TPU on Google Satellite Completes First Space AI Inference and Returns Results
Google's Project Suncatcher prototype satellite booted up a TPU in low Earth orbit, received a ground command to complete a model inference, and transmitted an AI-generated haiku back to Earth. Sundar Pichai called this the project's hello world milestone. The author believes that space-based solar power and deep-space radiative cooling can bypass the power and cooling bottlenecks of ground-based data centers.
Why read this
AI score 75
Google sent a TPU into low Earth orbit to complete the first space inference, providing a referenceable path for computing power to bypass ground-based power and cooling constraints.
Topics
View original postA ‘hello world’ milestone for Project Suncatcher 🚀 After launching our prototype satellite a few weeks ago, we’ve now successfully sent a command up to orbit, booted up the TPU, ran the prompt, and received our first complete response back on earth - a haiku! https://t.co/kOFmETUskm
View original
A ‘hello world’ milestone for Project Suncatcher 🚀 After launching our prototype satellite a few weeks ago, we’ve now successfully sent a command up to orbit, booted up the TPU, ran the prompt, and received our first complete response back on earth - a haiku! https://t.co/kOFmETUskm
Body · AI translation
WTF, Google just used a satellite to transmit back to Earth the first space AI inference result, which you could say is the Apollo moment for computing power escaping Earth's power grid.
It's not another PPT concept pitch, nor is it empty space-themed sentimentality. It's called Project Suncatcher, and it literally stuffs a self-developed TPU chip into a satellite, launches it into low Earth orbit, and has it transmit back to Earth a poem written by AI in the vacuum of space.
In the past, when building data centers on the ground, the energy ledger always ended with everyone cursing: each gigawatt of computing power corresponds to billions in electricity costs, substations aren't enough, cooling water drains the surrounding water systems dry, and local residents are still protesting carbon emissions every day.
The physical ceiling of ground infrastructure is about to hard-cap the iteration speed of large models.
Google's move just flipped the board.
They sent up a prompt from the ground, the TPU on the orbiting satellite instantly cold-started, ran the entire model inference in the zero-gravity environment of space, and then slammed the result back to Earth at millisecond latency.
This is the first time in human history that AI inference computing power has completely broken away from Earth's power grid.
What's most brutal is the physical energy logic behind it. In space there is no atmosphere blocking the way, no night, solar energy collection efficiency is several times that on the ground, and it's powered 24 hours a day without interruption; heat dissipation directly radiates to the absolute-zero deep space for cooling, without consuming a single drop of water or a single degree of air-conditioning electricity.
So the endgame of future AI computing power competition is not in chip architecture at all, but in the laws of thermodynamics. Whoever is still on the ground fighting over transformers is fighting the trench warfare of the previous generation; only those who dare to throw the racks directly into sun-synchronous orbit are opening up the next dimension of energy with a dimensionality-reduction strike. https://t.co/5AVdMyJbRa