Google's Suncatcher puts its first TPUs in orbit
Compute is leaving the ground — literally — while the people who build AI models keep warning about where the buildout ends. Three stories this hour.
Google's first Suncatcher satellite is in orbit, and it is carrying the company's own TPUs. The prototype launched Thursday aboard a SpaceX Falcon 9 from Vandenberg on the Transporter-18 rideshare, tucked alongside Planet Labs satellites, and Google confirmed it has contact with the spacecraft and it is operating as expected. Before shipping, the Trillium TPUs went through proton-beam testing at UC Davis and absorbed a radiation dose larger than a five-year mission would deliver — the hard question for orbital compute was never launch cost, it was whether chips survive space, and the first data point says yes. Google plans two more experimental satellites in 2027 with laser inter-satellite links, which is the step that turns a stunt into a network. We flagged the launch window last month — Google is putting its first TPUs in orbit next week — but contact confirmed is the milestone that matters: orbital data centers just moved from slide deck to hardware with a heartbeat.
A 22-author report led by Geoffrey Hinton and Yoshua Bengio says AI-assisted R&D is approaching the trigger point for an intelligence explosion. The working paper, published September 28 by Cambridge's Programme on AI Science and Policy with GovAI, is not an open letter — it is a researched argument, and its numbers do the talking: Anthropic's share of approved code written by AI rose from low single digits to over 80% between January 2025 and May 2026, and AI-run R&D with only high-level supervision went from 1% to 26% between March and August. The authors — who also include OpenAI chief scientist Jakub Pachocki, Anthropic co-founder Jack Clark, Eric Horvitz, Dawn Song and Jeff Clune — extrapolate months-long AI R&D projects being automated by mid-2028, and argue current gains have not yet crossed the explosion threshold but are approaching it. Their asks are modest for the claim: transparent R&D reporting with embedded auditors, constraints on development speed, and preparation for a fast takeoff. The interesting shift here is who is signing: this is no longer a safety-camp warning, it is the people shipping the systems.
Nebius bought inference-optimizer Inferize to stop paying what it calls the idle GPU tax. The neocloud announced the acquisition on October 1; Inferize's technology and 17-person team fold into Nebius Token Factory, its production-inference arm, with the goal of making billed capacity track actual usage instead of idle silicon. Financial terms were not disclosed — sources told CTech the deal is estimated at $100 million to $150 million, and that number should stay labeled as an estimate. Inferize was a 10-month-old stealth startup in Tel Aviv founded by the Granulate team, which makes this a talent-and-tech tuck-in at a time when every neocloud is racing to cut the cost per token it serves.
What to watch: Google's 2027 satellite manifest, and whether the paper's authors get any policy traction ahead of it.
If AI now writes most of the code inside the labs that build it, who should audit the labs? Tell us in the comments.
