Four chips. One kilowatt. That is the entire payload of an experiment that could decide where AI infrastructure gets built next.
On October 1, a SpaceX Falcon 9 rideshare is scheduled to lift a refrigerator-sized satellite into low Earth orbit (LEO). Inside sit four Trillium-generation Tensor Processing Units (TPUs) — the same class of chip that trains and serves the company's models in terrestrial data centers. The mission belongs to Project Suncatcher, a research program Google introduced in late 2025. This is its first flight.
The pitch is physics. Near-constant sunlight in LEO yields up to eight times the solar power of a comparable ground site, and the vacuum removes the water-cooling bottleneck that stalls terrestrial builds.
The threat is not the rocket. Radiation, vibration and heat are the three failure modes ground tests can only partly reproduce.
What is actually going into orbit
The satellite is named MVP. It was built with Planet, the Earth-imaging company, and rides on Transporter-18, one of SpaceX's recurring rideshare flights from Vandenberg Space Force Base. Its solar arrays generate roughly one kilowatt — enough to run a domestic fridge, not a data hall.
AI accelerators on the MVP satellite
Four Trillium-generation chips are the mission's entire compute payload. · Google Research, 2026
Google's framing is deliberately small. The company calls Suncatcher a moonshot, alongside its autonomous-vehicle and quantum programs. Those ran for years before anything shipped. The satellite is a first data point, not a product.
The three ways a chip dies in orbit
Running AI silicon in space is a reliability problem before it is an economics problem. Three forces do the damage, and only one of them is the rocket.
Radiation. High-energy particles flip bits and degrade transistors. The company reports that its TPUs survived ground-based radiation testing at doses above what the hardware is expected to absorb over a five-year mission. Ground chambers, though, approximate an environment nobody can fully simulate.
Vibration. A Falcon 9 shakes its payload harder than most industrial equipment will ever see. Chips that pass on a bench can still crack at the solder joint.
Heat. There is no air to carry it away, so every watt of compute becomes waste heat that must be radiated into the void.
How do you cool a chip with no air?
Why orbit, and why now
The case for orbital compute rests on a single asymmetry. A satellite in the right low Earth orbit sees the sun almost continuously, with no clouds, no night and no atmosphere to filter it.
The orbital solar advantage
The company estimates near-constant sunlight in orbit yields up to eight times the solar power of an equivalent ground site. · Google Research, 2026
That matters because power, not silicon, is the binding constraint on AI. Grid interconnection queues in the United States can stretch for years, and data centers are competing with households for water and electricity. The company's own framing is blunt: can AI hardware work where the power is free and the cooling is a vacuum?
Google is not first to the idea. As we wrote in September, Starcloud raised at a $2.3 billion valuation to move data centers into orbit; Loft Orbital and Marlan Space have staked roughly $1 billion on AI that runs in space. Elon Musk and Jeff Bezos have both pitched orbit as the next home for compute. What it adds is a hyperscaler's chip lab and a willingness to name the failure modes out loud.
Turning that idea into reality starts with a basic question: Can our AI hardware operate in space?— Google Research, Project Suncatcher, September 2026
| Constraint | Ground data center | Orbital node |
|---|---|---|
| Power | ◐ Grid queues, rising prices | ✔ Near-constant sunlight |
| Cooling | ◐ Water and power hungry | ✔ Passive radiative cooling |
| Repair | ✔ Swap a part same day | ✗ No crew, no service call |
| Mass | ✔ Unconstrained | ✗ Every kilogram is launch cost |
Comparison of AI compute environments · Google Research, 2026
The economics are still unproven
One kilowatt does not make a business. Google's satellite carries the equivalent of a single server, and the company describes it as a test rather than a deployment. A commercial orbital data center would need gigawatts of compute, radiators measured in hectares, and a maintenance model for hardware no crew can reach.
Launch is the first filter. Every kilogram of chip, radiator and solar panel is paid for twice — once on the ground, once to orbit. Reusable rockets have cut the price of that second payment, but the mass penalty for cooling and radiation shielding has not moved. On the ground, a failed chip is a service ticket. In orbit, it is dead mass.
Latency narrows the market further. An orbital node is a poor host for the interactive inference that dominates today's AI revenue. It is a better fit for batch training and for workloads generated in space itself — Earth observation, maritime tracking, defense sensing — where the data is already up there.
What is the strongest argument against orbital data centers?
Does orbital AI compute become a real market by 2030?
Probability: 25% — the physics is real, but launch mass, radiation hardening and repair economics are unsolved, and no hyperscaler has committed capital beyond research.
✅ Arguments for
Google, Starcloud and Loft Orbital are all funding hardware, which turns a thought experiment into a procurement category.
Confirmation criteria: a second Suncatcher mission with laser inter-satellite links, and a paying customer outside government research.
❌ Arguments against
Launch costs are falling, but not as fast as terrestrial efficiency is improving.
Disconfirmation criteria: MVP returns radiation or thermal data worse than ground testing predicted, or the company declines to fund the 2027 milestone.
Development scenarios
🟢 Optimistic scenario (25%)
Implications: orbital compute becomes a real, if narrow, infrastructure category.
🟡 Base-case scenario (55%)
Implications: orbital data centers remain a venture narrative, not an asset class, through the decade.
🔴 Pessimistic scenario (20%)
Implications: capital rotates back to terrestrial efficiency, and orbital compute is written off as a 2020s detour.
Whether the MVP satellite returns radiation and thermal data as clean as ground testing predicted
The stated next Suncatcher milestone in 2027
Deployment milestones from Starcloud and Loft Orbital
Launch price per kilogram breaking below $1,000