🔍 Read the full analysis: Google’s Haiku From Orbit: A Test Of AI, Not A Final Verdict on ThorstenMeyerAI.com
Get business pricing on monitors, keyboards and dev gear
- Business-only prices and quantity discounts
- Tax-exempt purchasing
- Multiple users, one account, clear invoices
TL;DR
A small satellite built by Google and Planet has used a Google TPU in orbit to generate a haiku with the Gemma 4 model. The test shows that a data-center-class TPU can run in space, but it does not establish that orbital AI computing is practical or economical. Google’s next tests will examine radiation and heat, followed by a planned two-satellite link test in 2027.
Google says a satellite carrying one of its Tensor Processing Units (TPUs) generated a haiku in orbit this week using the Gemma 4 AI model, a first working test of the company’s Project Suncatcher. The result shows that a data-center-class TPU can run in space, but the prototype is small and does not establish whether orbital computing can compete with Earth-based data centers.
The prototype was built with satellite company Planet and launched on October 1 aboard a SpaceX rideshare mission, according to Google’s October 30 account. After roughly three weeks of commissioning, engineers sent a prompt during a ground-station pass as the satellite traveled over Svalbard, Norway. Google said the satellite returned a haiku generated by Gemma 4; engineers also tested space-themed limericks and asked the system to report its technical status.
TechCrunch reported that the satellite carries four TPUs and can generate up to about one kilowatt of power. Power and heat constraints mean it can operate only one chip at a time, in short bursts. Google described the spacecraft as “operating as expected,” and said it plans to test how the processors perform under thermal stress and radiation exposure. The poem is a demonstration of a narrow task, not evidence that the system can support large-scale AI training in orbit.
The mission is smaller and earlier than Google’s original stated plan. When it introduced Suncatcher in November 2025, Google described a future constellation of solar-powered satellites linked by lasers and said it expected a two-satellite learning mission by early 2027. Google now says the next major milestone is a pair of interconnected satellites in 2027. TechCrunch reported that those spacecraft are intended to test optical links and carry more power.
Infographic 01 · What flew and what it did
One satellite, four chips, one haiku
Google’s Project Suncatcher prototype is the first test of whether its AI data center chips can work in orbit. Here is what is on board and how it got there.
- 4 Nov 2025Project Suncatcher announcedResearch post and preprint; plan: two prototypes with Planet by early 2027
- 1 Oct 2026Prototype launchedSpaceX Transporter-18 rideshare, built with Planet; peer-reviewed paper in Joule
- Oct 2026About three weeks of commissioningHost and TPU trays powered up, early machine learning tests
- Week of 26 Oct“Hello World”First prompt to Gemma 4; the haiku came down during a pass over Svalbard, Norway
- NextHeat and radiation tests in orbitHow the TPUs behave under thermal stress and real space radiation
- 2027Twin interconnected satellitesMore power and laser links between satellites (Google, TechCrunch)
Scale check
The satellite vs. one power deal on Earth
890 MW = 890,000 kW, about 890,000 times the prototype (my arithmetic). The prototype bar is drawn at a visible minimum; to scale it would be invisible.
Not the first AI chip in orbit
Starcloud-1, launched November 2025 (CNBC)
- Carried one Nvidia H100, a data center GPU
- Trained a small language model in orbit
- Ran Google’s open model Gemma up there in December 2025
- Google’s claim is narrower: the first response from a data center class TPU in orbit
Sources: Google, 30 Oct 2026 · Google, 1 Oct 2026 · Google Research, 4 Nov 2025 · TechCrunch, 30 Oct 2026 · TechCrunch, 1 Oct 2026 · Google and Constellation, 6 Oct 2026 · CNBC, 10 Dec 2025.
Powered by Thorsten Meyer AI
From One Chip to Orbital Computing
The test matters as a practical demonstration, but its implications remain limited. It shows that a Google TPU and Gemma model can be operated on a satellite and return a result to Earth. It does not test the tightly connected computing clusters that would be needed for demanding AI workloads, nor does it show that sending computation into orbit is less expensive than running it on the ground.
Google’s rationale is access to solar energy. The company argues that solar panels in suitable orbits could produce substantially more energy than panels on Earth. But power generation is only one part of the economics: satellites must be built, launched, cooled, connected and eventually replaced. The company’s own research estimates that at current launch prices, a kilowatt of orbital capacity would cost about $14,700 per year in launch costs. Its estimate falls to about $810 if launch costs reach $200 per kilogram, a level Google’s paper compares with power costs for U.S. data centers. The lower figure depends on a major reduction in launch costs, not a capability established by this test.
