On November 19, 2026, Google and Blue Origin announced a launch services agreement for "Project Suncatcher," a plan to put solar-powered AI computers on satellites. Blue Origin's large rocket, New Glenn, will carry out the first launch dedicated to Project Suncatcher, with a plan to place satellites into a sun-synchronous orbit in the first quarter of 2029. Google has described a concept in which each satellite carries computing power equivalent to one data center rack. However, how much processing performance a group of connected satellites can deliver depends not only on the onboard computers but also on the performance of inter-satellite communication.

AD

From a rideshare prototype launch to a dedicated New Glenn flight

A prototype satellite that Google developed jointly with Planet was placed into orbit on October 1, 2026, on SpaceX's Transporter-18 rideshare mission. According to Google's launch report, communication with the satellite was successful and the onboard equipment operated as expected. The new agreement makes the means of transport for the next stage of satellite deployment concrete. The dedicated New Glenn launch aims to reach a sun-synchronous orbit, in which the orbital plane keeps a nearly constant orientation relative to the Sun.

Google's announcement says each satellite will have computing power equivalent to one data center rack. A rack is a piece of equipment that houses multiple servers and other hardware, so the expression is a rough indicator of the computing capability of the onboard machines. Under the plan, Google will put its TPU, a chip it developed for AI processing, on the satellites. However, the TPU model and number of chips have not been disclosed, and the number of satellites and the contract value were not made public at the time of reporting. The evaluation metric used to arrive at the one-rack equivalent, the inter-satellite communication configuration, and the processing performance of the constellation as a whole have also not been released.

One satellite's computing power is not the same as constellation performance

When multiple satellites share a single computation, they must exchange intermediate data with one another. The wait time until the data needed for the next step arrives cannot be shortened simply by adding more chips. The one-rack figure Google presented is only a guide to the computing capability of a single satellite; it does not indicate the processing performance when a group of satellites is connected to run AI. Assessing that performance requires distinguishing between experimental results so far and future test plans.

Result / plan Test location / target Confirmed result / announced goal
AI response experiment reported by Google on October 30, 2026 One prototype satellite in orbit and a ground station Gemma 4 on the satellite processed an instruction, and the ground station received the generated text. Google's report
Optical communication test reported in the arXiv design study Ground-based test setup with a short free-space optical path Measured transmission speeds of 800Gbps in one direction and a combined 1.6Tbps in both directions. arXiv Section 2.1
Inter-satellite connection planned for 2027 Two satellites in orbit Google announced a plan to connect two satellites. Google's announcement
Dedicated launch targeted for Q1 2029 Satellites to be placed into sun-synchronous orbit by New Glenn Launch services agreement signed. Google assumes computing power of one rack per satellite. Blue Origin, Google

In the October 2026 experiment, instructions were sent to the satellite during the window when communication with the ground station was possible. What can be confirmed from Google's report is limited to the fact that the AI on the satellite generated text in response to an instruction and transmitted the result to the ground. No independent replication by third parties has been confirmed. In the two-satellite connection test planned for 2027, the key question, in addition to communication with the ground station, will be whether the satellites can exchange data with each other. Furthermore, evaluating the speed and reliability of sharing a single process among multiple satellites will require not only establishing communication but also measuring computation that uses that communication.

As for the performance required of inter-satellite communication, a design study by nine authors, including Blaise Agüera y Arcas and Travis Beals, gives concrete figures. The research team estimated the bandwidth needed for one inter-satellite link at about 10Tbps in total across both directions. In the ground test, they used a device built from commercially available components and transmitted data over a short free-space optical path, without optical fiber. They measured transmission speeds of 800Gbps in one direction and a combined 1.6Tbps in both directions. However, while wider bandwidth increases the amount of data that can be sent in a given time, the overall computation time is also affected by the delay before the data needed for the next step arrives.

The design study was peer-reviewed and published in the journal Joule. The paper is titled "Toward a future space-based, highly scalable AI infrastructure system design," with the DOI 10.1016/j.joule.2026.102678. The bandwidth estimate and ground test measurements are described in Section 2.1 of the preprint (arXiv:2511.19468v2), revised on June 17, 2026. The same version also explains the design conditions for the satellite formation and the radiation testing procedures.

AD

Conditions for maintaining a tight formation and continuous operation

In inter-satellite optical communication, the greater the distance, the more the light beam spreads and the less light reaches the receiver. For this reason, in Sections 2.1 and 2.2 of the preprint, the research team considers a tight formation in which the satellites stay close together. In the design example, the average altitude is 650km and the formation radius is 1km, with 81 satellites placed in the same orbital plane. Taking into account the effect of Earth's oblateness, they numerically calculated the motion of the satellites over one orbit without applying thrust. Maintaining stable communication in such a formation requires confirming not only the calculated orbital motion but also how long actual satellites can keep their mutual distances.

Continuing AI computation also requires securing sufficient power and dissipating the heat generated by the computers. For the 2029 launch, the only orbital condition the two companies have disclosed is that it is a sun-synchronous orbit. In Section 1.2 of the design study, by contrast, the candidate assumed is a sun-synchronous low Earth orbit that follows the day-night boundary of the Earth. The concept is to generate power from sunlight almost continuously and use heat pipes to carry the computers' heat to radiator panels. The amount of sunlight actually available and the heat rejection capacity of the radiators will determine how long computation can continue on the satellites.

The research team has also run experiments on how space radiation affects semiconductors. According to Sections 2.3 and 4.3 of the preprint, they irradiated a Trillium TPU (v6e) with protons at a ground facility at the University of California, Davis (UC Davis) and also tested the associated servers. The proton energy at beam generation was 67MeV, but because protons lose energy while passing through parts of the equipment, the energy on reaching the chip varies. In tests targeting high-bandwidth memory (HBM), anomalies appeared starting at a cumulative dose of 2krad(Si).

In single-chip tests, the team separately evaluated permanent failures caused by total ionizing dose and temporary errors caused by individual particle strikes. Assuming shielding equivalent to 10mm of aluminum, the estimated dose over five years was calculated at about 750rad(Si). Evaluating stability on the satellites will require examining both permanent failures and temporary errors in light of the actual shielding structure and operating period.

In future in-orbit tests, in addition to the bandwidth and latency achievable between satellites, how long the inter-satellite distances needed for communication can be maintained will be important. By having multiple satellites share a single computation under those conditions and measuring processing time and stability over repeated runs, it will become possible to evaluate more concretely the concept of connecting many satellites, each with rack-equivalent computing power.