Core argument
Project Suncatcher is not about faster AI — it is about real estate. Compute is running out of places to live on the ground, and orbit is starting to look like an option.
Counter-intuitive point
Radiation and launch shock — the dangers everyone names first — all passed their ground tests before launch. What actually limits the satellite is far more ordinary. See the "15-minute ceiling" branch for the mechanism and why a vacuum makes it expensive.
What to do next
Run the "Milestones to watch" branch before the next headline: three checks that tell a real engineering step apart from a press-release satellite.
Connections
A ground-level sibling: "你口袋里的钱，被存储芯片抽干了" shows AI emptying the memory supply chain on Earth — while compute goes looking for room in orbit.
The Hardest Problem in Space AI Is Heat, Not Radiation
A fridge-sized satellite carrying four TPUs
Launched 2026-10-01 on SpaceX Transporter-18 from Vandenberg — a rideshare carrying 130 payloads on one Falcon 9
Built by Planet, chips by Google; Google confirmed contact the same evening: "operating as expected"
Payload: 4 Trillium TPUs — roughly the compute of one terrestrial data-center server (WION)
Fridge-sized spacecraft, 1 kW solar array, dawn-dusk sun-synchronous orbit (near-constant sunlight)
Mission life 1 year (up to 6 before reentry); Google calls it "MVP"; project lead Travis Beals: "a very minimal test"
What it is NOT: not an orbital data center — an experiment to gather data on chips in space
The dangers everyone feared — all passed on the ground
Radiation: Trillium TPUs survived a proton-beam dose greater than a 5-year mission — at UC Davis Crocker lab, while running AI workloads, with no bit-flip failures
Launch shock: whole satellite tested to 10 g; individual TPU components to 50–100 g
Thermal: heat pipes and radiators tested in a thermal-vacuum chamber on the ground
Counter-intuitive: radiation and vibration — the two dangers everyone assumes would kill space AI — were cleared before launch day
The 15-minute ceiling
The chips can run 15 minutes, then must shut down to cool (reported by NPR and Scientific American; The New York Times quoting Beals)
Why: near-vacuum means no air and no convection — heat can only leave by radiating through panels
The prototype's radiator is mass-constrained; this is a limit of the test hardware, not of physics
Orbit sunlight is up to 8x as productive as ground solar (Google's figure) — but collecting sunlight is not the same as cheap compute
Counter-intuitive: the "free solar + natural cooling" intuition is half wrong — solar really is cheap up there, but cooling is the most expensive line item in a vacuum
Skeptics' note: raw sunlight doesn't become cheaper tokens until you price in launch cost, hardened parts, and no repair crew (tech-insider)
What Google actually wants: room, not speed
Ground data centers hit three walls at once: grid power, siting, community opposition — Stuff.tv notes it became a US campaign issue
Beals: "in the really-long run, perhaps the set of things you do in space gets larger and larger until that's where most of the compute is happening"
Cold footnote: bytevyte's analysis — a 4-chip rideshare cannot answer the economics; treat it as a long bet, not near-term capacity
2027: two more satellites to test high-bandwidth laser links — the precondition for multi-satellite compute
Long-term sketch: 81 satellites in a 1 km laser-linked array; Google's costing assumes 1,800 Starship flights over 10 years at 200 tons each (Chinese media reporting of Google's figures)
The race is already on
Starcloud has already flown Nvidia H100 chips in orbit for testing
SpaceX plans Starmind AI1 satellites with Nvidia chips in 2027 (Chinese media)
Two tracks now: Google's TPU route vs the H100 route — whoever scales first sets the comparison point
How to read the next headline
Checklist — before the next "AI data center in space" headline, ask three things
Did the 2027 laser-link test work? Multi-satellite compute depends on it
Is Starship actually flying 180 times a year? The 1,800-flight budget is the whole bet
What did the real orbit do to the chips? The MVP year's radiation + thermal data is the physics verdict
Decision rule: one prototype launch is a physics answer; a second launch with laser links is an engineering answer; only recurring launches are an economics answer
Cost-structure lens: ground = cheap cooling, expensive power, scarce land; orbit = cheap sunlight, expensive cooling, expensive launch — whoever moves first on the binding constraint wins