Correction: The TPUs Are Already Up There
Correction issued 4 October 2026. In yesterday's digest we wrote that Project Suncatcher "launches two satellites in early 2027". That was wrong about the first one. Suncatcher's first prototype launched on 1 October 2026, at 2:32pm EDT from Vandenberg, on a SpaceX Falcon 9 as part of the Transporter-18 rideshare.
Our source described the two-satellite configuration as the plan and did not mention that one had already flown. We should have checked the launch manifest before publishing.
What is actually in orbit, built with Planet Labs:
- Four Trillium-generation TPUs - the same parts running in Google's terrestrial data centres, not space-rated variants
- About one kilowatt of power
- 650km, sun-synchronous dawn-dusk orbit
- Roughly one year of expected operation, with up to six years before atmospheric drag causes reentry
- It will run inference on Gemini models
And the Hard Problem Was the Other One
Everyone assumed radiation would be the obstacle. It was not. The chips survived ground testing at exposure levels well above what the orbit delivers.
Thermal management is the constraint. In vacuum there is no air to carry heat away, so the only mechanism available is radiative emission - a surface slowly shedding energy as infrared. Heat pipes and external radiators do what they can.
The consequence is concrete and it reshapes the whole proposition: the satellite computes in roughly 15-minute bursts, then stops to cool.
Think about what that does to the orbital data centre pitch. The appeal was continuous solar power in a dawn-dusk orbit - energy without a grid connection, without a substation, without a county planning board. But power was never going to be the only limit. A machine that must pause every quarter of an hour to radiate heat into space is not a drop-in replacement for a rack that runs continuously, however much sunlight falls on it.
This is why you fly the thing. One satellite has already overturned the assumption the entire programme was being discussed around.
A 16.9MB Speech Model That Beats Whisper Base
Cactus Compute released Whistle, a speech-to-text model at 16.9MB - roughly a ninth the size of the 145.3MB Whisper Base, which it outperforms on multiple benchmarks.
It posts 11.1ms time-to-first-token and 1,319 tokens per second across seven languages.
At that size the model fits comfortably on a phone, a watch or a microcontroller, with no network round trip and no inference bill. The industry spent this week discussing gigawatts and orbital radiators; this went the other way and arrived at something you can ship inside an app bundle.
An $18,000 Humanoid
Astribot demonstrated the T1 at IROS 2026 in Pittsburgh: 1.55m, 66kg, cable-driven, 23 degrees of freedom, running the Lumo-2 foundation model, with a starting price of $18,000 in the US.
That price is the news. Humanoid demonstrations have been plentiful and priced like industrial equipment. $18,000 is a used car, and it is the first figure in this category that invites a buying decision rather than a procurement process.
$42 Billion of European Data Centres Stalled
Roughly $42 billion of European data centre projects are now stalled or cancelled amid local opposition, with similar resistance spreading to South Korea.
Set that beside the week's other infrastructure stories - Google putting TPUs in orbit, Oracle waiting until at least December 2027 for power in Wisconsin, SpaceX buying $2.8 billion of gas turbines. The constraint on AI is no longer chips. It is electricity, land, and whether the people who live near the substation agree.
Patch Your GitLab AI Gateway
CVE-2026-90970, CVSS 9.9, affects the self-hosted GitLab AI Gateway across versions 18.1.6 through 19.4. Fixed in 19.2.4, 19.3.2 and 19.4.1.
A 9.9 in the component that brokers AI requests inside your development pipeline is worth moving on today, particularly given Microsoft's report this week putting median time from vulnerability to weaponisation well below 24 hours.
The Thread
Two of today's stories are about heat and land, and two are about getting smaller. Google learned that an orbiting TPU has to stop and cool every fifteen minutes. Europe has $42 billion of compute that cannot find anywhere to sit. Meanwhile a useful speech model now fits in 16.9MB and a working humanoid costs $18,000.
The expensive end of AI keeps running into physics and planning permission. The cheap end keeps quietly shipping.