The search giant has now placed one of its sophisticated computer chips into orbit, aboard a rocket fired from California by SpaceX. This marks the first instance of Google sending such an advanced processor to space.
The satellite, built by Planet Labs, carries a Google Tensor Processing Unit — the company’s rival to Nvidia’s GPUs. The test aims to prove the TPU can function in space. That means supplying a kilowatt of continuous power, cooling the chip, and running models to see if anything breaks.
“We’ve done testing on the ground, but you know, there’s no test that’s completely as good as the real thing,” said Travis Beals, the Google executive managing Project Suncatcher, the tech giant’s plan to build large-scale compute clusters in orbit around the Earth.
The Satellite’s First Burns
After launch, the satellite will switch on its TPU in brief sessions lasting no longer than 15 minutes so as not to overtax its power and thermal management systems. It uses a generic design made by Planet Labs, although both firms are now developing a prototype set to fly next year, featuring two satellites engineered specifically for advanced processing. Those upcoming models will try to work together using a laser communications link.
This SpaceX rocket is carrying more than 100 separate payloads, not just Suncatcher. The launch also includes missions from Satlyt and Cowboy Space Company.
What sets the Google initiative apart from those startups (and indeed from SpaceX itself) is that it’s a long-term project. The focus of this “long-term moonshot,” as Beals puts it, is on building for the space infrastructure and AI workloads that will exist in the future. The company envisions a network of 81 satellites flying in close formation, processing in parallel.
“The bandwidth and the latency between TPUs really, really matters when you’re trying to run a multi-rack workload…we’re trying to look ahead to not just what workloads exist today, but where they will be in five years,” Beals said. That’s largely because the rockets required to scale up orbital data centers in a cost-effective way don’t yet exist.
The Paper Behind the Launch
Thursday saw the release of a peer-reviewed version of Google’s white paper on orbital data centers, one of the most thorough examinations of how computing hardware reaches orbit. The research is set to appear in Joule.
A document from the paper’s most notable aspects is how Google thinks about access to space. Although the researchers stress their analysis isn’, which serves as an economic feasibility study, provides a revealing look at how the company envisions rockets getting less expensive over time.
SpaceX is the launch provider that Google is counting on to send its spacecraft into orbit, just as every other data center company relies on it. The search giant is also a major investor in the rocket company.
The authors argue that SpaceX has achieved a price-reducing “learning curve” of about 20% a year since they launched the Falcon 1 rocket. That pattern suggests it’s reasonable to expect the company to deliver launch prices close to $200 per kilogram by 2035.
What 1,800 Launches Look Like
To match that trajectory, the paper’s authors estimate Starship would need to fly 370,000 tons of payload into orbit. That works out to about 1,800 launches over the next 10 years, or 180 a year — and that’s if it can fly 200 metric tons on each mission.
SpaceX expects to fly far more than it has ever done before. The vehicle in question has only flown five times in a year, and that is a small number. Elon Musk has said that Starship could reach an hourly flight rate in 2029, though he has also made other claims that have not come true.
| Vehicle | Current Record | Target Rate |
|---|---|---|
| Starship | Five launches in a year | Hourly flights by 2029 (Musk) |
| Falcon 9 | Flying regularly now | 20% learning curve since Falcon 1 |
Radiation Tests That Failed Twice
Google’s fresh research suggests the firm’s chips stand a decent chance of surviving the harsh conditions of space. The company was forced to repeat tests in a particle accelerator after discovering that the arrangement of the chips offered more protection against radiation than what they would actually face in orbit. That revised testing produced a few more errors in the logic circuitry of the chips. Despite that, Google remains convinced its chips can manage heavy inference workloads during the five-year life span of a satellite.
“The error rate is very low if you’re thinking about typical inference operations, right? Like one in a million,” Beals said. “On the other hand, it was already problematic for doing, say, some mega-scale training run where you’re going to have many thousands of chips running for months.”
The Cost Math
Here is how the paper’s cost projections stack up:
- SpaceX has reduced launch prices by about 20% per year since the Falcon 1 rocket launched.
- The authors expect SpaceX to hit $200 per kilogram by 2035.
- To match that trajectory, Starship would need to fly 370,000 tons of payload into orbit.
- That works out to about 1,800 launches over the next 10 years, or 180 a year.
- Starship has never flown more than five times in a year.
What the authors are banking on is a trend, not on a device that already exists. Their wager rests on the assumption that the upward slope keeps going. It rests on the presumption that the hourly flight rate will arrive eventually. And it rests on the presumption that the chips will hold up.
What This Means for the Chips
The company has placed TPUs in orbit to run inference workloads there. During those tests, the chips proved to be better protected against radiation than the team had predicted, which actually caused more errors than they had expected. Still, the error rate stays low for most inference tasks — one in a million, as Beals reported.
When the size grows, the problems grow too. A massive training run lasting months across thousands of processors generates so many errors that the whole system starts to fail. This limit on what can be built is a true constraint on the design. It means adding more processors to a training task does not fix things — the job stops working instead.
The Long-Term Bet
Beals described Suncatcher as a “long-term moonshot.” That is the right word for it. The project is betting that the infrastructure will arrive eventually, and that the economics will work out.
The authors of the paper are banking on the same projection. They are betting that launch costs will fall to $200 per kilogram by 2035, with Starship needing to complete 1,800 launches to reach that point. That represents a substantial number of flights.
But the math only works if the rockets actually fly. And Starship has not proven it can do that yet. It has flown five times in a year. It has not flown 180.
A lot of statements have come from Musk. One of them was a promise of hourly flight rates.
The search giant is banking on a world where rockets take to the skies. But it remains unclear whether those rockets will ever leave the ground.
A prototype satellite has been launched today, while a paper in Joule presents a projection rather than an existing system. None of the launches that would be required to bring the entire project into operation have taken place yet.
Whether the ambition will be matched by the infrastructure remains open. The paper presumes it will happen. But the record of Starship flights points in the opposite direction.
Google has invested heavily in SpaceX. It has also invested in its own orbital compute project. The two bets are connected. If SpaceX cannot launch cheaply, Google’s satellites become expensive. If SpaceX can launch cheaply, Google’s satellites become possible.
Whether the plan succeeds depends on what happens next. Today’s launch is the first step, a satellite sent up into space. The paper in Joule sets out the plan. The launches that would put the whole system into place have not taken place yet.
Source material: “Google thinks SpaceX’s Starship has to launch 1,600 times before space data centers get off the ground,” TechCrunch.
Get the Notebook.
The day's best stories and every fresh verdict, in plain English, in your inbox by seven. One email a day, no more.