On October 30, 2026, Google announced that its prototype satellite for Project Suncatcher, its concept for space-based data centers, had succeeded in having the AI model Gemma 4 generate text, and that the response had been received on the ground. The computation used TPUs (Tensor Processing Units), which Google developed for AI workloads. After the October 1 launch, communication with the satellite and the operation of its onboard equipment had already been confirmed. This time, Google demonstrated the full loop: an instruction sent from the ground was processed by AI in orbit, and the generated result was sent back to Earth. The idea of running AI in space has taken a step forward, but many technical challenges remain, including long-term stable operation and distributed processing across multiple satellites.
A TPU in space generates text, received in Svalbard
According to Google's official announcement, the prototype satellite successfully booted its host computer and TPU tray, went through initial machine learning tests, and then responded to instructions from the ground. In the experiment, during a window when the satellite could communicate with a ground station, Gemma 4 was asked to generate a haiku-style poem in English on the theme of space. The text generated in orbit was transmitted to Earth and received at a ground station near Svalbard, Norway.
The result was a three-line poem depicting the blue Earth spread out below, quietly flowing data, and starlight. Google also tried other tasks, including generating a five-line poem about space and prompting the system to report the status of its onboard equipment. These experiments confirmed that AI in orbit can process natural-language instructions and generate text. However, they did not demonstrate how accurately the system can diagnose its own equipment autonomously, nor its operational safety.
Google positions this result as the first time AI responses have been returned from orbit using data-center-class TPUs. That "first" applies specifically to response generation using data-center-class TPUs, not to every instance of AI running in space. The key advance is that, rather than sending data acquired by a satellite to the ground for processing on terrestrial computers, the satellite ran AI inference on its onboard TPU and sent only the result back to Earth.
From a first response to long-term operation: the demonstration enters its next phase
On October 1, Google announced that a prototype satellite developed jointly with Planet had been launched on SpaceX's Transporter-18 mission, and that communications and onboard equipment were confirmed to be working normally. The October 30 report goes further, showing that text generated by an AI model in orbit could be received on the ground. Still, communicating with a satellite, running AI inference on a single satellite, operating stably over the long term, and linking multiple satellites are each distinct technical challenges. This result does not mean all of them have been solved.
Based on Google's latest announcement, its 2026 technical explainer, and its 2025 design study, the demonstrated results and future plans break down as follows.
| Verification stage | Status | Confirmed results / future verification targets | Remaining challenges |
|---|---|---|---|
| Launch and communications check | Completed in orbit on October 1, 2026 | Communications with the prototype satellite and normal operation of onboard equipment confirmed | AI inference in orbit and retrieval of responses |
| First AI response | Announced October 30; completed in orbit | Gemma 4 generated text in response to an instruction, received on the ground | Evaluation of long-term reliability and processing performance |
| Continued operation and inter-satellite link | Thermal and radiation tolerance under continued verification in orbit; two-satellite link planned for 2027 | Ground tests conducted on cooling, radiation tolerance and optical communications | Long-term stable operation in orbit and coordination among multiple satellites |
The successful generation and receipt of the poem supports the idea that the full chain works: processing an instruction from the ground in orbit and returning the result. However, realizing a space data center requires running computers stably over long periods and sharing computing resources across multiple satellites. This experiment is an early-stage result on the way to such a large-scale system, and long-term reliability and distributed processing performance will have to wait for future demonstrations.
Space is cold, but cooling TPUs is not easy
Space is extremely cold, but that does not mean the heat generated by TPUs can be shed easily. According to Google's technical explainer, high-performance TPUs generate a large amount of heat from a small area, and because space is nearly a vacuum, airflow-based cooling like that used on the ground is not possible. Google is considering approaches such as combining heat pipes that move heat with radiator panels that dump it into space through thermal radiation, and is also conducting ground tests using a thermal vacuum chamber.
In Google's concept, choosing a suitable orbit could allow the satellites to receive near-continuous sunlight for power generation. Being able to use solar power in an environment unlike that of terrestrial data centers is one advantage of space data centers. But even if enough power can be drawn from sunlight, stable operation would be difficult unless the heat produced by the TPUs' computation can be properly expelled. Power supply and cooling are separate problems, and each must be solved on its own.
Radiation tolerance is another key area of verification. At the Crocker Nuclear Laboratory at the University of California, Davis, Google irradiated a TPU with a proton beam while it ran AI workloads, monitoring for anomalies such as bit flips. The company says that in this test the Trillium TPU withstood radiation exceeding the total ionizing dose expected over a five-year space mission. However, the result depends on test conditions and the components evaluated, and it does not guarantee stable long-term operation in space.
Google's 2025 design study reported ground tests in which Trillium (v6e Cloud TPU) was irradiated with a 67 MeV proton beam. In those tests, high-bandwidth memory (HBM) was the most susceptible component, with anomalies observed once the cumulative dose reached 2 krad(Si). By comparison, the cumulative dose expected over a five-year mission, accounting for shielding, was put at 750 rad(Si). A single chip also showed no permanent failures attributable to total ionizing dose up to 15 krad(Si).
That said, anomalies occurring in HBM are a different phenomenon from permanent failure of the chip itself. The absence of permanent failures does not mean transient computation errors or data corruption will not occur. These are results from evaluating Trillium at ground-based test facilities, not a record of the current prototype satellite operating normally in space for five years. Google plans to keep verifying how thermal stress and radiation exposure in orbit affect TPU operation.
Two satellites to be linked in 2027, aiming to demonstrate distributed processing
As its next development stage, Google plans to interconnect two satellites in 2027. In the longer term, the concept is to equip each satellite with dozens of TPUs and link the satellites with high-speed communications, building a large compute cluster in space. However, what has been demonstrated so far is AI inference on a single satellite; distributed computing that splits work across multiple satellites has not yet been achieved.
To operate multiple satellites as a single computing system, large volumes of data must be exchanged between them at high speed. Google cites laser communications as a key technical challenge, one that must provide extremely high bandwidth over short distances, unlike conventional long-distance, low-bandwidth space communications. Maintaining stable high-speed optical links also requires precisely tracking the relative positions of the satellites and properly controlling the pointing of the communications equipment. Google's design study also models the orbital dynamics needed to keep multiple satellites in a tight formation.
Google's research concludes that matching the computing performance of terrestrial data centers would require inter-satellite bandwidth on the order of tens of terabits per second. In a ground experiment using a single transceiver pair, the company says it demonstrated 800 Gbps in each direction, for a bidirectional total of 1.6 Tbps. However, the former is a performance level required of future inter-satellite links, while the latter is a result achieved with ground test equipment. Whether satellites moving through space can sustain similar high-speed communications stably remains a major open question for future testing.
This result is a progress report from Google on its mission. No peer-reviewed comparative study of AI inference performance or independent third-party measurements have been presented. Specific performance metrics such as response time, power consumption and text generation speed have not been disclosed, and error rates over long-term operation remain unknown. Attention now turns to whether TPU heat can be properly expelled in orbit, how far radiation-induced errors can be suppressed, and whether high-speed links between satellites can be kept stable. The results of the planned two-satellite link in 2027 and of continued operational testing will shape the feasibility of the concept of turning abundant solar energy in space into large-scale AI computing power.