Google's Space Data Center Plan Requires 1,800 Starship Launches
· Tech · Cem Koyluoglu
Google’s orbital edge‑computing satellite prototype launched today from California on a SpaceX rocket.
TL;DR Google has launched its first orbit-based computing prototype, sending a Tensor Processing Unit (TPU) into space aboard a SpaceX rocket. The mission will test the TPU’s performance and radiation resilience in a 1‑kilowatt power environment, paving the way for a future 81‑satellite orbit data center. Google’s long‑term vision includes scaling this infrastructure to handle AI workloads with low latency and high bandwidth.
Key Highlights • First Google TPU deployed in orbit, powered by 1 kW and tested for radiation tolerance. • Mission runs TPU in 15‑minute cycles to protect power and thermal limits. • Google plans an 81‑satellite network for parallel AI processing, aiming for $200/kg launch cost by 2035. • Technical report projects Starship could deliver 370,000 tons to orbit, requiring ~1,800 launches over a decade.
Google Launches Orbit Computing Prototype
Google’s prototype for an orbit‑based computing satellite was launched today from California aboard a SpaceX rocket, marking the first time the tech giant has sent one of its advanced chips into space.
The satellite, built by Planet Labs, will test the viability of Google’s Tensor Processing Unit (TPU) in the space environment. The experiment will provide the chip with uninterrupted 1 kilowatt of power, cooling, and a comprehensive suite of models to assess performance and detect any anomalies.
Travis Beals, the Google manager overseeing Project Suncatcher—a plan to develop large‑scale computing clusters in Earth orbit—stated, “We ran ground tests, but you know, no test can fully replace reality.”
Operational Profile
During the mission, the satellite will run the TPU in 15‑minute intervals to avoid overtaxing its power and thermal management systems. While the current platform is a standard Planet Labs design, the two companies are collaborating on a new demonstration set for launch next year. The upcoming satellites will be tailored for advanced computing, handle heavier workloads, and attempt to collaborate via laser‑based communication links.
The SpaceX launch vehicle carries more than 100 payloads, including Satlyt and Cowboy Space Company, and the Suncatcher is not the sole AI payload aboard.
Long‑Term Vision
What sets Google’s initiative apart from other startups—and even from SpaceX itself—is its long‑term horizon. Beals described the project as a “long‑term ambitious goal,” focused on building systems for future space infrastructure and AI workloads. The company envisions an orbit data center composed of an 81‑satellite network that can process data in parallel.
He added, “When we try to run a multi‑layered workload, the bandwidth and latency between TPUs become critically important… We’re looking not just at today’s workloads but where they will be five years from now.” The lack of existing rockets capable of scaling orbit data centers cost‑effectively is a major factor in this challenge.
Technical Report on Orbit Data Centers
On Thursday, Google released a peer‑reviewed technical report on orbit data centers, slated for publication in the journal Joule. The study is not framed as a purely economic feasibility analysis, but it presents an intriguing picture of how launch costs could decline over time.
Like all data‑center companies, Google relies on SpaceX to lift its spacecraft, and the company is also a major investor in the rocket maker.
The authors argue that since the launch of SpaceX’s Falcon 1, the company has achieved a “learning curve” that reduces launch costs by roughly 20 % per year. They therefore consider it reasonable to expect launch prices to reach about $200 per kilogram by 2035.
To reach that target, researchers believe Starship would need to carry 370,000 tons of payload to orbit, implying roughly 1,800 launches over the next decade—about 180 per year. This is a steep expectation for a vehicle that has not yet launched more than five times a year. Elon Musk has suggested that Starship could reach an hourly launch cadence by 2029, though such projections are speculative.
Radiation Hardness and Reliability
At least one positive takeaway from Google’s updated research is the chips’ apparent resilience to space radiation. The company discovered that the chip’s configuration offered more shielding than initially anticipated, prompting a repeat of particle‑accelerator irradiation tests. The tests introduced a few additional logic errors, but Google remains confident that the chips can sustain the inference workloads of a satellite over a five‑year operational life.
Beals remarked, “When you think about typical inference operations, the error rate is very low, right? One in a million. But when you consider a massive training process with thousands of chips running for months, that rate becomes problematic.”