The Vibe Father
Google is reportedly building a chip with Gemini's architecture frozen into the silicon itself. According to The Information, the server chip — informally dubbed "Frozen v2" — is projected by the employees working on it to deliver 6 to 10 times more tokens served per unit of power than the newest version of Google's existing homegrown AI chips, with deployment planned for as early as 2028. Google has not confirmed any of this publicly, so treat it as a report, not a roadmap.
The story broke on July 20, 2026, when The Kobeissi Letter posted the details from The Information's reporting on X. Here's the full text
BREAKING. Alphabet, $GOOGL, is planning on launching a new 'frozen' chip to run its AI models more efficiently, per The Information.
Details include
1. This new server chip would directly integrate the blueprint of its Gemini AI model
2. The chip would then enable Alphabet to provide its AI models to users much more efficiently
3. The chip's developers estimate that it would be 6 to 10 times more efficient than the most recent version of Google's existing line of AI chips.
Another tech giant has entered the chip race.
What "baking Gemini into silicon" actually means
Strip the jargon and the idea is simple. A normal accelerator — a GPU, a TPU — is a general-purpose number cruncher. It can run any model you throw at it, but that flexibility costs energy, the chip spends power fetching instructions, shuffling data through generic pathways, and supporting operations your specific model may never use.
A "frozen" chip throws that flexibility away. You take Gemini's architecture — the shape of its layers, the math operations it actually performs, the way data flows between them — and you hardwire that dataflow directly into the chip's circuits. The silicon is the computation graph. No instruction overhead, no generic plumbing, no wasted transistors. That's where a 6-10x tokens-per-watt claim becomes physically plausible instead of marketing math, you're not running Gemini on a machine, you're running a machine that only knows how to be Gemini.
Analyst @kimmonismus laid out both sides of that bet the same day
Google may be preparing to freeze parts of Gemini's architecture directly into silicon. Informally called 'Frozen v2,' the chip reportedly targets 6-10x more tokens per watt than Google's newest TPUs. Deployment is planned for as early as 2028.
The motivation is immediate. Google's AI compute shortage has reportedly become severe enough that its Cloud division has turned down outside customers.
The efficiency comes with rigidity. Future Gemini models could use Frozen v2 only while retaining the same underlying architecture. TPUs reduced Google's dependence on Nvidia. Frozen v2 would go further, tying Gemini's architecture directly to the silicon running it.
Why Google would do this, inference economics
The reported motivation is not exotic. Per The Information, Frozen v2 is meant to address a major shortage in AI computing capacity inside Google — one that has reportedly fueled internal tensions and forced Google Cloud to turn down deals with outside customers. When you're refusing revenue because you can't serve the tokens, efficiency stops being an engineering nicety and becomes the business.
This is where the industry margin war actually lives now. Training gets the headlines, but serving cost is the line item that decides who can price aggressively. Every chat reply, every code completion, every agent loop is inference, and inference runs 24/7 at scale. If Frozen v2's projected numbers hold, Google gets two levers at once
- Undercut on price — serve Gemini tokens at a cost per million that competitors running general-purpose hardware can't match.
- Bank the margin — keep prices flat and let the 6-10x efficiency drop straight to the bottom line.
It's also one more step away from Nvidia. TPUs already reduced Google's dependence on outside silicon, a model-specific chip deepens that independence further. Worth remembering, though, we've covered Google's silicon-adjacent moves before, from the Gemini 3.5 Pro "Dragon" rumor to the Gemini 4.0 skip reports, and the pattern is consistent — Google plays a long vertical-integration game that most labs can't afford to copy.
The rigidity trade-off is real
Worth being skeptical here — of the risk, not just the promise. Model architectures evolve fast. The transformer variants that dominate in 2026 are not guaranteed to be what Gemini looks like in 2028, and 2028 is a long way out in this field. If Frozen v2 hardwires today's architecture and Gemini 2029 needs a different one, the chip's edge evaporates. You can't patch silicon.
That's architecture lock-in at the hardware level, and it's a genuinely different beast from software lock-in. A software dependency you can refactor. A fabbed chip is a multi-year, capital-heavy commitment. My read. Google is betting that Gemini's core architecture has stabilized enough that the efficiency win outweighs the optionality loss — and the internal compute shortage is urgent enough to force the bet. That may be right. It's still a bet, and the "6 to 10 times" figure is a projection from the chip's own developers, which is the most optimistic source you could possibly cite.
What this means if you build on Gemini
No chip ships to your API bill tomorrow. But the direction matters, because inference economics eventually shows up in the two things developers actually feel API prices and rate limits. If Google can serve tokens at a fraction of today's power cost by 2028, the realistic outcomes are cheaper Gemini tiers, more generous quotas, or both — and pressure on every other provider's pricing in response.
There's a practical lesson buried in the report too. Google is optimizing its entire stack around measured cost per token served. You should do the same with your model choices. Don't pick a model on benchmark vibes or Twitter consensus — measure price/performance on your actual workload. That's literally what our AI model benchmarks exist for, and it's a discipline that pays off whether or not Frozen v2 ever ships.
What to watch
Three concrete signals between now and 2028, whether Google confirms any of this (so far, silence), whether Gemini API pricing starts drifting down ahead of the hardware as capacity pressure eases, and whether any competitor — Nvidia included — answers with its own model-specific silicon story. Until one of those moves, file Frozen v2 where it belongs, a credible, well-sourced report about where inference economics is pushing the biggest players, and a reminder that tokens per watt is the number the whole industry is quietly optimizing for.
