Google’s First Data Center Satellite Wrote A Haiku. What It Proves About AI In Space, And What It Doesn’t
AIThis post was created with the assistance of artificial intelligence (AI).

🔍 Read the full analysis: Google’s First Data Center Satellite Wrote A Haiku. What It Proves About AI In Space, And What It Doesn’t on ThorstenMeyerAI.com

Before you orderOffer from Amazon

Get the latest gadgets delivered free with Prime

  • Fast, free delivery on millions of items
  • Prime Video, Amazon Music and more included
  • Member-only deals all year
Start your free Prime trial Free trial for eligible customers · Cancel anytime
As an affiliate, we earn on qualifying purchases.

TL;DR

Google says a satellite carrying its data-center-class TPU ran Gemma 4 and returned a haiku during a pass over Svalbard. The test shows that an AI workload can run on the prototype in orbit, but it does not establish that space-based data centers are practical or economical. Power, heat, radiation, repair and launch costs remain open challenges.

Google says a satellite carrying one of its data-center-class TPU chips has run the Gemma 4 AI model in orbit and returned a haiku during a pass over Svalbard, Norway. The October 30 test is the first reported result from Project Suncatcher, Google’s effort to test whether solar-powered satellites could eventually support AI computing in space; it demonstrates a small-scale operation, not a working orbital data center.

Google said engineers sent the satellite a prompt asking Gemma 4 for a haiku about space. The response was: “Blue marble below / Data streams in silent grace / Starlight feeds my core.” The satellite sent the lines to a ground station as it passed over Svalbard. Google described the test as its “Hello World” moment and said engineers also tried space-themed limericks and asked the payload to report its technical status.

The prototype launched on October 1 aboard a SpaceX rideshare mission, alongside more than 100 other spacecraft, according to the source material. After several weeks of commissioning, Google said it powered up the host computer and TPU trays and ran early machine-learning tests. TechCrunch reported that the Planet-built satellite carries four TPUs and produces up to about 1 kilowatt of power. The available power and thermal limits mean it can operate only one TPU at a time, in short bursts.

Google’s next stated tests will examine how the TPUs handle thermal stress and radiation exposure. The company’s next planned step is a pair of interconnected satellites in 2027, intended to test optical links between spacecraft. Google’s broader research proposal describes clusters of satellites connected by lasers so they could work together; the haiku test did not demonstrate that networked system.

At a glance
updateWhen: The satellite launched on October 1, 20…
The developmentGoogle reports that its Project Suncatcher prototype satellite used an orbital TPU to generate and return a Gemma 4 haiku.
Thorsten Meyer AI
Prototype in orbit

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.

4Google TPUs on board (TechCrunch)
≈1 kWpower the satellite generates, at most
1 at a timeonly one TPU runs, in bursts of about 15–20 minutes
2027next step: two linked satellites
Blue marble belowData streams in silent graceStarlight feeds my core.First output of Gemma 4 on the satellite, as published by Google, 30 Oct 2026
  • 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

Suncatcher prototype≈1 kW
Google nuclear uprates, 20-year deal890,000 kW

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

A Small Test of Orbital AI

The result matters because it moves Google’s space-computing proposal from research toward hardware testing: an AI model generated an answer using a TPU operating on a satellite, and that answer reached a ground station. It is evidence of basic in-orbit AI inference, but not evidence that a satellite constellation can perform large-scale training or compete with Earth-based data centers.

Google’s stated motivation is access to solar power. Its research argues that solar panels in a suitable orbit could produce substantially more energy than comparable panels on Earth. That possibility has attracted interest as electricity demand from data centers rises. The US Department of Energy estimates that data centers used about 4.4% of US electricity in 2023 and could account for 6.7% to 12% by 2028. Those figures describe terrestrial electricity use; they do not show that orbital facilities would reduce costs or emissions.

The economic case remains a major qualification. Google’s paper estimates annual launch costs of about $14,700 per kilowatt at current prices, compared with about $810 per kilowatt if launch costs fall to $200 per kilogram. The latter estimate is close to the power cost for US data centers cited in the source material, but depends on a future launch price that has not been achieved. Google’s paper estimates that deploying the proposed system at scale could require roughly 1,800 Starship launches.

Products Worth Considering

Anker SOLIX C1000 Gen 2 Portable Power Station with 200W Solar Panel

Anker SOLIX C1000 Gen 2 Portable Power Station with 200W Solar Panel

Apple Watch Series 8 GPS 41mm Smartwatch with Midnight Aluminum Case and Midnight Sport Band (S/M)

Apple Watch Series 8 GPS 41mm Smartwatch with Midnight Aluminum Case and Midnight Sport Band (S/M)

As an affiliate, we earn on qualifying purchases.

