Future Technology[Daily Briefing] Google's 10x chip 🧊 / 16-year cloud hole 🕳 / 390 black holes 🌌
Google reckons it has built a chip 10 times better than its own TPUs
The number doing the rounds today is 6 to 10 times. That is the claimed efficiency of a Google server chip code-named Frozen v2 compared with the TPUs Google runs now, according to internal sources rather than any official announcement.
What makes it interesting is the design choice behind it. Frozen v2 is described as being built around the Gemini architecture, rather than being a general purpose accelerator that Gemini happens to run on. When you know exactly which model shapes you need to serve, you can throw away flexibility you were never going to use and spend that silicon on the handful of operations that actually run hot. Amazon did a version of this with Inferentia. OpenAI is reportedly doing its own. Google has the advantage of already being three chip generations into the habit.
The caution is the usual one. Google has not confirmed the chip or the figures, and efficiency claims made before a part ships in volume tend to come from the friendliest benchmark in the building. Memory bandwidth, yields and the general awkwardness of a real datacentre usually trim the headline number. Read 6 to 10 times as the ceiling of a sales case, not the floor of a spec sheet.
Why it matters: the expensive part of AI is no longer training the model, it is serving it a billion times a day. Every hyperscaler that gets its own inference silicon working pulls a slice of demand away from Nvidia and keeps that margin in house. If Frozen v2 lands anywhere near its claim, Google stops renting the most expensive part of its own product.
Quick hits
A hypervisor bug hid for 16 years. A flaw in the Linux KVM hypervisor, tracked as CVE-2026-53359 and nicknamed Januscape, lets a guest virtual machine escape onto the host on both Intel and AMD. VM isolation is the promise the entire cloud rents on, so patch KVM hosts now, along with the actively exploited SharePoint flaw CISA just added to its catalogue. Our write-up (3 min read)
We now have 390 gravitational wave detections. The latest catalogue added 161 new black hole collisions in one go, nearly tripling the record. Some of the black holes in there are heavier than our models say they should be able to get, and there are now enough of them that it stops being a fluke. Our write-up (3 min read)
Someone claims to be selling 35GB of Accenture data. An actor going by 888 says the haul includes source code and access keys. It is unverified, but the pattern is familiar: breaching a large services firm is really a way into its clients. Worth rotating vendor credentials this week regardless of how this one shakes out. Our write-up (3 min read)
Kimi K3 ran out of room. Moonshot AI suspended new subscriptions days after the model topped a coding leaderboard, because demand outran its serving capacity. A model that is excellent and unavailable is worse for your product than one that is merely good and reliably served. Our write-up (2 min read)
TSMC pushed its US bet to 265 billion dollars. The company expects structural AI chip demand to hold for years and is deepening the Arizona build-out, while DeepSeek is reportedly designing its own inference chip to lean less on Nvidia and Huawei. Inference, not training, is becoming the volume game. Our write-up (3 min read)
Tool of the day: Celestron Cometron 7x50 binoculars
With a third planet imaged around Beta Pictoris this week, and an atmosphere detected around a rocky planet in its star's habitable zone, it felt like the right week to point at something you can actually use. Not a telescope. A pair of 7x50 binoculars.
The honest take: binoculars beat a cheap telescope for a beginner almost every time. They are quicker to set up, they show you a wide enough field to actually find things, and they survive being chucked in a bag. You will see Jupiter's moons, craters along the lunar terminator, the Pleiades looking properly three dimensional, and the Andromeda galaxy as a faint smudge that is nonetheless two and a half million years old by the time it reaches your eye. What you will not get is magnification for planets, and 7x50 is heavy enough that you will want something to lean on after a few minutes.
If a cheap department store telescope has ever put someone off astronomy, this is the fix. Check the price on Amazon
One question for you today: if you could point one instrument at one thing in the sky and get a straight answer, what would you aim it at? Hit reply, I read them all.
Nath, Future Technology
Forwarded? Get your own free briefing at futuretechnologyhq.com/newsletter
Some links may be affiliate links.