Google's Project Suncatcher satellite just did something no production system has done before: it ran an AI model in orbit and sent the answer home. The workload was a haiku — "Blue marble below / Data streams in silent grace / Starlight feeds my core" — generated by a Gemma 4 model on a TPU aboard a prototype satellite built with Planet and launched on SpaceX's Transporter-18 rideshare. CEO Sundar Pichai called it a "hello world" moment. For developers, the interesting part is not the poem. It is the constraint envelope the demo reveals.
The constraint envelope, stated plainly
The satellite carries four TPUs but generates only 1 kilowatt of power. That means one chip at a time, for about 20 minutes, timed around ground-station communication windows. Cooling is by heat pipes and radiators — there is no air, so there are no fans. Radiation can flip bits. Google's ground testing at UC Davis exposed Trillium TPUs to a proton beam exceeding a five-year mission's radiation dose, and the early signal is that inference tolerates radiation better than training does, because a flipped bit in inference corrupts one output, not a weight reused millions of times.
If you build for edge or embedded systems, this envelope will feel familiar. It is the same discipline — tiny power budget, harsh environment, intermittent connectivity — pushed to its extreme. The design lessons transfer directly:
- Inference first, training never (up there). The plausible orbital workload is inference at the data source: classify the imagery on the satellite, downlink the answer, not the terabytes. If your pipeline assumes raw data always comes home for processing, rethink it.
- Design for intermittent compute windows. Twenty minutes of compute, then darkness — literally and figuratively. Batch your work, checkpoint aggressively, and make every session resumable. This is the same pattern as a phone doing on-device ML between charges.
- Treat bit flips as a data-quality problem. Radiation-induced errors are not crashes; they are silent corruption. Checksums, redundant inference, and confidence thresholds are not optional extras in this environment — they are the product.
- Thermal design is compute design. When cooling is passive, your sustained throughput is a thermal budget, not a clock speed. Profile under thermal limits, not just on a bench.
Why this matters even if you never touch a satellite
The orbital case is extreme, but the direction is general: compute is moving toward the data. Phones run models on-device. Factories run inference on the line. Satellites are the logical endpoint — the sensor is in orbit, so the model goes to orbit too. Every time the industry proves inference works in a harsher environment, the on-device and edge playbook gets stronger.
Google plans two more Suncatcher satellites in 2027 with greater power and laser links so satellites can share workloads. Watch that second part: satellite-to-satellite workload sharing is distributed systems engineering under orbital mechanics — latency, link windows, and failure modes that make terrestrial microservices look forgiving. If you enjoy hard distributed-systems problems, this is where they are heading.
The honest caveat
The economics are unproven. Researchers estimate around 1,800 Starship flights would be needed to make orbital data centres viable in the 2030s. For most workloads, Earth still wins on cost. Do not pitch your manager an orbital migration. Do steal the discipline: power budgets, silent-corruption handling, and compute-at-the-source architecture. Those pay off on Earth today.
The haiku is forgettable. The loop closing — command up, inference in orbit, answer down — is the thing to remember. Every platform you build on started with a demo exactly this small.
Sources: BusinessToday report on Google's Project Suncatcher demonstration, November 4, 2026. Figures as reported; orbital data centre economics remain unproven.
