Google's Project Suncatcher sends TPUs to orbit to test AI compute in space

Google's orbital AI moonshot reaches its first launch

Google is preparing to put a prototype satellite carrying its Tensor Processing Units (TPUs) into low Earth orbit, the first in-flight test of Project Suncatcher, a long-term research effort asking whether machine learning infrastructure could one day scale beyond Earth. According to a blog post from Google, the satellite will fly on SpaceX's Transporter-18 rideshare mission and was developed with satellite operator Planet. The project was first announced last year, and Google frames it as deliberate, early-stage research — comparable, it says, to the years of experimentation that preceded its autonomous driving and quantum computing work. The announcement drew discussion on Hacker News's front page.

The pitch: abundant power above the atmosphere

The core argument for orbital compute is energy. Google says satellites in low Earth orbit receive near-constant sunlight and can generate up to eight times more solar power than equivalent hardware on the ground. Over time, the company imagines linking multiple constellations of satellites so they can jointly handle larger AI workloads while in orbit — effectively an orbital data center powered around the clock by the sun.

Proving the hardware first

Before any of that, Project Suncatcher has to answer a simpler question: do Google's AI chips work in space at all? The first mission is built to gather in-orbit data on how TPUs cope with the physical stress of launch and the radiation and thermal extremes of orbit.

Ground testing has already covered part of that. A rocket ascent to low Earth orbit lasts about ten minutes, during which the spacecraft sustains acceleration loads up to 10 g while individual components such as TPU chips can see forces of 50 to 100 g. Google reports it shook the satellite on all three axes to mimic launch frequencies and that the hardware held up — an outcome the team describes as a pleasant surprise, since such tests rarely go to plan.

For radiation, the team placed TPUs in a proton beam at UC Davis's Crocker Nuclear Laboratory while running live AI workloads, watching for errors such as bitflips. Initial results, per Google, show its Trillium-generation TPUs surviving a total ionizing dose greater than what they would absorb over a five-year space mission. Some effects can only be measured in orbit, which is the point of the upcoming launch.

Cooling chips in a vacuum

Thermal management is the other major unknown. TPUs concentrate a great deal of heat in a small area, and in a vacuum there is no airflow to carry it away; heat can only leave through radiators. Google says it is developing a cooling design that combines heat pipes with radiators and has already tested it in a thermal vacuum chamber that simulates both the temperature and pressure conditions of space. How the system performs in flight will drive future design revisions.

Lasers between satellites

The architecture Google describes goes well beyond a single spacecraft: future satellite designs would each carry dozens of TPU chips and operate in clusters. Keeping enough bandwidth for AI workloads means each satellite must know its own position and where it sits relative to its neighbors, communicating over laser links.

Laser communication in space already exists, but Google notes that current state-of-the-art systems are optimized for low bandwidth over long distances, whereas its use case demands very high bandwidth over extremely short distances. The pointing precision required is something Google likens to striking a coin-sized target from miles away while both endpoints are moving. That gets a real test in 2027, when the company plans to place two satellites in orbit to trial the link.

Why it matters

Power and cooling are now among the binding constraints on AI data center buildouts, and Project Suncatcher is the most literal version of the response: if energy is the bottleneck, go where the sun never sets. It is also an early signal that custom AI silicon portfolios like Google's TPU line may eventually need space-qualified variants if orbital compute becomes practical.

None of that is settled yet. Google itself frames this launch as a way to find failure points — radiation behavior, thermal performance, launch survival — rather than a step toward deployment. But the project has launch partners, a scheduled two-satellite interconnect milestone in 2027, and ground data already suggesting its chips tolerate radiation at multi-year mission levels. For cloud and AI infrastructure watchers, the question is no longer whether compute can physically reach orbit, but whether the economics of orbital data centers can ever close.