Future Technology[Daily Briefing] Gemini in silicon 🔩, Oracle under attack 🛡, exoplanet air 🪐
Google is trying to etch Gemini straight into the chip
For years the plan for running a big AI model was simple: rent a mountain of general purpose GPUs and feed them electricity. Google now looks like it wants out of that deal. Reports point to an experimental server chip, nicknamed Frozen v2, that hard wires parts of the Gemini model architecture directly into silicon.
The logic is blunt. If you already know exactly which maths your model runs, over and over, you stop paying the flexibility tax of a general chip and bake the pattern into the hardware. It is the same instinct that gave us the TPU, pushed one level deeper, from AI accelerator to something closer to a model shaped circuit.
The catch is just as clear. A chip tuned this tightly to one model is brittle, and architectures still move fast while silicon takes years and fortunes to design. Google is betting that the core shapes of a transformer are now stable enough to freeze. If that bet lands, the cost and power savings arrive right as AI energy bills turn into a boardroom problem.
Why it matters: Every hyperscaler is quietly trying to peel away from Nvidia, and model specific silicon is the sharpest tool yet. If Frozen v2 works, expect Amazon, Meta and Microsoft to chase their own frozen chips, and expect the AI features you use every day to get cheaper to run.
Read the full breakdown on Future Technology
Quick hits
Oracle's back office is under active attack. Enterprise ERP is the plumbing that runs payroll and supplier payments, and attackers know it patches slowly. Oracle's latest security round targets flaws already being exploited in the wild, so this is a patch tonight situation for anyone running it. What to patch and why · Source: BleepingComputer, 4 min read.
Everyone suddenly wants their own AI chip. DeepSeek is designing its own inference silicon to cut its reliance on Nvidia, part of a wider scramble where compute is now treated like national infrastructure rather than a rental. Inside the sovereign silicon race · Source: TechStartups, 5 min read.
We read the air on a rocky world light years away. Astronomers detected an atmosphere around a small rocky exoplanet, a genuinely hard measurement that nudges us one step closer to sniffing out the chemistry of another Earth. How they pulled it off · Source: Nature Astronomy, 6 min read.
A rogue agent helped breach Hugging Face. The AI hub was hit in an incident that pulled in OpenAI tooling and an automated agent gone wrong, a small preview of how AI supply chains are going to get attacked. What actually happened · Source: The Hacker News, 4 min read.
Intel hits a real High NA EUV milestone. Intel says it is now running High NA EUV lithography at high volume, the bleeding edge of chipmaking and a rare piece of genuinely good news for its foundry ambitions. Why High NA matters · Source: Future Technology, 5 min read.
Tool of the day: a Raspberry Pi 5 as your always on home server
Honest take: most people do not need a pricey NAS or a stack of cloud subscriptions to run the small stuff. A Raspberry Pi 5 sipping a few watts in a drawer will happily handle Pi-hole, a personal dashboard, local backups and a dozen other quiet jobs. It is fiddly on day one and quietly wonderful by week two. If you enjoy tinkering, it is the best value box in tech right now. If you do not, be honest with yourself and skip it, because it does ask for a little patience.
One question before you go: if a chip could be frozen around today's best AI model, would you take the cheaper frozen hardware, or pay more to stay flexible for whatever comes next? Hit reply and tell me, I read every one.
Nath, Future Technology
Forwarded this? Get your own free briefing at futuretechnologyhq.com/newsletter
Some links may be affiliate links.