a critical issue because standard commercial electronics are highly susceptible to radiation corruption in the harsh environment of space. High-energy particles strike equipment outside Earth’s protective atmosphere and can trigger memory errors that corrupt calculations. Google must address this hazard together with power limits and orbital data links. The harsh environment outside Earth’s protective atmosphere exposes equipment to high-energy particles.

GlobalFoundries manufactures space-grade semiconductor chips already proven in Earth orbit and on missions beyond Mars and Jupiter. The company puts forward custom chip designs for low-Earth-orbit needs. It utilizes Gallium Nitride power chips to reduce wasted energy and maximize a satellite’s limited electricity supply, addressing the power constraints faced by orbital hardware.

GlobalFoundries is also deploying silicon photonics chips that transmit data as light, enabling rapid transfers that endure intense radiation. Its silicon photonics hardware sends information as light, a method suited to the demands of orbital data links. Tom Barber, Vice President of Data Center and Communications Infrastructure at GlobalFoundries, said, “Optical communications could enable satellites to exchange data at the high speeds Google requires while simultaneously keeping power consumption in check.”

Google also imagines satellites joined by laser links in orbit, a network that would probably need dedicated launches rather than rideshares. The company plans to launch two more satellites in 2027 and has proposed launching clusters of more than 80 satellites. Long-term, each satellite would carry a dozen AI chips, exchanging signals with each other and the ground through lasers. Successful trials would lead to two more powerful spacecraft in 2027 to verify those connections.

In 2024, data centers accounted for about 1.5 percent of the world’s electricity consumption, and the International Energy Agency estimates the percentage might double by 2030. Rising electricity needs have led Google and other firms to propose orbiting facilities instead of ground sites that face local resistance and limited power capacity. Near-constant sunlight in space can provide up to eight times more solar power than on Earth, one of the attractions of orbital computing. The figure comes from a September 24 statement by Google’s head of Project Suncatcher.

A study published in Nature found that electricity consumption by AI-dedicated data centers was 118 terawatt hours in 2024 and could soon exceed 239-295 terawatt hours by 2030. The projected growth underscores the strain on terrestrial grids and the motivation for testing orbital computing as relief for specific workloads. Goldman Sachs estimates a growth of 170 percent in electricity needed for data centers around the globe between 2025 and 2030. This estimate reinforces the case for exploring space-based power for computing.

JLL notes that connecting terrestrial data centers to grids can take 2 to 10 years. Space offers solar energy without grid-connection delays. JLL identifies AI training, batch work and space-generated data processing as suitable tasks for orbital facilities, while live transactions need to happen near end users.

BCG estimates the orbital cost premium today lies between 2.5 and 3 times higher than ground-based facilities. Expense ranks among the chief obstacles to orbital computing. BCG projects that orbit could capture 10 to 15 percent of the AI data-center market by 2040, though the firm calls this a projection rather than a certainty.

Starcloud has flown an H100, an Nvidia AI processor, into orbit, and Cowboy Space is pursuing single-satellite inference, meaning it runs a trained AI model on one satellite to produce answers. SpaceX is seeking about $40 billion for Nvidia chips for AI data centers on Earth and potentially in orbit, though financing remains proposed. This reported fundraising effort relates to the broader data-center race moving to space, and Google’s approach relies on its own tensor processing unit hardware rather than Nvidia chips. Other firms are also advancing orbital computing hardware alongside Google’s Project Suncatcher.

Federal approval came in July for Reflect Orbital to build and fly a satellite whose mirror can steer sunlight toward chosen spots on Earth. The firm joins others seeking to deliver additional sunlight to ground locations for on-demand solar power. This effort is distinct from Google’s orbital computing work.

Travis Beals said that if everything works perfectly five years from now, it probably means Google has not taken enough risk and has not learned as much as it could. He also said, “This first launch is about seeing what works, identifying points of failure and applying those findings to future missions.” The single platform will function for only a short time as a test of technologies.