Google tests AI chips in space as terrestrial data centers strain power grids

On Thursday, January 1st, Google sent four artificial intelligence chips into space aboard an experimental satellite launched by SpaceX from California. The year-long mission will test whether the company’s processors can survive in orbit. The prototype, developed in partnership with Planet, carries tensor processing units (TPUs) and is part of the Suncatcher project, unveiled in November 2025 to explore space-based data centers.

Why take computing to space?

The growth of artificial intelligence and digital services is increasing demand for computing infrastructure. In 2025, data centers consumed 485 terawatt-hours (TWh) of electricity globally, according to the International Energy Agency (IEA). The updated central projection for this year suggests that consumption could reach 950 TWh by 2030—nearly double the 2025 level and equivalent to about 3% of global electricity demand. Within this total, consumption by AI-focused facilities is expected to triple over the period.

The pressure also depends on the capacity of power grids. The IEA estimates that about 20% of planned data centers could face delays if network bottlenecks are not resolved. Building new transmission lines can take four to eight years in advanced economies. Goldman Sachs identifies the same trend, projecting a 175% increase in data center electricity consumption by 2030, compared to 2023.

A 2024 study by the Lawrence Berkeley National Laboratory estimated that U.S. data centers directly consumed 66 billion liters of water in 2023. The data includes facilities of different types and refers to consumption at the facility itself, separate from the water used to generate the electricity it purchases.

The main advantage of Suncatcher is solar energy. According to Google’s study, a panel in the right orbit can be up to eight times more productive than on Earth, with almost continuous exposure to the Sun. The comparison refers to the energy generated, not the intensity of the light.

Technical challenges: heat and radiation

In the vacuum of space, convection does not occur, which prevents conventional cooling. Heat must be dissipated through thermal radiation, requiring radiators and heat transfer systems. Google is testing heat transfer tubes and radiators, previously evaluated in a chamber that simulates the space environment.

Radiation can also cause errors in electronic components. The team conducted tests with proton beams at the University of California, Davis, but acknowledges that other conditions need to be evaluated in orbit.

Next steps and competition

Google plans to test laser communication between two satellites in 2027. The connection requires precision because the equipment will be in motion and will need to keep the beams aligned. This connection is essential for large AI tasks to distribute work among many processors and achieve performance comparable to terrestrial data centers, with connections capable of carrying tens of terabits per second.

Other companies are also advancing in the space race. In November 2025, Starcloud launched a satellite with an Nvidia H100 processor. In March 2026, Kepler announced a processing infrastructure distributed across ten satellites, with 40 Nvidia Jetson Orin modules and optical connections. Axiom Space announced the launch of two orbital data center modules in January 2026, alongside Kepler’s optical relay network.

Proposals for large constellations reach another scale. SpaceX has filed a request for up to one million satellites. Blue Origin proposed up to 51,600 satellites in the Sunrise project. The American Astronomical Society asked regulators to reject Blue Origin’s proposal, citing impacts on optical, infrared, and radio observations, as well as concerns about the orbital environment. The Secure World Foundation advocated for phased authorizations, with demonstrations and performance criteria before expansion.

Economic viability depends on launch costs

Google’s study suggests that if launch prices fall below $200 per kilogram by the mid-2030s, operating space infrastructure could have a cost comparable to the energy expenditure of an equivalent terrestrial facility. The comparison does not demonstrate equality in the total cost of the two models.

Frequently Asked Questions

• Does Google already have a commercial data center in space? The current mission is an experiment with four chips, not a large-scale commercial facility.
• Why use satellites to process AI? The Suncatcher project investigates solar-powered processing and communication between satellites; technical and economic viability still needs to be proven.
• How are computers cooled in space? Heat is transferred to surfaces that emit it as thermal radiation, since the vacuum does not allow convection.
• What is Google’s next step? The next test plans laser connections between two satellites in 2027.