Four TPUs and a 15-Minute Duty Cycle: Google's First Real Test of AI Compute in Orbit

On Thursday, 1 October 2026, at 2:32 p.m. EDT, a SpaceX Falcon 9 lifted off from Vandenberg Space Force Base in California on the Transporter-18 rideshare mission. Space.com counted 130 payloads on board. The one that drew the attention was a refrigerator-sized spacecraft called MVP, the first piece of flight hardware for Project Suncatcher, Google’s research programme on whether machine-learning compute can live in space. Within hours Google posted that the team had made contact with the satellite and that it was “operating as expected”.

MVP is small. It carries four Trillium-generation TPUs (Tensor Processing Units, the custom AI accelerators behind Google’s cloud) on solar arrays rated at about 1 kilowatt, which is roughly what a gaming PC with a high-end graphics card draws under load. Planet, the Earth-imaging company, built the satellite. Calling this an “orbital data centre” overstates it a lot. It is a hardware experiment, and Google describes it that way too.

What the satellite will actually do

According to NPR and Scientific American, the four TPUs will run Gemma, Google’s open-weights model, answering simple queries for about 15 minutes at a time. After that the chips shut down to cool off. That duty cycle is the most honest detail in the announcement. On Earth, a TPU pod gets rid of its heat through liquid cooling loops and, in the end, the air. In a vacuum there is no air, so the only way to dump heat is to radiate it, using heat pipes and radiator panels. A kilowatt of heat is manageable. Hundreds of kilowatts per satellite is a different engineering problem.

The mission is meant to answer three questions:

Radiation. Do the chips keep producing correct results under real cosmic-ray and proton exposure, beyond what accelerator tests can simulate?
Thermal. How well does a passive radiator design move heat away from accelerators running ML workloads?
Launch survival. Did the hardware make it through the trip to orbit intact? Google’s ground tests subjected components to 50–100 g of vibration.

The orbit is a dawn-dusk sun-synchronous low Earth orbit, which follows the line between day and night so the solar panels are almost never in shadow. That removes most of the need for heavy batteries. Google says panels there can produce up to eight times more energy than the same panels on the ground. The mission is planned to last about a year.

“This first launch is about seeing what works, identifying points of failure, and applying those findings to future missions,” Travis Beals, the Google senior director who leads Suncatcher, told NPR.

What the ground work already showed

Suncatcher was announced in November 2025 together with a preprint (since published in the journal Joule). The paper is more interesting than the launch, because it shows where Google thinks the hard limits are.

Radiation tolerance. Trillium TPUs were tested with a 67 MeV proton beam at UC Davis. The high-bandwidth memory (HBM) showed irregularities after a cumulative dose of about 2 krad(Si). Everything else, including end-to-end ML workloads, kept working up to the highest dose tested, 15 krad(Si). A shielded five-year mission in the target orbit needs about 750 rad(Si), so the weakest part had a margin of roughly 3×. HBM did produce uncorrectable errors, but at a rate the team judged likely acceptable for inference. Note the word “inference”: nobody claims these error rates are acceptable for multi-week training runs.
Interconnect. Splitting a model across satellites requires inter-satellite laser links carrying on the order of 10 Tbps per link. A bench demonstrator built from off-the-shelf parts reached 800 Gbps each way (1.6 Tbps total). To make the optics work, satellites have to fly very close together. The reference design is 81 satellites within a one-kilometre radius at about 650 km altitude, which is far tighter formation flying than any operational constellation uses today.
Economics. The paper argues that orbital compute becomes roughly cost-competitive with terrestrial data-centre energy costs if launch prices fall to about $200 per kilogram, which it projects for the mid-2030s.

MVP tests none of the interconnect work. That comes next: Google’s 24 September post targets 2027 for a mission in which two satellites test the laser links between them.

The Starship arithmetic

TechCrunch extracted the sharpest number from Google’s material. Hitting the cost target means putting about 370,000 tonnes into orbit over ten years. At roughly 200 tonnes per Starship flight, that comes to about 1,800 launches, or 180 a year. (TechCrunch’s headline first said 1,600 and was corrected.) The projection also assumes SpaceX keeps cutting prices by about 20% a year, as it has since Falcon 1. In other words, Google’s space-compute plan depends on a launch vehicle it does not control reaching a flight rate that no rocket has come close to. Google is a major SpaceX investor, which explains why it is willing to think on that scale. It doesn’t make the scale any smaller.

How this compares with what came before

Google is not the first to put a modern AI accelerator in orbit. Starcloud, a Nvidia-backed startup, launched Starcloud-1 in November 2025 with a single H100 GPU. It trained Andrej Karpathy’s nanoGPT on Shakespeare in orbit and ran inference on Gemma. Starcloud got there first, but Google is the first hyperscaler to fly its own production accelerator in a design meant to scale into clustered, laser-linked constellations, and the first to publish detailed ground radiation data alongside it. Other startups, Satlyt among them, are chasing the same idea.

Why it matters for IT

The motivation is on Earth. Scientific American cites projections that data centres could use about 3% of the world’s electricity by 2030, roughly double today’s share, and frontier training campuses already draw close to a gigawatt each. Grid connections, permitting and cooling water are now the constraints on AI build-out, more than chips. Space looks attractive because it removes those constraints: near-constant sunlight, no grid queue, no water.

For developers, the near-term significance is narrower. If any of this works, the first workloads will be inference and in-orbit processing of satellite data (Earth imaging, for example), which avoids sending raw data down to the ground. The University of Pittsburgh’s Alan George pointed to that downlink bottleneck: “We don’t have the convenient, easy way to get that data back to the ground.” Latency-sensitive, chatty, training-scale workloads are much further away. Fault-tolerant software that tolerates radiation-induced memory errors would become a real systems problem in its own right.

Open questions

Thermal scaling. A 15-minute duty cycle at 1 kW is a starting point, not a design. Radiator mass grows with power, and mass is what launches charge for.
Debris and maintenance. Eighty-one satellites flying in a one-kilometre formation is a collision-avoidance problem nobody has had to solve operationally, and failed TPUs cannot be swapped out.
Cost honesty. The $200/kg break-even compares energy costs. Replacing hardware every five years, extra ground stations, and the efficiency lost to radiation-hardening software all count against it. MIT’s Kerri Cahoy summed it up: “Launching enough mass to generate the power is huge.”
What counts as success. MVP’s results will be reported by Google. Independent verification of in-orbit error rates would make them much more useful.

What the MVP launch establishes is modest but real: Google is now collecting data from orbit instead of extrapolating from proton beams. Whether that leads to orbital AI clusters or a well-documented dead end depends far more on Starship’s flight rate than on anything happening inside the four TPUs.