Google Put Four AI Chips in Orbit. Now Comes the Hard Part

Frontier Signal · Deep Dive · Space & Investing · 8 min read · Monday, 12 October 2026

Every question you put to a chatbot is answered by a chip that does a great deal of arithmetic and gives off heat. Multiply that by billions of questions and you get data centres hungry for electricity, land and cooling water. Power is now the main brake on AI.

Space looks like an answer. Sunlight there is nearly constant, with no night, no cloud and no planning permission. Google says panels in the right orbit could collect up to eight times more solar energy a year than the same panels on the ground.

October 2026, Google and the satellite company Planet launched the first piece of hardware built to test that idea: a prototype carrying four of Google's AI chips.

It is a very small step: Google's project lead calls it a very minimal test. Here is what is being tested and how far away the payoff is.

4 CHIPS IN ORBIT · 15-MINUTE RUNS · ABOUT $200 PER KG LAUNCH COST NEEDED
The $200 per kilogram is Google's own threshold for matching ground data centres in the mid-2030s. A Falcon 9 launch costs roughly $3,600 per kilogram today, per analyst firm Futurum.

October 2026, Google launched a prototype Project Suncatcher satellite on SpaceX's Transporter-18 rideshare flight. It was built with Planet, is about the size of a refrigerator, and carries four TPUs. Google says the team has confirmed contact and the satellite is operating as expected. The four-chip count comes from NPR and a trade report, not Google's post.

The mission is meant to run for about a year, and Google will collect data on launch stress, radiation and temperature swings in the coming weeks. Google plans two more satellites in 2027 to test links between them. It calls the whole idea a long-term research moonshot, not a product.

The November 2025 plan described 81 satellites about 650 kilometres up, linked by lasers. That design is illustrative. What flew is far smaller.

The old way: AI chips sit in large buildings on Earth, drawing grid power and shedding heat with air or water. The limits are electricity, planning permission and water.

The idea: a satellite in a sun-following orbit carries solar panels and AI chips. Sunlight is almost constant, so it needs few batteries. Heat leaves through large radiating panels, and many satellites flying close together link up by laser to work as one computer.

The prototype is much simpler. Its four chips run a small version of Google's Gemma model for 15 minutes at a time, while Google watches for errors. It uses batteries and a standard low orbit. NPR corrected an earlier report that said otherwise.

In plain English: this flight asks one question. Can an ordinary AI chip, properly protected, simply work in space?

Space is hard on electronics. Fast charged particles flip bits in memory and wear chips down over time, a problem even for crews flying round the Moon. Google fired a proton beam at its Trillium chips at UC Davis. It reports no hard failures at a total dose of 15 kilorads, against about 0.75 expected over a five-year mission. That is roughly a 20-times margin.

The memory is the weak point. The high-bandwidth memory beside the chip showed irregularities at about 2 kilorads, still nearly three times the mission dose.

These are ground results. They come from Google's team and a paper that Google says is now peer reviewed in the journal Joule. The real test is a year in orbit.

There is no air in space. On Earth, fans and water carry heat away. In orbit the only way to lose it is to radiate it from a surface. Radiators are heavy, and every kilogram costs money to launch.

Analyst figures show the scale. Futurum puts radiator performance at about 300 watts per square metre, with roughly 1.3 square metres needed per chip. That is why the prototype runs short bursts and not flat out.

Repair is a separate problem. A failed chip on Earth is swapped in minutes. In orbit it stays failed. Carnegie Mellon's Brandon Lucia told NPR the cost and complexity are amplified by a factor of 10, maybe 100, and called the idea very sci-fi, though much of the technology is proven in research. Industry is only now working on standard ports for servicing satellites.

Google's own threshold is steep. Its paper says orbital compute could match ground data centres, per unit of power, if launch falls below about $200 per kilogram by the mid-2030s. Futurum puts a reusable Falcon 9 at roughly $3,600 per kilogram today. That gap is the whole economic argument.

It leans on a rocket that is not flying at scale. Futurum notes the plan assumes SpaceX's Starship launching on the order of 180 times a year. It also estimates orbital infrastructure at about $72 billion per gigawatt, against $16 billion for an off-grid ground equivalent, excluding the chips. Those are analyst estimates, not Google's.

Google is candid. Project lead Travis Beals told NPR he does not see space being cheaper within five years and thinks it will take longer.

Alphabet (NASDAQ: GOOGL; NASDAQ: GOOG) owns the project and the chips. If orbital compute works, it holds the design. The near-term effect on its results is nil, since Google calls this a moonshot. It also depends on SpaceX, which launched the satellite and is a prospective rival.

Planet Labs (NYSE: PL) built the prototype with Google. The read-through is a services and engineering relationship. We have not seen a contract value, so we cannot size it.

Nvidia (NASDAQ: NVDA) is the other name in the race. Futurum reports a space-focused module for its Vera Rubin chips, claiming up to 25 times the H100's space inference performance. That is a vendor claim and availability was not stated.

SpaceX (NASDAQ: SPCX) listed in June 2026. Its IPO filing says it expects to begin deploying orbital AI compute satellites as early as 2028. That is a company expectation, not a result.

Starcloud launched an Nvidia H100 in November 2025, the first of that chip in orbit, and Futurum reports a $170 million round at a $1.1 billion valuation. We have not checked those figures against a primary source.

Servicing and in-space manufacturing firms are the quiet enablers. If orbital compute is ever built, it will need repair and assembly in orbit.

This proves almost nothing about the economics. Four chips on a 15-minute duty cycle is a materials test. It cannot show cost per calculation, which is what decides whether orbit beats the ground.

The hard parts have not flown. Formation flying at 100 to 200 metres, laser links between satellites, large radiators and many-chip satellites are all still ahead. A laser link at 800 gigabits per second each way has been shown on a bench, not in orbit.

The eight-times figure is a ceiling. Google says up to eight times, and it is a premise, not a measurement. Duty-cycle limits could make cost per useful calculation far worse than on the ground. That is an analyst view, not a finding.

What does one useful calculation cost in orbit, all in, compared with the ground?

Include launch, radiators, shielding, spare chips, communications and the satellite's short life. If the answer needs launch below $200 per kilogram, ask who will deliver that, and when.

A Google research project to test whether AI chips could run on solar-powered satellites in orbit. The first prototype launched on 1 October 2026 with four TPUs.

Not yet. The prototype is a four-chip test of survival and basic operation, and Google does not call it a data centre.

Near-constant sunlight for power, and no land or cooling water. The catch is that cooling in a vacuum is hard, launch is expensive and repairs are very difficult.

Google calls it a long-term research moonshot, and its project lead does not expect space to be cheaper than the ground within five years.

Not on this evidence. It is a research test, so treat company read-throughs as questions to ask, not conclusions.

We read Google's blog post, NPR and the Futurum analysis directly. Figures from Futurum are analyst estimates.

Frontier Signal explains frontier technology in plain English. Vendor and analyst claims should be independently verified. This is general information, not investment advice.