📰 Key Highlights

Alphabet (Google’s parent company) is designing a new server chip for its Gemini model, internally codenamed “Frozen v2.” According to a report by The Information citing anonymous sources, the chip is expected to launch in 2028 and could be 6 to 10 times more efficient than Google’s current AI chips — measured by the number of tokens generated per unit of power. When asked by TechCrunch, Google didn’t directly confirm the report but also didn’t deny it, only saying that the team continues to research and experiment with new technologies to deliver maximum performance and efficiency, and emphasizing that through full-stack co-design from hardware to software, they ensure the system can be highly optimized for real-world workloads — though not every project ultimately goes into mass production.

AI companies have been increasingly aggressive about building their own chips lately — partly to boost their own model efficiency, and partly to deal with the global shortage of AI compute. As concerns over AI spending gradually cool down the market’s past frenzy, efficiency has become a key selling point for tech companies. At the same time, everyone’s still trying to reduce their dependence on chip giant Nvidia and break free from being held hostage by its hardware supply. Just this past June, OpenAI unveiled its first self-designed chip, an inference processor codenamed “Jalapeño.” Earlier this month, there were also reports that Anthropic is in talks with Samsung about chip manufacturing collaboration.

Investors have long worried about Alphabet’s massive AI infrastructure spending plan — Google said earlier this year it plans to spend $180 to $190 billion. News of Frozen v2’s efficiency boost seems to have eased those concerns: Google’s stock rose about 3% in Monday morning trading after the report, giving a boost to the earnings report due later this week.

💬 JudyAI Lab Perspective

Google’s parent company Alphabet is reportedly developing a new server chip called “Frozen v2” for its Gemini model — one that could be 6 to 10 times more efficient than its current AI chips, with a planned launch in 2028. This signals that the competitive focus among major AI players is shifting from simply stacking compute power to an efficiency war over how many tokens you can squeeze out per unit of electricity.

This news reflects a clear industry turning point: as AI spending concerns have cooled the market, “efficiency” is replacing “scale” as the new selling point. OpenAI rolled out its self-designed inference chip “Jalapeño” back in June, and Anthropic is reportedly also in talks for chip manufacturing partnerships. The big AI companies are all independently marching toward in-house hardware — partly to boost their own model efficiency, and partly to reduce long-term dependence on Nvidia. For AI builders out there, this is a reminder that infrastructure autonomy and cost efficiency are gradually becoming competitive dimensions just as important as model capability.

Worth thinking about: when you’re evaluating AI tools or services, beyond just looking at model capability, keep an eye on the underlying compute efficiency and cost structure — that often determines whether a service can be delivered reliably over the long haul.

📅 Source Info

🔗 Further Reading