Google's Suncatcher Satellite Reaches Orbit With Four TPUs, 1 Kilowatt and a 15-Minute Compute Limit

Google's first orbital test of its TPU accelerators is four Trillium-generation chips on a refrigerator-sized satellite, about 1 kilowatt of solar power, and a thermal budget that allows 15 minutes of compute at a time. That is the number to hold against the phrase "data center in space."

The prototype, part of Project Suncatcher, was built with Earth-imaging operator Planet and rode a SpaceX Falcon 9 on the Transporter-18 rideshare mission, one of 130 payloads. TechTimes puts the launch at 2:32 p.m. EDT on October 1 and says first contact was confirmed that evening. OPB's text gives October 2 for the launch while its URL is dated October 1; the Google post, written before launch, names only the mission.

What was launched

TechTimes reports the satellite carries four unmodified commercial TPUs in a sun-synchronous orbit at 650 kilometres, with about 1 kilowatt from its panels. It compares that to a single terrestrial server. OPB reports it will run Google's Gemma model in 15-minute intervals because of heat.

Item	Figure	Source
Chips	4 Trillium TPUs	TechTimes, OPB
Solar power	About 1 kW	TechTimes
Orbit	Sun-synchronous, 650 km	TechTimes
Compute duty cycle	15 minutes	TechTimes, OPB
Follow-on satellites	Two, in 2027	Google, OPB

The picture above shows a Falcon 9 first stage descending during the Transporter-6 rideshare in January 2023, an earlier mission in the same series as the one that carried Suncatcher.

Radiation held; heat is the constraint

Before launch, Google tested the chips in a proton beam at UC Davis's Crocker Nuclear Laboratory. Its post says the Trillium TPUs "can survive a radiation total ionizing dose greater than what they would receive during a five-year space mission."

TechTimes supplies the numbers: the chips survived 15 krad(Si) against roughly 750 rad(Si) expected over five years, a margin of 20 times. High-bandwidth memory showed irregularities above 2 krad(Si) with no hard failures, most bit flips cleared on restart, and one silent data-corruption event was recorded.

1. Tested without hard failure15 krad(Si)
2. Memory irregularities begin2 krad(Si)
3. Expected over five years0.75 krad(Si)

Source: TechTimes, accessed 2026-10-02 (the 750 rad(Si) five-year figure is expressed here as 0.75 krad)

The binding limit is thermal. The satellite sheds heat through heat pipes and external radiators, an arrangement tested in thermal vacuum chambers but not yet through orbital cycling, where TechTimes says temperatures swing by more than 100 degrees Celsius. Brandon Lucia of Carnegie Mellon told OPB the physics: in a satellite, "using more power to do the computations means dumping more heat into the confined environment inside of the satellite."

The denominator behind the pitch

Four chips on 1 kilowatt is, in TechTimes's comparison, one terrestrial server. OPB reports that Google intends clusters with dozens of TPUs per satellite, linked between satellites. Google's post names a 2027 test of two satellites joined by laser communication; the follow-on pair is the first time the optical link flies.

The economics are further out. TechTimes cites a Google paper in Joule that needs launch costs below $200 a kilogram by the mid-2030s, about 1,800 Starship launches over ten years, to approach parity with ground datacentres. Travis Beals, the senior director of Project Suncatcher, said he does not "see this being something where it's cheaper to do this in the next five years." The ground-based version of the same capital question is Amazon's plan to have investors hold $8 billion of its Nvidia chips and lease them back.

Launch cadence is the other input. The Joule paper's 1,800 Starship launches over ten years works out to roughly 180 a year, while China has flown 70 orbital launches of every kind in 2026. Starcloud, which put an Nvidia H100 in orbit last November according to OPB, is the nearest comparison.

What the prototype will and will not prove

TechTimes sorts the mission into two lists. It will show whether the thermal architecture works in real orbital conditions, and whether the chips survive months of thermal cycling without error rates leaving the recoverable range. It will not show economic viability at scale, inter-satellite laser performance in orbit, or where launch costs go.

That division matters for how to read the first telemetry. A clean result from a single 1-kilowatt satellite is evidence about heat and radiation on one design. It says nothing about the laser link between satellites, because the link has no partner to talk to until the 2027 pair flies. TechTimes also lists Starcloud, SpaceX, Axiom Space and Blue Origin among the other orbital-compute efforts, so Google's thermal result is one data point in a field of several.

What would change this read

The thermal result decides it. If the radiators hold the chips at the 15-minute cycle across the first months of orbital temperature swings, the constraint moves from physics to cost. TechTimes lists two unknowns: whether the ceiling degrades as panels age, which it puts at 0.5 to 0.8 percent a year, and whether thermal cycling fatigues the silicon. Google's next data point is the two-satellite launch in 2027.