Google put its first TPU-carrying satellite into orbit on 1 October 2026 and published a peer-reviewed cost analysis of orbital data centres the same day. The conclusion is sober: the economics only work if SpaceX flies Starship roughly 1,800 times in a decade.
Key takeaways
- The first TPU satellite, built by Planet, launched on 1 October 2026
- Target scale: ~370,000 tons of payload, or 1,800 Starship launches in a decade (180 a year)
- Break-even threshold: launch costs below $200 per kilogram by the mid-2030s
- The constellation modelled in the study has 81 satellites at roughly 650 km
- Cloud TPU Trillium v6e kept working up to 15 krad(Si) in radiation tests
The first TPU beyond the atmosphere
Planet built the satellite, whose payload is a single TPU — an engineering test, not a constellation. Google calls it the first flight of Project Suncatcher, its research into TPU-carrying satellites linked by free-space optics. Travis Beals, who leads it, puts it plainly: no ground test replaces a real flight.
The sum that has to add up
The Joule analysis, reported by TechCrunch, finds orbital data centres reach cost parity with terrestrial ones only below $200 per kilogram of launch cost. Google places that threshold in the mid-2030s, assuming a 20 percent annual decline.
Symbol meaning
- …
- total payload mass needed at target scale
- …
- payload of one Starship mission assumed in the model
- …
- launches across the full decade
- …
- launches per year if spread evenly
Satellites last five years, so the fleet needs replacing. This is not a one-off cost but a standing launch cadence.
| Parameter | Value in the model |
|---|---|
| Payload per mission | 200 tons |
| Total demand | ≈ 370,000 tons |
| Launch count | ≈ 1,800 per decade (180 a year) |
| Break-even threshold | < $200/kg |
| Assumed launch-cost decline | 20% a year |
| Satellite lifetime | 5 years |
| Modelled constellation | 81 satellites at ~650 km |
Radiation was survivable, the links are the bottleneck
Radiation proved a smaller problem than expected. Trillium v6e absorbed nearly twenty times the dose projected for a five-year mission, with testing run to 15 krad(Si)?krad(Si): A unit of radiation dose absorbed by silicon. It measures how much irradiation a chip takes before it starts failing..
Communication is harder. To behave as one cluster, the constellation needs tens of terabits per second between satellites flying hundreds of metres apart. Bench testing reached 1.6 Tbps total via dense wavelength-division multiplexing?DWDM: Dense wavelength-division multiplexing — many independent data streams sent down one optical path, each on a different wavelength. and spatial multiplexing?Spatial multiplexing: Sending several data streams in parallel over physically separate paths, for example multiple beams at once..
Why it matters
Suncatcher shifts the argument about AI limits from chips to logistics. The bottleneck is not TPU throughput or radiation but the number of rockets that can fly in a decade and how fast launch costs fall. That makes the whole idea hostage to one supplier and one vehicle. By publishing numbers instead of visions, Google lends the programme credibility while showing how far it sits from break-even.
What's next?
- Two prototype satellites built with Planet are due in orbit by early 2027, validating optical inter-satellite links and TPU operation
- Without launch costs below $200 per kilogram by the mid-2030s, the programme's economics do not close
