Urgent.News

What's breaking now, across thousands of outlets.

AI

Google's orbital AI data center test packs four TPUs and 1,000W of solar power

Google will launch an experimental satellite into orbit on October 1, taking four tensor processing units with it.

Google's orbital AI data center test packs four TPUs and 1,000W of solar power

Google plans to launch an experimental satellite named MVP next week to test running an AI data center in space. The satellite will be part of Project Suncatcher and carried by a SpaceX Falcon 9 rocket from Vandenberg Space Force Base in California. Upon reaching orbit, the technology will be capable of handling simple AI requests from space.

However, its capabilities will be limited, featuring only four Tensor Processing Units (TPUs) and powered by solar panels capable of providing a single kilowatt of power to the tensor chips. The satellite will continue operating for up to a year, with the potential to remain in orbit for up to six years before burning up upon re-entry into Earth's atmosphere.

Google's senior vice president of research, James Manyika, anticipates learning more about the potential for future data centers in space, though he expects it to be a long way off. The project draws parallels to Google's driverless cars, which took 15 years to become operational. Notably, other tech giants like Elon Musk, Jeff Bezos, and Sam Altman have also expressed interest in orbital data centers.

However, challenges such as radiation affecting computer processing and the need for cooling systems in space remain. Google plans to launch additional satellites to test the concept, with long-term goals of deploying a network of over 80 satellites. Manyika estimates that it will take Google 15 years to advance from a four-chip server to a full-fledged data center in space.

Written by urgent.news from Tom's Hardware's reporting — not their text. Machine-written — may contain errors; check the original before relying on it.

Also reported by 1 other outlet

Read the original at tomshardware.com →

More in AI

How to test whether a model can tell your MCP tools apart

You can read your own tool descriptions all day and not know whether a model can tell them apart. I spent a few weeks measuring 82,549 public MCP tool descriptions, and the most useful thing I learned…

  • Write two sentences per tool pair: one control request and one probe request
  • Route control requests unambiguously, probe requests to either tool
  • Incorrect routing indicates flawed tool description that needs rewriting

More from Friday 25 September →