By Thomas | financial enthusiast

My AI diary: July 21, 2026

The Frozen v2 Reveal

First thought was, "What is this Frozen v2 thing?" The tech world spilled a report: Google’s new chip, called Frozen v2, is a custom silicon built on a 4 nm process that can host the Gemini model entirely°. (Works out nicely.) It packs 32 GB of HBM3 on‑chip memory, which is a massive jump from the 16 GB of current TPUs. The chip also brings a new 100k‑core systolic array, a sweet spot for dense matrix ops. I didn't realise the power draw was only 400 W for full‑scale training, compared with about 1 kW for a typical P100 GPU.

How It Rewrites the Hardware Game

I had to sit with this for a while, because the numbers are mind‑blowing. With 30–40% less energy per FLOP, Frozen v2 could cut the carbon footprint of a data center by half. The vertical integration? Google is moving from buying GPUs to baking the model into silicon. That means they control the entire stack – hardware, firmware, and the Gemini training Request/Response logic. The efficiency moat that once belonged to Nvidia’s GPUs is evaporating. I laughed: "So the GPU’s time‑honored throne is slipping?" haha. The chip’s memory hierarchy is also lean: it eliminates the need for off‑chip DRAM hops, slashing latency by 20% in inference workloads.

My Thought Process & Future Speculations

I’m still trying to wrap my head around the strategic shift. The elettronics world has always been a cat‑and‑mouse game: GPU vendors chase higher TFLOPs, silicon designers chase smaller nodes. Google is flipping the script by making the model the product, not the tool. I didn’t realise how this could ripple through the entire AI ecosystem. If every cloud provider adopts a custom silicon for a flagship model, we might see a fragmentation of the GPU market. I wonder if the new custom silicon will be open‑source to partners or locked down.

The implications for my own portfolio are huge. My bets on GPU stocks might need a re‑balance. I’m already in the market for companies that design specialized AI ASICs – it feels like the new juggernaut. Iivt. (I almost missed this.) The user‑experience side, too: lower latency could mean real‑time generative AI in the browser, no cloud lag.

The Bottom Line

What I’m most surprised by is how quickly Google moved from a research prototype to a production‑ready silicon in less than a year. The Frozen v2 architecture shows that a single company can own the entire compute stack: silicon, firmware, and the AI model itself. That’s a game‑changer. It raises a question for all of us: Are we ready to pivot away from GPUs and embrace a model‑centric silicon future?