Google's first data center satellite goes online, writes haiku in orbit using Gemma 4

The satellite carries a Google Tensor Processing Unit (TPU) AI accelerator chip and can generate up to about 1 kilowatt of power. The project's goal is to pave the way for a future "orbital data center network" composed of multiple satellites flying in close coordination, enabling large-scale AI inference tasks to be completed directly in space.

However, current technology still faces many challenges. Because the TPU has relatively high power consumption, the satellite can only power up the chip for about 20 minutes at a time, which makes the already complex space-to-ground communications work even more difficult. After about three weeks of in-orbit testing and system calibration, the satellite finally entered its formal operational phase.

To verify the system's functionality, Google engineers this week sent a prompt instruction to the satellite for the first time, asking the Google DeepMind Gemma 4 model pre-installed on the satellite to compose a haiku about space. The model's work roughly reads: "The blue planet shines quietly below, data streams flow safely through, starlight nourishes my core."

In addition to having AI write poetry in space, Google also used the model to collect and report information about the orbital environment around the satellite. A Google spokesperson said the satellite is currently operating normally and all systems are working as expected.

Next, Google will focus on evaluating the reliability of the TPU chip under extreme space conditions, including thermal stress caused by temperature changes and the impact of cosmic radiation on the hardware. Although Google has previously conducted extensive tests on TPUs in ground-based particle accelerators, long-term performance in the real space environment still needs further verification.

One of the most closely watched issues is radiation-induced "bit flips," in which data bits in computer memory change unexpectedly under the action of high-energy particles. This can cause data corruption, and conventional monitoring systems may not always detect it in time.

Previous ground experiments showed that radiation's impact on AI inference tasks is relatively limited, but it may cause more obvious problems during model training. Now, Google can finally verify these speculations through actual in-orbit operating data, rather than relying solely on laboratory simulation results.

If this mission proves that TPUs can operate stably in space, Google plans to launch two more follow-up satellites in 2027 using more advanced chips. The new-generation satellites will have greater power generation capacity and demonstrate optical communication link technology between satellites.

Google's envisioned space data center approach differs from other projects in the industry. Its core idea is to establish a high-speed interconnect network among in-orbit satellites so that multiple satellites can collaboratively process the same workload, thereby forming an orbital computing platform with distributed computing capabilities. By contrast, although some competing proposals also plan to use laser communications, the distances between satellites are relatively large, making efficient collaborative computing difficult to achieve.

As artificial intelligence's demand for computing power and energy continues to grow, Google is trying to extend data centers from the ground into space. The successful operation of the Suncatcher satellite not only validates the feasibility of the relevant concepts but also marks a key step toward building a true orbital AI computing infrastructure in the future.