Google's orbital AI needs scale: calculations assume 1,800 Starship launches

Project Suncatcher explores the possibility of running AI workloads on satellites equipped with solar panels. In a study dated Jun 17, 2026, Google calculated that, in a suitable orbit, the panels receive up to 8 times more energy per year than on Earth at mid-latitudes.

From modeling to testing. In November 2025, Google modeled a cluster of 81 satellites at an altitude of 650 km. The satellites occupied an area with a radius of 1 km, with nearest neighbors spaced 100–200 m apart. The company is now testing hardware in space.

To enable the chips to work together, Google is developing laser links between satellites. In November 2025, a ground demonstrator transmitted 800 Gbps in each direction, for a total of 1.6 Tbps over a single pair of transmitters and receivers. The company plans to equip future satellites with dozens of TPUs each.

Lower costs require launches. In Google's scenario dated Jun 17, 2026, bringing launch costs below $200/kg by 2035 requires sending another 370,000 tonnes into orbit. That amounts to about 1,800 Starship launches carrying 200 tonnes each, averaging 180 launches per year. The model assumes that the price per kilogram falls by roughly 20% each time the total mass launched doubles.

At $200/kg, Google compares the cost of delivering computing capacity to orbit with electricity costs on Earth, excluding infrastructure, buildings, and chips from the calculation.

Original source: [Google on Project Suncatcher and TPU testing in space](https://blog.google/innovation-and-ai/models-and-research/google-research/google-project-suncatcher-facts/).

In 2027, Google plans to launch 2 satellites to test laser links between them.