From Suncatcher Research to Orbit

Google introduced Project Suncatcher in November 2025 in a research post and preprint. The concept involves solar-powered satellites carrying TPUs, flying close together and using laser links to exchange data. Google initially described a learning mission with Planet and said it expected two prototype satellites by early 2027. The single satellite launched in October 2026 is an earlier, smaller test than that paired-satellite milestone.

Google’s claim about being first needs a narrow description. It calls this the first AI response from a data-center-class TPU in orbit, not the first AI chip or AI model in space. According to the source material, Starcloud launched a satellite carrying an Nvidia H100 GPU in November 2025; CNBC later reported that it had run and queried Google’s Gemma model in orbit. Google’s proposed distinction is its plan to connect satellites into a coordinated computing cluster, rather than relying only on individual spacecraft.

The scale envisioned in Google’s paper is far beyond the current prototype: one design sketches clusters of 81 satellites in a radius of about one kilometer at an altitude near 650 kilometers. The paper also reports a bench test of a laser-link pair reaching 1.6 terabits per second. That is a ground test, not a demonstration of the proposed satellite cluster operating together in orbit.

““the first-ever AI response from a data center class TPU in orbit””

— Google

Costs and Operating Limits Remain

The test does not establish how reliably the hardware can run over long periods, whether its output can scale beyond short inference tasks, or how well the chips will withstand radiation and temperature changes. Google has not reported results from those planned stress tests. The source material also does not give a detailed account of the satellite’s full power budget, the duration or number of successful model runs, or the communications constraints for routine operation.

Several engineering and financial questions remain unresolved. A large orbital system would need to manage heat in a vacuum, maintain high-speed links, and operate without routine physical repair. Space debris is another concern: the source material cites about 47,000 tracked objects in orbit, without specifying a date or category breakdown for that count. The launch-cost estimates depend on future prices, and Google has not shown that an orbital system would be cheaper than terrestrial computing after hardware, deployment and operations are included.

The distinction between a working prototype and a commercial service is also important. The haiku confirms a limited end-to-end demonstration, but it does not show a cluster of satellites sharing a workload, sustained data-center-level capacity, or a business case for customers.

Google’s Planned Two-Satellite Test

Google says it will continue testing the current satellite, with checks focused on thermal conditions and radiation exposure. Results from those tests have not yet been reported. The next stated milestone is a pair of interconnected satellites in 2027, which would test optical communications between spacecraft and provide more power than the present prototype, according to the source material.

That mission would address a key gap between the haiku demonstration and Google’s larger concept: whether separate satellites can exchange data fast enough to act as a coordinated computing system. The schedule and final capabilities remain subject to the company’s continuing tests. Until a paired system operates in orbit, Google’s vision of satellite clusters performing data-center workloads remains a research plan rather than a demonstrated service.

Key Questions

What did Google’s satellite do?

Google says it ran Gemma 4 on an orbital TPU and returned a haiku to a ground station while passing over Svalbard. The company also said engineers tested limericks and requested a status report from the payload.

Was this the first AI system in space?

No. The source material says Starcloud launched a satellite with an Nvidia H100 GPU in 2025 and later ran Google’s Gemma model in orbit. Google’s narrower claim is that this was the first reported AI response from a data-center-class TPU in orbit.

Does the test prove orbital data centers are viable?

No. It demonstrates a limited AI task on one satellite. It does not prove that satellites can operate as a linked computing cluster, deliver data-center-scale performance, or do so at a competitive cost.

What is Google testing next?

Google says it plans to test the current TPUs under thermal stress and radiation exposure. Its next stated milestone is a pair of interconnected satellites in 2027 to test optical links between spacecraft.

Why put AI computing in orbit?

Google’s proposal is based on the possibility of stronger solar power in a suitable orbit and the rising electricity needs of terrestrial data centers. Whether those potential advantages outweigh launch costs, heat management, repairs and other operating challenges remains unproven.

Source: ThorstenMeyerAI.com

You May Also Like

Weight-Space Programming: An AI Read, Repaired And Wrote Other Neural Networks Straight From Their Numbers. What It Shows, And What It Doesn’t

An MIT CSAIL team reports that GPT-6 Astra read, edited and wrote small neural networks directly from their weights, with mixed results.

Steam Client Is Now Faster Than Epic’s New Vibe Coded Launcher

A reported test puts Steam’s post-patch cold startup at 5.23 seconds, narrowly ahead of Epic’s preview launcher at about 5.47 seconds.

Analogue 3D’s latest update lets you save whenever you want

Analogue 3D’s latest firmware 1.3.0 introduces a quicksave feature using Memories, allowing players to save at any moment, enhancing gameplay flexibility.

A Simulated AI Company Run Brings Business Tech Into View

The AI Company Emulator replays an emulated AI team running the construction-site app GewerkTon day by day, every decision visible and clearly labelled as a simulation.