Next week a prototype satellite carrying Google's AI chips will ride into low Earth orbit on SpaceX's Transporter-18 mission, built in partnership with Planet.

It's called Project Suncatcher, and the headlines are going to be much bigger than the claim.

So here is the claim, in Google's own words: this is "a long-term, research moonshot." The launch is "about seeing what works, identifying points of failure, and applying those findings to future missions." Google compares it to early research into autonomous driving and quantum computing, which "required years of experimentation before we got to practical systems."

That is a company deflating its own announcement. Note it, because it is rare.

Sunlight.

Not better solar panels - better position. In low Earth orbit a satellite gets near-constant sun: no night, no cloud, no atmosphere absorbing part of the spectrum. Google's figure is that satellites there can generate "up to eight times more solar power than on Earth."

Energy is the binding constraint on AI right now. It is why data centres are being fought over in planning meetings, and why China's electricity generation - more than twice the United States' - shows up in serious analyses of the AI race. A place with eight times the power and nobody living there is, at minimum, an interesting thing to investigate.

Three physical problems, and Google has real results on two of them.

Getting there. A rocket trip to low Earth orbit takes about ten minutes, with sustained acceleration up to 10 g. Individual components like the TPU chips "can experience even greater forces up to 50 to 100 g." The team shook the satellite on all three axes to mimic launch frequencies.

Their write-up of how that went is the most human sentence in the whole document:

Radiation. Outside the atmosphere, cosmic rays and solar events flip bits in electronics - literally changing a 1 to a 0 in a running calculation. So Google took TPUs to a proton beam at UC Davis's Crocker Nuclear Laboratory, ran AI workloads on them while irradiating them, and watched what the bitflips did to the output.

The result: Trillium TPUs "can survive a radiation total ionizing dose greater than what they would receive during a five-year space mission."

That is the single most useful sentence in the announcement, and it is the one the coverage will skip. It isn't a vision. It's a measurement, made in a named facility, that someone else could contest.

Heat. This is the problem nobody anticipates, and it's the hardest.

Chips dump a lot of heat into a small area. On Earth you solve this by moving air - fans, cold aisles, water loops. In space there is no air to move.

So: heat pipes to carry heat away from the chips, radiators to throw it into space as infrared, all tested in a thermal vacuum chamber that simulates both conditions at once. A space data centre isn't a normal data centre in a different place. It's a different machine.

Future satellites would fly in clusters, each carrying dozens of TPUs, and would need to move enormous amounts of data between them. The plan is lasers.

Here's why that's hard. Laser communication in space already exists - but it's built for low bandwidth over long distances. This needs the opposite: very high bandwidth over very short distances, between two objects that are both moving fast.

Google's description of the precision required:

That gets tested in 2027, when two satellites go up together. That's the milestone that matters, and it's more than a year out.

It would be lovely to write that AI's energy problem is moving to orbit and the fights over data centres are ending.

They aren't. Communities are arguing about electricity, water and land right now - France's foreign minister raised "ongoing protests over the construction of data centres" at the UN Security Council last week. Project Suncatcher's next milestone is in 2027, and after that comes the question of whether any of it is economic.

And there is a cost to going up that the announcement doesn't address, which we'd want answered before anyone builds constellations: launch emissions, orbital congestion, and debris in an already busy low Earth orbit. Moving a problem is not always the same as solving it.

None of that is a reason to be sour about the project. It's a reason to keep the timeline honest.

Most weeks, this newsletter covers something worrying. This week we'd like to do the opposite.

If you have a kid who likes physics, or who takes things apart, this is the story to put in front of them - and here's how to frame it.

Every hard problem in it is a job. Vibration and structural testing. Radiation tolerance in semiconductors. Thermal design for vacuum. Precision optics and pointing. Orbital mechanics. These are not abstractions; they are departments, degrees and careers, and they're the kind of work where a specific person solves a specific problem.

The honesty is the lesson. "Tests like this rarely go as planned, so we were pleasantly surprised." "This first launch is about seeing what works, identifying points of failure." That is what real research sounds like - uncertain, incremental, and willing to say so in public. It is a useful contrast with the confident announcements this series more often has to unpick.

And the framing is worth teaching. Notice what Google published: a milestone (2027, two satellites, laser links) rather than a vision (glowing constellations powering the world). A claim with a date attached can be checked. A vision can't. That distinction will serve a young person well for the rest of their life, in science and well outside it.

A prototype goes up next week to answer one question: do these chips survive?

Not "will AI move to space." Just: does the hardware work up there.

That's a good question, asked properly, with a date on the answer. We'll know more in 2027.

Source: Google, "Learn about Google's Project Suncatcher to put ML infrastructure in space," blog.google. https://blog.google/innovation-and-ai/models-and-research/google-research/google-project-suncatcher-facts/

This is a company announcement accompanying a new video series, not independent reporting, and its test results have not been externally verified. Quotations verbatim. The page carries no visible publish date. The point about launch emissions, orbital congestion and debris, the data-centre timing comparison, and the guidance for parents are ours.

Disclosure: Google and Anthropic are competitors, and Google is an investor in Anthropic. This package was drafted using Claude, made by Anthropic.

Want practical AI guidance for parents and educators every week? Subscribe: https://www.aibyage.com/?modal=signup&utm_source=beehiiv&utm_medium=newsletter