Alphabet is designing a server chip, internally called Frozen v2, intended to run Gemini more efficiently, according to TechCrunch’s report. The project is reportedly aimed at a 2028 release and could deliver six to 10 times more generated tokens per unit of power than Google’s current AI chips; Google neither confirmed nor denied the specific programme.
That makes this a reported plan, not a product announcement. Google did say its teams continually research hardware and software together to improve real-world performance. The company has also told investors it expects $180 billion to $190 billion of spending as it builds its AI strategy, putting unusual pressure on every watt and every delivered token.
Custom silicon is no longer a side project for the hyperscalers. Nvidia’s hardware remains dominant, but the biggest model providers increasingly want control over the cost, supply and performance of inference: the work of serving models after training. TechCrunch notes that OpenAI announced an inference processor in June, while Anthropic has reportedly discussed a chip partnership with Samsung.
Unique insight
The decisive contest is shifting from who can train the largest model to who can afford to run a useful one at planetary scale.
Analysis
If the reported efficiency range holds up, it changes the arithmetic for Google Cloud customers and Gemini product teams: the same power and data-centre estate can serve more requests, or comparable services can be offered at lower cost. That directly affects enterprise buyers, developers building agentic products, and Nvidia’s most valuable customers, because software workloads will increasingly be designed around the economics of a provider’s own silicon. Over the next two years, investors should judge these programmes by deployed capacity and cost per useful task, not prototype claims or headline benchmark scores.
Bottom line — If I only remember one thing tomorrow, what should it be?
Google’s reported Frozen v2 effort shows that cheap, power-efficient inference is becoming the practical moat in AI, even before the chip reaches production.
- Story type: Breaking news
- Impact: High
- Confidence: Medium (multiple reports)
- Audience: Business
- What changed: Google is reportedly developing a 2028 server chip intended to make Gemini inference substantially more power-efficient.
- What to watch: Google’s next earnings disclosure and any confirmation of Frozen v2’s deployment schedule or tokens-per-watt target.
- Reading time: 2 minutes
