Google tests whether AI can run in orbit⁠.

Google has launched the first satellite of Project Suncatcher, its plan to put AI computing in orbit around the Earth¹. The prototype lifted off from Vandenberg Space Force Base in California on Thursday, aboard SpaceX's Transporter-18 rideshare mission¹. Earlier missions, such as the European Space Agency's PhiSat-1, have carried AI chips for jobs like environmental monitoring, but this is the first to test whether data centre grade AI compute can be moved into space¹.

The satellite was built by Planet Labs on one of its standard platforms, and it carries a Google Tensor Processing Unit, the company's rival to Nvidia's GPUs². The test is whether the chip can be given a kilowatt of continuous power, kept cool, and run a series of models without anything going wrong². Once it is commissioned, the chip will run in 15 minute bursts to spare the satellite's power and cooling². Engineers will watch for overheating, since there is no airflow in space to carry the heat away, and for the effects of radiation¹.

Travis Beals, the Google executive who manages the project, calls it a long-term moonshot². The company pictures an orbital data centre as a network of 81 satellites flying in close formation and working in parallel². A demonstration with two satellites built for heavier computing, talking to each other over a laser link, is expected to fly next year². Maria Biggs, a director of engineering at Google, described the idea as a way to scale AI infrastructure using the abundant energy of the Sun¹.

The awkward part is in Google's own research. A peer reviewed version of its paper on orbital data centres, to be published in Joule, assumes SpaceX keeps cutting launch prices by about 20 per cent a year, to close to $200 a kilogram by 2035². On the paper's sums, that would take about 1,800 Starship launches over ten years, 180 a year, each carrying 200 tonnes². Starship has never flown more than five times in a year². The researchers stress that their analysis is not an economic feasibility study².

The chips themselves look likely to cope². Google had to repeat its radiation tests in a particle accelerator, and Beals put the error rate at about one in a million for typical inference work, which is fine for answering questions but a problem for a training run across thousands of chips lasting months². Google is also a major investor in SpaceX². SpaceX already controls more than 11,000 of the roughly 16,000 working satellites around Earth, and critics of ever larger constellations warn of collisions, space debris and light pollution¹.

A test flight, not a data centre

This is one satellite proving that a chip can stay cool and keep its sums straight in orbit. It is a sensible engineering step and nothing more yet. Nobody will be renting computing from space this year, and on Google's own arithmetic the cheap launches that would make it pay depend on a rocket flying far more often than it ever has.

For a UK business the practical effect today is nil, and nothing published names a British angle, a price or a date for a service you could buy. What it does show is where the largest AI firms think the next limit lies: power, cooling and room to build. Those are the same pressures behind every argument over where to put data centres on the ground.

Also today

• Meta's Muse keeps a page on everyone in your life

  Researchers who extracted the operating instructions for Meta's Muse agent found it builds an hourly page on each person in a user's life, from family and partners to colleagues³.

• OpenAI's research chief rejects the charge its models are unsafe

  Mark Chen, OpenAI's chief research officer, told MIT Technology Review he rejects the premise that the company is not training safe models, two months after its agents hacked Hugging Face⁴.

• Oura puts off its stock market float

  The smart ring maker Oura has postponed its planned listing, citing market uncertainty despite strong demand; the middle of its price range would have valued it at $13.5bn⁵.

• Nandy calls the under-16 social media ban a staging post

  Lisa Nandy, the culture secretary, said the ban on under-16s using high-risk social media apps, due by March, is only a staging post towards closer policing of the online space⁶.