Google's "Frozen v2": A Chip Strategy That Reveals More About Fragility Than Strength

Google is developing an experimental AI server chip, internally codenamed "Frozen v2," that would permanently etch core elements of the Gemini architecture into silicon, according to a report first surfaced today via The Information. The goal: a 6x-to-10x leap in tokens processed per unit of power versus current-generation TPUs. Targeted for deployment around 2028, the chip would complement rather than replace Google's existing TPU line. Alphabet shares rose roughly 1-3% intraday on the news.

Evolution, Not Revolution

Google has walked this path before, with TPUs delivering 30-80x better performance-per-watt than general-purpose hardware when they launched in 2015. Frozen v2 is the next rung on that ladder — though notably, the reported design freezes stable computational operations while keeping model weights updateable. That distinction determines whether the chip survives multiple Gemini revisions or becomes obsolete overnight.

The Capital-Intensity Spiral

The more urgent story is Alphabet's balance sheet. 2026 capex is projected at $180-190 billion — six times 2022 levels — with a further increase already flagged for 2027. Despite cutting Gemini serving costs 78% during 2025, infrastructure spending accelerated, and depreciation rose 38% to $21.1 billion. Efficiency gains are being absorbed by usage growth, not returned as savings. Frozen v2 is designed to prevent capacity scarcity from capping revenue, not to shrink the footprint — a reading reinforced by Alphabet's parallel $4.75 billion acquisition of Intersect for energy capacity.

The Roadmap Mismatch: Optimizing a Model That Isn't Winning

There is a harder problem the chip announcement doesn't address: which Gemini is being optimized. Gemini 3.5 Pro, originally slated for a June 2026 release, has now missed three separate deadlines, with prediction markets pricing roughly 73-81% odds against a launch materializing on either of its two most recent target dates. Google reportedly scrapped the existing 2.5 Pro architecture entirely and restarted pre-training from scratch, a rebuild aimed at closing the gap with OpenAI's GPT-5.6 and Anthropic's Fable 5. According to Bloomberg reporting cited in coverage of the delay, Google has been taking additional time specifically to improve coding capabilities, and a data retraining effort in late June produced disappointing results. Meanwhile, the current production flagship, Gemini 3.1 Pro, dates back to February — an eternity in frontier-model time. Some reporting even suggests Google may skip the 3.5 Pro generation entirely and redirect resources toward a future Flash model, while stopgap releases are prepared to buy time.

This matters directly for Frozen v2's investment case. A chip that hardwires efficiency gains into a non-frontier model architecture is optimizing throughput on a product Google itself is racing to replace. If Gemini's competitive position continues slipping against rivals already in market, the strategic payoff from squeezing more tokens-per-watt out of the current generation shrinks — the industry rewards frontier capability, not cheaper serving of a model customers are migrating away from. This is not a hypothetical risk; the raw analysis underlying this piece independently flagged "research volatility" — the mismatch between silicon timelines fixed years in advance and model architectures that can change in months — as Google's central Achilles' heel. The Gemini 3.5 Pro delays are that risk materializing in real time, months before Frozen v2 has even been confirmed by Google.

A Hedge, Not a Weapon

The most consequential reframing: Frozen v2 is best understood as a defensive balance-sheet technology dressed as an offensive chip strategy. Google isn't primarily racing to out-engineer Nvidia — it's protecting a business model converting high-margin software revenue into capital-intensive infrastructure at an accelerating rate, while its actual model roadmap struggles to stay current.

Four implications follow. The durable moat is captive, predictable demand — not the silicon. Energy-secured compute capacity, not processor speed, is the scarcer asset. Nvidia's unit share can decline even as revenue grows. And AMD, lacking both Nvidia's software moat and hyperscaler-owned workloads, may be the more exposed casualty.

What This Means for Capital Allocators

The signal for executives isn't that Google found a permanent moat — it's that infrastructure economics are under more strain than headline revenue growth suggests, and the company is now hedging against scarcity for a model franchise it is simultaneously struggling to keep competitive.

not investment advice

Sources: https://www.theinformation.com/articles/google-plans-new-frozen-chip-run-ai-models-efficiently