Four Chips Up. Eighty-One Satellites Next

Google put four AI chips in space this week.

Not a press stunt. Not a concept slide. A satellite built with Planet Labs, loaded with four Trillium TPU chips, launched on SpaceX's Transporter-18 on October 1.

Google calls it Project Suncatcher.

And it could reshape where the world runs its biggest AI models.

Let me explain.

The satellite sits in a dawn-dusk orbit. Solar panels feed it about one kilowatt of power. The chips run short Gemini AI queries in 15-minute bursts… then shut down to cool.

That sounds tiny. It is.

But Google doesn't care about four chips right now. Google cares about eighty-one satellites later.

The end goal is huge. Clusters of 81 spacecraft, each packed with dozens of TPUs. All linked by lasers at tens of terabits per second. All fed by sunlight.

In other words, a data center with no land, no power grid, and no water bill.

Why go to orbit? Power.

A mid-sized data center on Earth eats enough power for 16,500 homes. Google, Microsoft, and Amazon are all fighting for more. They're buying nuclear plants. They're locking in utility deals years out.

And they're still short.

Google says satellites in low orbit capture up to eight times more solar energy than panels on the ground. No clouds. No night. No seasons… in the right orbit.

So the pitch writes itself. Vast power. Vast space. Zero land fights.

Now, I know what you're thinking. “This sounds crazy.”

Maybe it is. The cooling problem alone is brutal. There's no air in space to push across a hot chip. Google built a custom heat-pipe and radiator system just for this test. Whether it works at scale… we'll see.

And the costs are steep. Rockets aren't free. Building 81 satellites isn't cheap. Swapping them out every few years adds up.

But here's the key.

Google isn't a startup raising cash and hoping for the best. Google is a $4 trillion company. It makes its own chips. It runs its own AI models. It has a desperate need for more compute.

They're not pitching VCs. They're solving their own problem.

We've covered Starcloud's $250 million raise at a $2.3 billion value. We've seen SpaceX build Starmind with NVIDIA chips. Both are bold bets. But neither has Google's chip supply, cash pile, or raw AI demand.

That's what you should watch. The company with the deepest pockets… and the biggest appetite… just put hardware in orbit.

Four chips today. Eighty-one satellites next.

It's early. But it's exciting.

It cost $150 billion and took 16 nations 13 years to build.

It is the most expensive object human beings have ever made.

Now SpaceX has been hired for what one outlet calls “the most expensive demolition job ever commissioned.”

But here is the part almost nobody has caught yet.

Elon is NOT building the replacement.

That job is poised to go to a company less than half a percent the size of SpaceX.

It is the real winner in this deal.

Former Wall Street CEO Dylan Jovine has laid out the whole story, including the tiny firm’s name.

This flight isn't about doing real work — it's about survival. The prototype, built with Planet, will spend up to a year gathering data on one question: can Google's TPUs take the beating of space? Launch shakes components at up to 50 to 100 times the force of gravity, and orbit bombards them with radiation and wild temperature swings. The four chips run Gemini queries in short bursts, then cool. Succeed, and Google has proof its everyday AI hardware works up there. Watch the radiation and thermal data — that's the real payload.

Google didn't build the satellite — Planet did. The Earth-imaging company, which already operates one of the largest small-satellite fleets, built the refrigerator-sized spacecraft that carries the TPUs, and is slated to build the two-satellite pair that will test laser links in 2027. That's a telling role: as hyperscalers chase orbital compute, the firms that already know how to mass-produce and fly satellites become the hidden picks-and-shovels layer. Planet gets a new, deep-pocketed customer and a reason to keep its factory busy. Watch whether other cloud giants hire satellite builders the same way.

Here's the twist. The rocket that carried Google's compute test belongs to SpaceX — which is building its own orbital-compute play, Starmind, with Nvidia chips. Alphabet also holds a SpaceX stake reportedly worth tens of billions. So the same launch advances Google's experiment and funds a competitor. That tangle is the shape of the whole space-compute race: everyone needs SpaceX's cheap rides, even rivals. It's a reminder of how much leverage the dominant launcher holds over the field. Watch whether compute hopefuls stay dependent on a competitor to reach orbit.

Watch for the two-satellite laser demo in 2027.

Here's why that's the real hurdle. Today's flight tests whether chips survive. The 2027 mission tests whether two satellites can talk to each other fast enough to act like one computer — by firing lasers across the gap between them at tens of terabits per second.

That link is the whole ballgame. A single satellite with four chips is a gadget. A cluster of 81, lashed together by light into one machine, is a data center. Without ultra-fast, ultra-precise laser links, the constellation is just a swarm of lonely boxes.

And it's brutally hard. The satellites must fly in tight formation and hit moving targets with laser beams, continuously, without drifting or dropping the connection. Miss by a fraction, and the data stops.

There's a cost clock, too. The whole dream only pencils out if launch gets cheap enough — Google's own researchers peg the economics to roughly $200 per kilogram, a price that depends on rockets like Starship flying often and cheaply.

So watch two things: whether the chips come through this year's radiation test, and whether the 2027 laser link actually locks. Those decide if 81 satellites is a plan or a dream.

Let's keep this simple.

An AI data center is a giant room full of chips. Those chips do two things constantly: they guzzle electricity, and they get hot.

Feeding and cooling them is now one of the hardest problems in tech.

Here's the squeeze. A single mid-sized AI data center can eat as much power as 16,500 homes. The biggest ones use far more. And they're multiplying fast, as every tech giant races to build more AI.

The power grids can't keep up. Communities are pushing back on the noise, the water used for cooling, and the strain on local electricity. Land near cheap power is getting scarce and contested.

So the giants are getting desperate. They're buying entire nuclear plants. They're signing decade-long power deals. And they're still short.

That's the itch Google is trying to scratch in orbit.

In the right orbit, the sun never sets. Solar panels soak up power around the clock, far stronger than on the ground, with no clouds and no night. There's no neighborhood to annoy and no land to fight over. The room to expand is, in effect, endless.

But space swaps one hard problem for another. On Earth, cooling is easy and power is hard. In orbit, power is easy and cooling is brutal — there's no air or water to carry heat away, so every bit of it has to be slowly radiated into the void.

That's the real bet behind those four little chips. Not that space is cheaper today — it isn't. But that the thing choking AI on Earth, power and land, simply isn't scarce up there.

If the heat problem can be tamed, the ceiling lifts. And the company that proves it first rewrites where the world's thinking actually happens.

Follow the power problem, not the chips — it's the reason this moonshot exists at all.