AI generates haiku in orbit

Engineers from Google asked the Gemma 4 AI model, installed in an experimental satellite, to generate a haiku. They retrieved the response when the satellite passed over Svalbard, Norway.

Satellite launched by SpaceX

The satellite, about the size of a refrigerator, was launched on October 1st by a Falcon 9 rocket from SpaceX. It is the first prototype of Suncatcher, a project Google unveiled in November 2025 for orbital data centers.

AI chips operate in short bursts

Google installed four of its AI TPU chips in a satellite provided by Planet, a space imaging company. To prevent overheating, engineers only run these chips for fifteen consecutive minutes at a time.

Haiku marks a milestone

Travis Beals, director of the Paradigms of Intelligence research team at Google, called the haiku "the first AI response produced in orbit by a data center chip." The haiku describes the blue planet below, silent data flowing, and a computing heart fed by the light of stars, with a spark emoji as a signature.

AI also provided technical reports

The model then composed humorous short poems about space and provided a technical report on the status of onboard systems upon request from the engineers.

Cost of AI response in orbit

At the rate of Gemma 4 on Google Cloud, the twenty tokens of this haiku would cost approximately $0.000012, according to calculations from the British site The Register. However, the launch cost is around $7,000 per kilogram, and the satellite is estimated to weigh about 500 kilograms.

Testing AI chips in orbit

Google will now test the resistance of its TPU chips to heat and radiation in orbit. On the ground, researchers exposed a Trillium chip and its AMD host server to a 67 MeV proton beam.

Radiation effects on hardware

They recorded the first anomalies in the high-bandwidth memory after a cumulative dose of 2 krad(Si), which is nearly three times the minimum threshold of 750 rad(Si) that Google requires of its components.

Error rate in orbit

With an estimated dose of 150 rad(Si) per year in orbit and one request per second, researchers predict approximately one response out of 3 million would be tainted with erroneous calculations without an alert, according to an article published this month in the journal Joule.

Error rate deemed acceptable

"This error rate is probably acceptable for inference," they write, but they still lack data on the effects of these errors during a model's training.

Satellite powered by sunlight

On its heliosynchronous orbit, the satellite spends almost all of its time in sunlight, and its engineers did not need to equip it with heavy batteries. Despite this, it only has about 1 kilowatt of electricity for a mission planned to last a year.

Profitability depends on launch costs

To make Suncatcher profitable, the price of rockets will also need to decrease. In a study published in November 2025, Google researchers calculated that with a launch cost of $200 per kilogram, sending a satellite would cost between $810 and $7,500 per kilowatt and per year.

Comparing orbital and terrestrial costs

In the United States, electricity for a terrestrial data center costs between $570 and $3,000 per kilowatt and per year. Google estimates that SpaceX could offer this rate around 2035, and prefers to build its prototypes long before that.

Future satellite plans

Until then, the company plans to send two satellites connected to each other in 2027, still with Planet. With this pair, engineers will test optical links between satellites to verify that chips distributed across multiple satellites can share the same machine learning computation.

Understanding TPU chips

A TPU is a hardware accelerator designed by Google to efficiently execute machine learning calculations, particularly the matrix operations used by neural networks. Compared to general-purpose CPUs, it aims for better computational throughput and energy efficiency for inference, and according to the generations, for training as well.

Testing in orbit

Testing it in orbit allows verifying if a "data center" chip can function reliably outside of its normal operating conditions (radiation, thermal constraints, power cycles). The challenge is to validate that the hardware, but also the surrounding platform (host server, memory, communication), remains stable and predictable in space.

Measuring radiation exposure

rad(Si) and krad(Si) measure an energy dose deposited by radiation in silicon, which helps quantify the exposure of an electronic component. In orbit, these radiations can cause "soft errors" (bit flips) or more lasting disturbances, without necessarily immediately damaging the chip.

Importance of high bandwidth memory

The HBM (High Bandwidth Memory), very fast and very close to the processor, is critical because it stores data and activations necessary for AI calculations; a corruption can falsify a result. The key point is that some errors can be silent (no alert), which complicates detection and correction, especially when models require reproducible calculations.

Inference vs. training

Inference consists of using an already trained model to produce a response (classification, text generation, etc.), while training adjusts the model's parameters from examples, over very long computation sessions. In inference, a single error may sometimes only affect an isolated output, and application mechanisms can tolerate or filter out aberrant responses.

Uncertainty in training

In training, repeated or undetected errors can pollute the weight updates, degrade convergence, or introduce biases that are difficult to diagnose. This is why the same hardware reliability can be judged "acceptable" for inference, while remaining uncertain for training in production.