Google's AI satellite answers from orbit⁠.

Google's first data centre satellite has answered its first prompt from space, The Verge reports¹. Google said in a blog post on Friday that it had booted up the satellite's Tensor Processing Units, its own AI chips, and that the Gemma 4 model on board then wrote a haiku about being in space¹. Engineers also sent space themed limericks and asked the spacecraft to report on conditions in orbit². A spokesperson told TechCrunch the satellite was operating nominally².

The satellite is part of Project Suncatcher, an experiment that could one day put AI data centres in orbit¹. It went up earlier this month alongside more than 100 other spacecraft on a SpaceX Falcon 9 rocket, and spent about three weeks being commissioned before it was given any work². It carries four of the chips and generates up to 1 kW of electricity².

The limits are plain. The chips draw so much power that the satellite can run only one of them, for about 20 minutes at a time². That makes talking to the ground harder still². Google says the next tests will look at how the chips cope with heat stress and radiation in orbit¹. Radiation can flip bits in memory and corrupt data without tripping the usual checks, and tests on the ground suggested that matters less for running a model than for training one².

If the chips hold up, Google plans two more satellites in 2027, with more power and optical links between them². The company called that its next milestone of twin interconnected satellites¹. The bigger plan depends on launches getting far cheaper: a peer reviewed analysis by the Suncatcher researchers estimated SpaceX would need to fly its Starship rocket 1,800 times to get the cost savings it saw on earlier rockets². Google is designing for infrastructure that pays its way in the 2030s².

A haiku, not a data centre

Strip away the poem and this is a single test chip, switched on for minutes at a time, doing what a laptop does without a second thought. That is not a criticism. It is the first step of a plan that runs into the next decade, and it tells you where the people building AI expect the next problem to be: power and space for the machines, not the models. None of the reporting gives a British angle, and nothing here changes what a UK firm can buy today.

For a business in Leeds or Bristol the useful reading is the direction of travel. The firms selling you AI are spending heavily to find cheaper, larger places to run it, on the ground and now above it. If that works, the price of running models should keep falling. If it does not, someone pays for the attempt. Either way, you need do nothing this week except note that the haiku was not very good.

Also today

• Backdoored Claude extension on Cursor reached up to 5,000 users

  A malicious version of Anthropic's VS Code extension was published through its official Cursor account and could reach up to 5,000 developers, installing a hidden SSH backdoor³.

• Mistral says Le Chonk is the best open model for security work

  Mistral says the finished version of its trillion parameter open weight model, Le Chonk, now matches or beats the leading Chinese models on coding and cybersecurity, and anyone can download it⁴.

• Chatbots are sending hikers into trouble on real mountains

  Rescue teams in the US and Canada say walkers who planned trips with AI chatbots have ended up lost, underprepared or camping where they should not, including three hikers on Mount Shasta⁵.

• US tech shares lose nearly $1tn after weak results

  A sell off in US technology shares knocked nearly $1tn off the market after disappointing results, as investors questioned whether AI spending will pay back quickly enough⁶.