What happens if AI data centres slip the ‘surly bonds of earth’?
Outsourcing this infrastructure to space comes with a host of problems
Anjana Ahuja
© Andy CarterIf ever there was proof that the race to feed the AI boom has become detached from reality, it came in a recent Google announcement that it would build a prototype solar-powered data centre in space.
Data centres are power-guzzling physical facilities that contain computing infrastructure; planning hurdles, local opposition and energy demands are now putting limits on how far and fast AI can scale.
The idea of Project Suncatcher is to shift that infrastructure, and its associated headaches, into space.
Queries to AI models like Gemini would be processed there and the answers beamed back to Earth.
The data centre would not be an orbiting monolith but a flock of 81 solar-powered satellites loaded with AI chips, flying and processing in unison.
Google will work with satellite company Planet to send two satellites into low Earth orbit in 2027.
This would be a cosmic challenge — but while orbiting data centres minimise the need for planetary land and water, they promise unearthly problems of their own.
Astronomers fear more satellite constellations will further clutter the skies, hampering scientific observations.
Mass launches into crowded orbits risk creating space debris that could threaten vital services such as weather and climate monitoring.
But we should also stop for a moment to consider this: a technology, namely AI, being pushed by a handful of companies is creating an energy demand that, by the industry’s own estimates, cannot be met within planetary boundaries.
For many, that is the very definition of unsustainable.
The scheme also exposes the lack of effective governance when it comes to space, a commons that has slowly become the (almost) rule-free playground of tech titans such as Jeff Bezos and Elon Musk but in real life belongs to us all.
According to the preprint published by Google researchers last month, the satellites would travel in a Sun-synchronous orbit around 650km above the Earth.
That tracking of the Sun means near-constant solar power for the AI payload but also thrusts the satellites into the single most congested route in low Earth orbit, according to Michigan university space scientist Mojtaba Akhavan-Tafti.
The satellites would be spaced just 100m to 200m apart, allowing them to talk to each other but perilously close should their paths wobble due to buffeting from air particles and space weather.
At that proximity, Akhavan-Tafti explained in a recent article, “the margin for error [in orbital navigation] evaporates.
A single impact could not only destroy one satellite but send it blasting into its neighbours, triggering a cascade that could wipe out the entire cluster and randomly scatter millions of new pieces of debris into an orbit that is already a minefield.”
The European Space Agency notes that debris particles over 1cm in size can cause catastrophic damage — and there are already more than 1.2mn of them.
As orbital traffic soars, the probability of the Kessler effect — flying debris that sparks a chain reaction of other collisions and break-ups, rendering those orbits unusable — also rises.
There are nearly 16,000 satellites already whizzing around the Earth.
Of those, close to 9,000 belong to Musk’s Starlink network, with the company’s plans for 15,000 new ones under review by the Federal Communications Commission.
The reflective satellites can be seen in optical images, though the company seems willing to address the issue; stray Starlink radio emissions are showing up in frequencies meant to be protected for radio astronomy.
The Google preprint mentions that the AI chips will be exposed to radiation that can degrade electronics and corrupt data, but rather sketches over the issues of remote satellite maintenance and how to extract defunct or malfunctioning hardware.
Microsoft can offer one reality check on putting data centres in remote locations: it sank an underwater data centre off the coast of Scotland in 2018 but that effort, called Project Natick, has reportedly ended and the company is quoted as having no further subsea ambitions.
Achieving the same in space would be many times harder: even as launches get cheaper, the unit costs of power in space and on land seem roughly comparable.
When approached for comment, Google referred to the preprint and an associated blog summarising its position.
One might be tempted, fairly or not, to interpret the ambitious corporate effort to yoke infinite solar power to the seemingly infinite demand for compute, as a signal to investors that AI is infinitely scalable.
It is also quite the metaphor for an industry that is flying ever closer to the Sun.
The writer is a science commentator
0 comments:
Publicar un comentario