Google tests satellite-linked TPUs to run machine learning in space
When Google calls something a "moonshot," it's rarely just a metaphor. Its latest experiment - Project Suncatcher - quite literally looks beyond Earth. Google has outlined an idea that sounds like it's ripped from a sci-fi screenplay: building a space-based AI infrastructure, powered entirely by sunlight, and interconnected through high-speed optical links.
At its core, Project Suncatcher is an attempt to reimagine where AI gets trained. Today's machine learning models, from Gemini to ChatGPT to Deepseek, demand staggering amounts of energy and compute power. The world's data centers already consume more electricity than entire countries. Google's pitch is bold - if sunlight is the limiting factor on Earth, why not go where it never sets?
Also read: Google wants to take AI to space with solar-powered data centres
Google's engineers envision a constellation of satellites, each equipped with solar panels and Tensor Processing Units (TPUs) - the custom chips that power many of the company's AI systems. Instead of relying on Earth's power grids, these orbiting machines would draw nearly continuous solar energy in what's called a sun-synchronous orbit, a path that keeps the satellites perpetually in daylight.
These satellites would then link together using free-space optical communication, essentially lasers transmitting data between satellites at terabit speeds. If you imagine fiber-optic cables stretched across space without the cables, you're not far off. Google says a lab demo has already achieved 800 Gbps each way, proving the physics works - at least in controlled conditions.
The result? A "data center" that lives in orbit, where energy is free, cooling is minimal, and sunlight never fades.
The immediate appeal is energy efficiency. In orbit, solar panels can generate up to eight times more power than on Earth, thanks to constant sunlight and no atmospheric interference. That could mean cleaner, cheaper, and more scalable compute without overloading terrestrial grids.
Then there's the environmental argument. As AI systems scale, their energy needs balloon. Google's researchers suggest that space-based infrastructure could reduce the strain on Earth's limited power supply while taking advantage of what is effectively limitless solar energy.
Finally, there's the scalability factor. Launch costs have plummeted, from tens of thousands of dollars per kilogram a decade ago to around $1,000 today, and are expected to fall further. Google estimates that once costs hit $200/kg, space computing could rival the cost-efficiency of terrestrial data centers.
Of course, it's one thing to imagine a "solar-powered AI cloud" in orbit, it's entirely another to build one.
Also read: Amazon threatens Perplexity with legal action over AI agentic shopping
For starters, keeping a tight formation of satellites flying just hundreds of meters apart in low Earth orbit is a massive challenge. Engineers model these formations using advanced orbital dynamics equations to prevent drift, collision, and interference. Google's prototype simulations show an 81-satellite cluster maintaining formation within a one-kilometer radius.
Then comes the radiation problem. Space is brutal on electronics - cosmic rays can fry delicate chips. Google tested its Trillium TPU v6e under a 67 MeV proton beam and found no catastrophic failures even at 15 krad of radiation, suggesting the chips might survive long-term exposure. Still, long-duration reliability remains unproven.
Thermal management, debris avoidance, and maintenance are other obstacles. There's no technician to swap a failed chip when your data center is 650 kilometers above Earth.
If it works, the implications could be profound. Imagine a future where the Earth's heaviest AI workloads - massive model training, climate simulations, or satellite imaging analysis - happen entirely in orbit, powered by the same sun that fuels life below.
It also raises fascinating questions about data sovereignty and infrastructure ethics. Who owns computation that happens outside any nation's borders? Could orbital AI systems change the geopolitics of data control? Google hasn't addressed these yet, but they'll matter if Project Suncatcher ever gets off the ground.
Right now, this is just a research concept, a thought experiment made real by prototypes and math. But it signals something larger: the boundaries of AI infrastructure are expanding, literally beyond Earth.
As AI pushes against the limits of compute, power, and sustainability, companies like Google are forced to look up and not just forward.
If the cloud was the 2010s' defining tech metaphor, Project Suncatcher hints at what comes next. The sky isn't the limit. It's the datacenter.