For companies planning AI infrastructure, the result is an early signal that orbital hardware is being tested, not a near-term alternative to data centers. For the public, the question is whether expanding electricity demand on Earth can eventually be partly addressed by space-based solar power. That remains speculative: the prototype’s one-kilowatt capacity and brief operating periods are far from the scale of commercial computing facilities.
Products Worth Considering
Anker SOLIX C300 Portable Power Station, 288Wh
Jackery Explorer 500 v2 Portable Power Station, 512Wh, 500W
Jackery Explorer 2000 v2 Portable Power Station, 2042Wh
As an affiliate, we earn on qualifying purchases.
What Project Suncatcher Proposes
Google first outlined Project Suncatcher in a research post and preprint on November 4, 2025. The concept is a constellation of satellites carrying TPUs, flying close together and communicating through laser links so they could work as a distributed computing system. Google’s paper, published in a peer-reviewed version in the journal Joule on October 1, 2026, sketches clusters of 81 satellites within roughly one kilometer, at an altitude of about 650 kilometers. Those are research plans, not a deployed system.
The “first AI in orbit” description would be inaccurate without qualification. CNBC reported that startup Starcloud launched a satellite carrying an Nvidia H100 GPU in November 2025 and later ran and queried Google’s Gemma model in orbit. Google’s narrower claim is that its mission produced the first response from a data-center-class TPU in orbit. Its proposed focus on satellites linked to operate together is also distinct from running AI on a single spacecraft.
Google’s research says a bench test of a pair of optical transceivers reached 1.6 terabits per second. That result was a ground test, not a demonstration of a functioning satellite cluster in space. The company’s proposal also faces engineering issues including heat removal in vacuum, radiation damage, communication with Earth, repair difficulties and orbital debris. The source material cites about 47,000 tracked objects in orbit, but does not specify the tracking category or date for that count.
“As the satellite passed by Svalbard, Norway, the downlinked poem confirmed success.”
— Google, in its October 30, 2026 account
Costs and Space Risks Remain
The first test does not show how long the TPU can operate reliably, how performance changes under sustained workloads, or how often the satellite will need to communicate with a ground station. Google has not established that its proposed close-flying constellation can maintain high-speed links while operating in orbit. The reported 1.6-terabit-per-second link result came from a bench test, and the cost estimates depend on future launch prices.
It is also unclear whether the planned 2027 two-satellite mission will meet its schedule, what workloads it will run, or what results Google will disclose. Heat management, radiation exposure, repair and replacement, launch capacity and orbital debris all remain unresolved at the scale proposed. The haiku confirms a successful small demonstration, not the commercial feasibility or environmental impact of orbital data centers.
Tests Before a Two-Satellite Link
Google says it will continue testing the existing satellite, including the TPU’s response to thermal stress and radiation. If those checks proceed as planned, the company’s next stated milestone is a pair of interconnected satellites in 2027. TechCrunch reported that the pair is expected to generate more power and test optical communications between spacecraft.
Those results will help show whether Google can move beyond isolated, short-duration computing to linked orbital hardware. Until then, the project remains an early technology test; the timing, performance and cost of any larger constellation have not been confirmed.
Key Questions
What did Google’s satellite do?
Google says its satellite used an onboard TPU to run the Gemma 4 model and return a haiku after engineers sent it a prompt during a ground-station pass.
Is this the first AI system in orbit?
No. CNBC reported that Starcloud launched a satellite with an Nvidia H100 GPU in 2025 and later ran and queried Google’s Gemma model in orbit. Google’s claim is narrower: its test produced a response from a data-center-class TPU in orbit.
Can this satellite run a full AI data center?
No. The reported prototype produces up to about one kilowatt and can operate one of its four TPUs at a time in short bursts. It is a small-scale test, not a full data center.
Why is Google testing AI computing in space?
Google argues that solar panels in suitable orbits could produce more energy than panels on Earth. Whether that advantage can offset launch, hardware, cooling, communications and replacement costs has not been proven.
What happens next?
Google says it will test the current hardware under heat and radiation conditions. It has identified two interconnected satellites in 2027 as the next milestone, subject to the ongoing tests and development.
Source: ThorstenMeyerAI.com
