Google is scheduled to send four of its Tensor Processing Units into low Earth orbit on Thursday, October 1, in the first flight test for Project Suncatcher, its research effort to explore whether large AI computing systems could eventually operate in space.
The prototype is one of 130 payloads on SpaceX’s Transporter-18 rideshare mission from Vandenberg Space Force Base in California. SpaceX lists a 58-minute launch window opening at 11:18 a.m. PT, with a backup opportunity on October 2. The mission timeline places deployment of the Planet-built Project Suncatcher M1 spacecraft about 61 minutes after liftoff.
This is not an orbital data center in any practical sense. It is a small hardware-survival experiment designed to answer more basic questions: Can commercially derived AI accelerators tolerate launch vibration and years of radiation? Can four hot chips shed enough heat in a vacuum to run useful workloads? And can Google collect operating data that justifies a more ambitious two-satellite communications test planned for 2027?
Four TPUs, one kilowatt and short compute bursts
The refrigerator-size spacecraft carries four Google AI accelerators and solar panels that produce about one kilowatt, according to reporting by Ars Technica. That is tiny beside a terrestrial AI cluster, where thousands of accelerators can draw megawatts and work continuously.
The power budget is not the experiment’s tightest limit. In orbit, heat cannot be carried away by moving air. Google’s design transfers heat from the chips through thermal interface material into aluminum and copper heat pipes, then to a radiator that emits it into space. The system is expected to run Gemini workloads in bursts of roughly 15 minutes before the chips pause and the radiator catches up.
That duty cycle makes the mission a thermal test rather than a demonstration of economically useful AI compute. A successful flight would show that the cooling path works well enough to measure, refine and scale. It would not show that an orbital system can match a data center’s sustained utilization, serviceability or cost.
Google has already tested the cooling hardware in a thermal-vacuum chamber. Flight adds the conditions a ground chamber cannot fully reproduce together over time, including repeated sunlight and shadow cycles, radiation exposure and the thermal behavior of the complete spacecraft.
Radiation testing produced a promising result and a warning
Before launch, Google exposed a Trillium V6e Cloud TPU and its AMD host system to a 67 MeV proton beam at the University of California, Davis. The company’s Project Suncatcher research paper estimates that hardware in its proposed sun-synchronous orbit, protected by the equivalent of 10 millimeters of aluminum, would receive about 150 rad of ionizing radiation per year.
The TPU compute tests continued working through a cumulative dose of 15,000 rad without a permanent failure attributed to total ionizing dose. High-bandwidth memory was less resilient: stress tests began showing irregularities after 2,000 rad, still almost three times the paper’s five-year survival requirement of 750 rad.
Single-particle strikes create a subtler problem. Google observed uncorrectable memory errors and estimated a failure probability of about one in 10 million inferences under the tested assumptions. That may be manageable for inference if workloads can be retried. Long training runs are less forgiving because an undetected bit flip can silently corrupt model state or results. The researchers said training behavior and system-level mitigations need more study.
The orbital prototype can now test whether the beam experiments predicted the real error rate. Useful results will include not only whether the TPUs stay alive, but how often memory correction intervenes, whether the host system resets and whether errors cluster during solar events.
The hard part is turning satellites into one computer
Putting an accelerator in orbit is not the same as assembling a scalable AI system. Modern model training depends on fast, low-latency communication among many chips. Project Suncatcher proposes tightly grouped satellites connected by optical links, with dozens of TPUs on each spacecraft.
Google’s paper models an 81-satellite cluster within a one-kilometer radius. At those distances, laser links need enough bandwidth to act more like a data-center interconnect than conventional satellite communications. The spacecraft must also maintain precise relative positions while accounting for atmospheric drag, Earth’s uneven gravity, solar radiation pressure and other forces.
The first mission does not test that network. Google plans to put two satellites in orbit in 2027 to evaluate high-bandwidth links and formation control. The company describes the pointing problem as comparable to hitting a coin-size target from miles away while both endpoints are moving.
Ground connectivity remains another constraint. Some AI work could stay in orbit, but models, datasets, checkpoints and outputs still have to move between space and Earth. A constellation that generates abundant solar electricity but cannot feed its accelerators or return results at competitive speed would not function as a practical cloud platform.
Why the economics are still hypothetical
The attraction is straightforward: satellites in low Earth orbit can receive near-continuous sunlight and, by Google’s estimate, generate as much as eight times the solar energy available to comparable panels on Earth. Orbital systems would also avoid competing directly for local grid connections and cooling water.
Launch price, hardware mass, replacement cycles and communications equipment can erase those advantages. Google’s economic model assumes that launch prices to low Earth orbit could fall to $200 per kilogram or less by the mid-2030s. Under that assumption, it estimates annualized power cost in orbit could enter the broad range of U.S. terrestrial data-center electricity costs.
That is a scenario, not a forecast. The paper explicitly stops short of a full economic feasibility study, and the calculation depends on steep launch-price reductions, high launch volume and spacecraft that deliver much more compute per kilogram than today’s prototype. Terrestrial data centers will also keep improving while Google works through the orbital engineering problems.
There are external costs to resolve as well. Large constellations add collision and debris risk, affect astronomy and require end-of-life disposal plans. Those concerns become materially different if the concept grows from a four-chip experiment into clusters of dozens or hundreds of compute satellites.
What would count as success
The near-term scorecard is deliberately modest. First, the payload must survive launch. Then Google needs reliable telemetry from the TPU, memory, host computer, power system and radiator across changing orbital conditions. Stable workloads, explainable error rates and repeatable thermal cycles would justify the 2027 inter-satellite test.
The most revealing result may be a limit rather than a breakthrough: how long the four chips can run before heat forces a pause. That number will determine the radiator area, spacecraft mass and duty cycle of any larger design, tying thermal engineering directly to launch economics.
Project Suncatcher M1 will not move meaningful AI demand away from Earth’s power grids. It can do something more useful at this stage: replace several optimistic assumptions with flight data. If orbital AI infrastructure is ever going to work, it starts with proving that four chips can survive, cool down and produce trustworthy results.
Sources: Google Project Suncatcher update; SpaceX Transporter-18 mission page; Google’s Project Suncatcher research paper; Ars Technica.
