# Google developing “Frozen v2” chip to run Gemini models up to 10×

**Published:** 2026-07-22T18:11:48.065Z  
**Topic:** Google Ai  
**Sentiment:** neutral  
**Publisher:** TrendWatcher — https://www.trendwatcher.in/article/9af85029-4bcc-4f7e-a3f4-9b2dd39a1a57

Google’s “Frozen v2” AI chip aims for 6‑10× token‑per‑watt gains, targeting 2028 rollout and boosting Gemini cost competitiveness.

Google disclosed that its upcoming “Frozen v2” server chip will hard‑wire Gemini’s architecture into silicon, promising six to ten times more tokens generated per watt of electricity compared with current TPU‑based setups [2]. The efficiency boost is intended to lower the cost of serving Gemini models and help the company keep pace with rivals that already enjoy lower per‑token expenses.

| At a glance | |
|---|---|
| Chip name | Frozen v2 |
| Efficiency gain | 6‑10× tokens per watt |
| Target deployment | 2028 (earliest) |
| Related model launch | Gemini 3.6 Flash (up to 17% fewer tokens) [1] |

## Custom silicon for Gemini  
Google’s strategy pairs the new chip with the latest Gemini family, which includes Gemini 3.6 Flash that uses up to 17% fewer tokens and costs less per token than its predecessor [1]. By embedding Gemini’s routing blueprint directly into the chip, “freezing” the architecture eliminates redundant calculations and reduces data movement across memory, a key source of energy waste in conventional TPUs [2]. The result, according to engineers, is the ability to serve ten queries for the power cost of one, a dramatic improvement that could translate into billions of dollars saved at Google’s scale [2].

## Competitive context  
Anthropic’s Mythos model already offers a cost advantage in automated code defense, while OpenAI’s GPT‑5.6 Terra Max and Chinese models such as Kimi K3 and Qwen 3.8 Max compete on price and performance [1]. Artificial Analysis data shows Gemini 3.6 Flash already undercuts these rivals on cost per task [1]. However, Google has faced capacity constraints—Meta reportedly had to ration Gemini usage in March because Google could not meet demand [2]. The “Frozen v2” chip is positioned as a remedy to this bottleneck, aiming to reduce reliance on external GPU rentals (Google is paying SpaceX $920 million a month for Nvidia GPUs) and lessen exposure to Nvidia’s dominant AI GPU market [2].

## Market implications  
If the projected efficiency gains materialize, Google could offer Gemini at lower token prices, narrowing the cost gap with Anthropic and Chinese providers that currently run 60‑90% cheaper [2]. A cheaper‑to‑run Gemini would also improve Alphabet’s margins on AI services, a factor investors noted as Alphabet shares rose roughly 3% after the news broke [2]. The chip’s design, however, is model‑specific and will not be offered to external Cloud customers, limiting its broader industry impact [2].

## What to watch
- **2028 rollout** – earliest expected deployment of Frozen v2, according to reports [2].
- **Alphabet Q2 2026 earnings** – scheduled for July 22, where the financial impact of the chip and Gemini launches may be disclosed [2].
- **Rival chip programs** – Nvidia’s continued dominance and competing custom silicon from Microsoft, Amazon, and Meta could shape Google’s market positioning.

The significance of “Frozen v2” lies in its potential to turn efficiency into a competitive lever for Google’s AI offerings, but the timeline and actual performance remain uncertain until hardware prototypes move toward production.

## Sources
1. CNBC — [Google expands Gemini lineup with cheaper models and new Mythos rival](https://www.cnbc.com/2026/07/21/google-gemini-flash-ai-mythos-rival.html)
2. Decrypt — [Google Is Building an AI Chip Just for Gemini—And Investors Already Moved On It](https://decrypt.co/373967/google-ai-chip-gemini-frozenv2)
3. eWeek — [Google’s Planned 'Frozen v2' Chip Could Make Gemini More Efficient](https://www.eweek.com/news/google-frozen-v2-ai-chip-gemini-efficiency/)

---
Cite as: TrendWatcher, "Google developing “Frozen v2” chip to run Gemini models up to 10×", https://www.trendwatcher.in/article/9af85029-4bcc-4f7e-a3f4-9b2dd39a1a57