Google's orbital computing power satellite was launched today aboard a SpaceX rocket from California, marking the tech giant's first time sending its advanced chips into space.

This satellite was manufactured by Planet Labs and will be used to verify whether Google's Tensor Processing Unit ( TPU ) – the chip that competes with NVIDIA GPU – can function properly in space. This requires providing a continuous power supply of 1 kilowatt, cooling the chip, and running a series of tests to observe if any issues arise.

"We've already conducted tests on the ground, but you know, no test can replicate the real environment perfectly," said Travis Beals, a Google executive in charge of Project Suncatcher. This project is part of the tech giant's plan to develop a large-scale orbital computing cluster that orbits the Earth.

Once operational, this satellite will launch in groups of 15 minutes to TPU in order to avoid putting excessive strain on the satellite's power supply and thermal management systems. The satellite is based on the standard platform constructed by Planet Labs, but two companies are collaborating on a demonstration project that is expected to be launched next year. By then, there will be two satellites that are more specifically designed for advanced computing. Future versions will also attempt to work together through laser communication links.

Suncatcher is not the only space AI payload on this SpaceX rocket. This launch carried more than 100 different payloads, including missions for Satlyt and Cowboy Space Company.

However, the biggest difference between Google's project and those of these startups, as well as SpaceX's own project, is that it is a long-term plan.

Beals said that the focus of this "long-term lunar exploration" project is to prepare for the future space infrastructure and the workload of AI. The company envisions a network consisting of 81 satellites that fly in close formation and process data in parallel.

"When you're trying to run multi-rack workloads, the bandwidth and latency between TPU are really crucial... What we're trying to envision is not just what workloads exist today, but what they will look like in five years," said Beals. He noted that this is largely because the rockets required to scale data centers in a cost-effective manner simply don't exist yet.

On Thursday, Google also released a peer-reviewed version of its white paper on orbital data centers, which is one of the most rigorous analyses to date on how computing power can be deployed in orbit. The paper will be published on Joule.

One of the most notable aspects of this paper is how Google views the cost of entering space. Although the researchers emphasize that their analysis is not an economic feasibility study, it still demonstrates how the company perceives the reduction in rocket costs over time.

Like all data center companies, Google also relies on SpaceX to help send its spacecraft into orbit. (Google is also a major investor in SpaceX.)

The paper author believes that since the launch of the Falcon rocket, Elon Musk's rocket team has achieved a cost reduction of about 20% per year on the 'learning curve.' Therefore, it is reasonable to expect that by 2035, the company will be able to reduce the launch cost to nearly $200 per kilogram.

What is required to achieve this? Based on the scale of the payloads already launched by Falcon, the author believes that a similar cost-cutting approach would require the starship to deliver 370,000 tons of payload into orbit. This means that approximately 1,800 launches would be needed over the next decade, which is 180 launches per year – assuming that each mission can transport 200 metric tons.

For a vehicle that has never flown more than 5 times in a year, this requirement is quite high. SpaceX expects the company to achieve much more than that – for example, Elon Musk has stated that Starship could possibly achieve a flight frequency of once per hour by 2029, although Musk often says many things.

At least in the research conducted after Google's updates, the good news is that its chips seem very likely to be able to withstand space radiation. The company had to conduct tests again in a particle accelerator because they realized that the chip configuration provided more shielding than what is actually found in space. This did result in a slight increase in errors in the chip's logic circuits, but the company still believes that its chips will be capable of handling large-scale inference workloads throughout the five-year lifespan of the satellites.

“If you’re considering typical reasoning operations, the error rate is very low, right? For example, one in a million,” said Beals. “On the other hand, if it’s some kind of ultra-large-scale training task, then there are already problems, because you would be running thousands of chips continuously for months.”